Prévia do material em texto
Intelligent DC Series Arc Fault Detection using Deep Learning
in Photovoltaic Systems
Author:
Lu, Shibo
Publication Date:
2021
DOI:
https://doi.org/10.26190/unsworks/2236
License:
https://creativecommons.org/licenses/by-nc-nd/3.0/au/
Link to license to see what you are allowed to do with this resource.
Downloaded from http://hdl.handle.net/1959.4/70737 in https://
unsworks.unsw.edu.au on 2024-02-08
http://dx.doi.org/https://doi.org/10.26190/unsworks/2236
https://creativecommons.org/licenses/by-nc-nd/3.0/au/
http://hdl.handle.net/1959.4/70737
https://unsworks.unsw.edu.au
https://unsworks.unsw.edu.au
INTELLIGENT DC SERIES ARC FAULT
DETECTION USING DEEP LEARNING IN
PHOTOVOLTAIC SYSTEMS
Shibo LU
Supervisor: A/Prof. Toan PHUNG
Secondary supervisor: Dr. Daming ZHANG
A thesis in fulfilment of the requirements for the degree of
Doctor of Philosophy
School of Electrical Engineering & Telecommunications
Faculty of Engineering
University of New South Wales
November 2020
Surname/Family Name : LU
Given Name/s : Shibo
Abbreviation for degree as give in the University calendar : Ph. D.
Faculty : Engineering
School : School of Electrical Engineering and Telecommunications
Thesis Title : Intelligent DC Series Arc Fault Detection Using Deep Learning in Photovoltaic Systems
Abstract 350 words maximum:
Grid integration of renewable sources including solar energy is growing faster than ever before. Nowadays, solar
power development is increasing throughout the world, and solar photovoltaic (PV) systems play an important role to
support the main loads and micro-grids. However, one needs to consider the long-term performance of PV components.
Their deterioration can be caused by various factors such as ageing, weathering, the higher DC operating voltage level,
improper installation, inadequate maintenance, etc. The consequence is a growing potential of electrical arcing incidents
especially the series arc fault in PV systems. Without timely detection and interruption, such dangerous events can cause
catastrophic fires, posing a severe threat to human safety and properties.
In this thesis, a comprehensive review of DC arc fault and their diagnosis methods in PV systems is presented.
Experimental study of DC series arc fault characteristics is carried out. The feasibility of applying deep learning (DL) in
series arc fault detection in PV systems is systematically investigated. Specifically, convolutional neural networks (CNN)
are successfully applied and demonstrate superior diagnosis performance over conventional machine learning algorithms
and other popular DL algorithms. For cost-effective real-time deployment, a lightweight CNN structure is designed to
achieve a good balance between model complexity and detection accuracy. Moreover, novel frameworks, including
domain adaptation and deep convolutional generative adversarial network (DA-DCGAN) and lightweight transfer
convolutional neural network with adversarial data augmentation (LTCNN-ADA), are proposed. They aim to address the
challenges when applying DL to practical applications, including lack of fault data from the field, data inconsistency
between laboratory and field, and limited computation resources in edge devices. The proposed methods are validated
through comprehensive offline analysis using pre-recorded data. In addition, the trained DL classification models are
deployed in an embedded system and tested in single-phase and three-phase PV systems in real-time under different test
conditions. Both offline and online experimental results show that the proposed methods can accurately and reliably
detect series arc fault in PV systems.
Declaration relating to disposition of project thesis/dissertation
I hereby grant to the University of New South Wales or its agents the right to archive and to make available my thesis or dissertation in whole
or in part in the University libraries in all forms of media, now or here after known, subject to the provisions of the Copyright Act 1968. I retain
all property rights, such as patent rights. I also retain the right to use in future works (such as articles or books) all or part of this thesis or
dissertation.
I also authorise University Microfilms to use the 350 word abstract of my thesis in Dissertation Abstracts International (this is applicable to
doctoral theses only).
……………………………………………………………
Signature
……………………………………..………………
Witness Signature
……….……………………...…….…
Date
The University recognises that there may be exceptional circumstances requiring restrictions on copying or conditions on use. Requests
for restriction for a period of up to 2 years must be made in writing. Requests for a longer period of restriction may be considered in
exceptional circumstances and require the approval of the Dean of Graduate Research.
FOR OFFICE USE ONLY Date of completion of requirements for Award:
I
Originality Statement
I hereby declare that this submission is my own work and to the best of my knowledge it
contains no materials previously published or written by another person, or substantial
proportions of material which have been accepted for the award of any other degree or diploma
at UNSW or any other educational institution, except where due acknowledgement is made in
the thesis. Any contribution made to the research by others, with whom I have worked at UNSW
or elsewhere, is explicitly acknowledged in the thesis. I also declare that the intellectual content
of this thesis is the product of my own work, except to the extent that assistance from others in
the project's design and conception or in style, presentation and linguistic expression is
acknowledged.
Signed……………………………….
Date……………………
II
COPYRIGHT STATEMENT
I hereby grant the University of New South Wales or its agents the right to archive and to make
available my thesis or dissertation in whole or part in the University libraries in all forms of media,
now or here after known, subject to the provisions of the Copyright Act 1968. I retain all
proprietary rights, such as patent rights. I also retain the right to use in future works (such as
articles or books) all or part of this thesis or dissertation.
I also authorise University Microfilms to use the 350 word abstract of my thesis in Dissertation
Abstract International (this is applicable to doctoral theses only). I have either used no
substantial portions of copyright material in my thesis or I have obtained permission to use
copyright material; where permission has not been granted, I have applied/will apply for a partial
restriction of the digital copy of my thesis or dissertation.
Signed ..............................................................................
Date ..............................................................
AUTHENTICITY STATEMENT
I certify that the Library deposit digital copy is a direct equivalent of the final officially approved
version of my thesis. No emendation of content has occurred and if there are any minor
variations in formatting, they are the result of the conversion to digital format.
Signed ..............................................................................
Date................................................................
III
A THESIS DEDICATED TO MY PARENTS.
IV
Acknowledgement
It has been a long and hard journey towards a Ph. D., and I have finally reached the
end.
I would like to express my deepest gratitude to my primary supervisor, A/Prof.
Toan Phung. Without your expert guidance, insightful suggestions, and encouragements
during the past few years, the accomplishment of this thesis would not be possible.
Your inspiring minds and firm support have undoubtedly taken my research to another
level.Arc Model
Cassie and Mayr arc models are the most popular and widely used, developed
based on the principle of energy conversion. The Cassie arc model can be expressed as
below:
1
𝑔
𝑑𝑔
𝑑𝑡
=
1
𝜏
(
𝑢𝑎𝑟𝑐𝑖𝑎𝑟𝑐
𝑉𝑐
2
− 1) (2.2)
𝑔 =
𝑖𝑎𝑟𝑐
𝑢𝑎𝑟𝑐
(2.3)
where 𝑔 denotes the arc conductance, 𝑢𝑎𝑟𝑐 is the arc voltage, 𝑖𝑎𝑟𝑐 is the arc current, 𝑉𝑐
is constant arc voltage, and 𝜏 is the arc time constant. In this model, Cassie assumed the
power loss is caused by forced convection, which means the area of arc cross section is
proportional to the arc current; thus, Cassie arc model is better for simulation of high
arc current level. On the contrary, the Myer arc model is more suitable for low current
arcs, because Myer assumed the power loss is caused by thermal conduction and it
remains constant [45]. The equation is shown below:
1
𝑔
𝑑𝑔
𝑑𝑡
=
1
𝜏
(
𝑖𝑎𝑟𝑐
2
𝑃
− 1) (2.4)
where 𝑔 denotes the arc conductance, 𝑖𝑎𝑟𝑐 is the arc current, 𝑃 is the static cooling
power, and 𝜏 is the arc time constant determined empirically. The unknown constant 𝑃
in (2.4) can be calculated and determined through observation of the experimental
results [46]. There are also many other physics-based arc models that can potentially be
used for PV application, such as Lowke’s model and modified Mayr model [42], [47].
However, those models involve more parameters, which are not easy to be implemented
in the simulation.
20
2.3.2. V-I Characteristic-based Arc Model
From 1902 till now, many empirical models such as Hertha Ayrton model,
Steinmetz model, Van and Warrington model, Paukert model, etc. have been proposed
[43]. Some of the more popular models are described in the following.
2.3.2.1. Nottingham Arc Model
Nottingham developed an arc equation as shown in (2.5):
𝑉𝑎𝑟𝑐 = 𝐴 +
𝐵
𝐼𝑎𝑟𝑐
𝑛 (2.5)
where the value of 𝐴, 𝐵 , and 𝑛 depends on the arc length and type of electrode material.
It covers the arc length between 1 mm to 10 mm (0.0394 in. to 0.394 in.), and current
level up to 10 A. When the electrode material is copper, which is the most common
material used in PV systems, the parameters of the equation are: 𝐴 = 27.5, 𝐵 = 44, and
𝑛 = 0.67 [48]. This equation is only suitable to simulate series arc fault at below string
level because of the limited range of current and arc length.
2.3.2.2. Hall, Myer, and Viicheck Arc Model
Nottingham’s arc model only covers low current levels because of the lack of high
power DC source in early years. Hall et al. carried out arc testing with high current level
ranged from 300-2400 A, and air gap widths ranged from 4.8-152 mm in 1978 [49]. It
was found that the experiment results match the estimation proposed by Nottingham in
(2.5). Hence, it is capable of simulating series arc faults with high current level (i.e. at
the DC side of the solar inverter in a large PV system) and parallel arc faults.
21
2.3.2.3. Stokes and Oppenlander Arc Model
Stokes and Oppenlander Model was developed in the most exhaustive way among
other existing models as it covers arc current range from 0.1 A to 20 kA and air gap
widths from 5 mm to 500 mm with electrodes in series [41]. It is widely used in
incident-energy estimation for DC arc fault in industrial applications [50]. The V-I
characteristic of arc current levels above the transition point line (constant voltage
region) is shown below:
𝑉𝑎𝑟𝑐 = (20 + 0.534𝐿)𝐼𝑎𝑟𝑐
0.12 (2.6)
𝐼𝑡𝑟𝑎𝑛𝑠𝑖𝑡𝑖𝑜𝑛,𝑎𝑟𝑐 = 10 + 0.2𝐿 (2.7)
where 𝐿 denotes the air gap width in mm. This arc model is very suitable for high
current arc simulation such as the series arc fault in the combiner box and solar inverter,
and the parallel arc fault between two strings.
2.3.2.4. Paukert Arc Model
In order to cover both “constant power” region and “constant voltage” region,
Paukert has developed a model that covers arc current range from 0.3 A to 100 kA and
air gap widths range from 1 mm to 200 mm as follows:
𝑉𝑎𝑟𝑐 = {
𝐴
𝐼𝑎𝑟𝑐
𝐵 , 𝐼 ≤ 100𝐴
𝐴𝐼𝑎𝑟𝑐
𝐵 , 𝐼 > 100𝐴
(2.8)
where A and B are positive arc constants, varying different gap width [51]. Because of
its wide current range, this arc model can be used to simulate both series and parallel
arc faults in PV systems at different levels with the fixed gap width. However, it is quite
difficult to be implemented when gap width changes continuously as it is not expressly
defining the gap width.
22
2.3.2.5. Modified Paukert Arc Model
In [52], to further reveal the dependency between the V-I curve and gap width, a
modified Paukert arc model with gap width integrated in both arc constants A and B has
been proposed for low-current series arc:
𝑉𝑎𝑟𝑐 =
𝐴 + 𝐶𝐿
𝐼𝑎𝑟𝑐
𝐵+𝐷𝐿 (2.9)
where A, B, C, and D are arc constants dependent on experimental condition, and L is
the gap width in inch. This model is proposed for modern DC power system with
several hundred volts of operating voltage, which is often the case for current PV
systems. Because of the low current and small gap width, it is very suitable for the
simulation of the series arc fault at below PV string level.
2.3.3. Heuristic Arc Model
The Heuristic model is also based on experimental observation. However, it has
additional ad-hoc-parameter to improve the correlation between simulation and
experimental results. In [53], a practical series DC arc model independent of electrical
time constants has been developed based on a large set of experimental data. The arc
voltage can be represented by the sum of two hyperbolic approximated equations: a
nonlinear voltage component (the arc column voltage) and an electromagnetic force
component (the anode and cathode voltage) as follows:
𝑉𝑛 = 𝑉𝑑𝑐,𝑠𝑜𝑢𝑟𝑐𝑒(0.5 + 0.5tanh (𝛼(𝑞 − 1))) (2.10)
𝑒𝑔𝑎𝑝 = 0.5(𝑎 + 𝑏𝑥𝑔𝑎𝑝)(tanh(𝜆𝑞) − tanh λ(𝑞 − 1)) (2.11)
𝑞 =
𝑥𝑔𝑎𝑝
𝑥𝑒𝑥𝑡𝑖𝑛𝑐𝑡𝑖𝑜𝑛
(2.12)
where 𝑞 represents the separation ratio (jitter), 𝑥𝑔𝑎𝑝is the air gap width, 𝑥𝑒𝑥𝑡𝑖𝑛𝑐𝑡𝑖𝑜𝑛 is
the critical distance between two electrodes when arc will quench, parameters 𝑎, 𝑏, 𝛼,
23
and 𝜆 are relevant to offset of 𝑒𝑔𝑎𝑝, voltage gradient of 𝑒𝑔𝑎𝑝, slope of nonlinear voltage
𝑉𝑞, and slope of 𝑒𝑔𝑎𝑝, respectively. An arc will be established when 𝑞 = 0+, while it
will be extinct when 𝑞 = 1 . Obviously, 𝑥𝑔𝑎𝑝 can be determined by measurement
whereas 𝑥𝑒𝑥𝑡𝑖𝑛𝑐𝑡𝑖𝑜𝑛 cannot be precisely measured because it is a random variable. Thus,
𝑥𝑒𝑥𝑡𝑖𝑛𝑐𝑡𝑖𝑜𝑛 could represent the random behaviors of simulated arc in terms of 𝑞, which
is given by:
𝑞|𝑡=𝑘𝑇 = 𝑞|𝑡=𝑘𝑇−1 + 𝑟𝑎𝑛𝑑(𝑐)10 (2.13)
where 𝑘 ∈ 𝑍, 𝑇 is the current simulation time step, and 𝑟𝑎𝑛𝑑(𝑐) is a random number
between 0 and 𝑐. Note that 𝑞 will be set to a new random number when 𝑞 > 𝑞𝑡ℎ (𝑞𝑡ℎThe high-frequency variation is closely relevant to the
discharge of the plasma because of the chaotic nature of the arc. Therefore, it is
important to account those factors in the arc noise modelling.
In order to describe the arc random behaviour, zero-mean Gaussian noise could be
added into the arc signal [55], [56]. The high-frequency part in the arc fault signal can
be fitted into the equation as shown in (2.14).
𝑓(𝑥) =
1
√2𝜋𝜎
exp (−
(𝑥 − 𝜇)2
2𝜎2
) (2.14)
where 𝜇 is the mean, and 𝜎 is the variance of the noise. The value of 𝜇 is assumed to be
0 in the noise modelling. It is found that the arc noise in PV systems can be
characterised by the pink noise (1/𝑓 noise), where the magnitude has an opposite trend
of the frequency [16], [57]. In [58], pink noise is used to simulate the AC component of
the arc fault. In [59], a modified pink noise model is introduced to better simulate the
arc faults in different physical states. It is also useful for simulations involving analysis
of the propagations of arc fault signal in wires and devices that attenuate noise. Based
on evaluations, the simulated signals match the experimental signals in the frequency
domain. In [53], an additional equation is established to represent the random
behaviours of the arc as mentioned earlier.
25
Table 2.1 Summary of DC arc fault models for simulation
Arc model Current level
Gap width
(mm)
Series arcing?
Parallel
arcing?
High frequency
variation
Myer Low Depends
Yes, at below string
level
No
Gaussian/Pink
noise
Cassie High Depends Yes, at array level Yes
Gaussian/Pink
noise
Nottingham
Low (below
10 A)
1-10
Yes, at below string
level
No
Gaussian/Pink
noise
Hall, Myer,
and Vilicheck
High (300 A
to 2400 A)
4.8-152 Yes, at array level Yes
Gaussian/Pink
noise
Stokes and
Oppenlander
High (Above
transition line)
5-500 (5, 20,
100, 500)
Yes, at array level Yes
Gaussian/Pink
noise
Paukert
Low/High (0.3
A-100 kA)
1-200 Yes Yes
Gaussian/Pink
noise
Modified
Paukert
Low (below
25 A)
1-3
Yes, at below string
level or at combiner
box with small
current
No
Gaussian/Pink
noise
Heuristic arc
model [53]
Low/High Depends Yes No
Additional
equation
Heuristic arc
model [54]
Low/High Depends Yes Possible
Additional
equation
2.4. DC Arc Faults Detection Methods in PV Systems
AC arc fault recognition and detection have been widely researched for a long time,
while DC arc fault is far less developed [60]. With the release of the outline and
standard related to DC arc fault protection in PV system in 2013 and 2018, respectively,
the demand for effective DC arc fault detection algorithms and products is rapidly
increasing [14], [15], [61].
Besides PV-specified detection methods, some detection methods for other DC
systems such as electrical vehicle and DC microgrids will be reviewed in this section as
well, because they can be adopted in PV systems without or with some minor
modifications.
26
2.4.1. Sensors for Measurement
There are several types of signals available for arc fault detection purpose in
different DC system applications. Current is more commonly used than other signals
such as voltage and electromagnetic signals. Voltage is less popular mainly because the
fault locations are unknown, and the arc voltage cannot be directly measured.
Furthermore, many voltage sensors are typically required than current sensors in order
to cover the entire system [20]. Radio frequency sensors capture the induced
electromagnetic wave when series arc fault occurs. In the higher frequency band, there
are less impacts from the common interferences such as power electronics noise.
However, the electromagnetic signals can be attenuated by physical obstructions
between the fault source and the sensors. For example, in practical conditions, a series
arc fault can occur in MC4 connectors at the backside of the PV panel. In this case, the
PV panel itself can be the obstacle between the arc fault and the sensors. Other sensors
such as optical, acoustic, and thermal imaging sensors also have potential for series arc
fault detection. However, similar to radio frequency sensors, their detection ranges are
generally limited. To apply those sensors in large PV applications, studies of optimal
sensor locations are required. Therefore, the measured system current signals are used
for the majority of detection methods.
2.4.2. Fast Fourier Transform
Fourier analysis is a classic approach to frequency domain analysis. It is found that
the noise floor will increase after arc fault occurrence [33], [57], [62]–[65]. J. Johnson
et al. from Sandia Laboratory carried out a series of tests related to DC arc faults in PV
systems, and the frequency band of 1-100 kHz is recommended for fault detection [35],
[57], [66].
27
Fast Fourier transform (FFT) algorithm is convenient to implement since it is
generally available for majority of software libraries. The detection methods based on
FFT can be found in [61], [63], [64], [67]–[76] . Most methods compare the amplitude
or power of the frequency spectrum between the arcing state and non-arcing state (or a
pre-defined threshold value calculated based on the frequency spectrum of non-arcing
state) to make decision. In [67], the sum of frequency spectrum of the CT current in the
frequency band of 0-12.5 kHz is employed using the sampling frequency of 250 kHz. In
[61], [68], [69], the contents in frequency spectrum of the CT current within the
frequency band of 40-100 kHz are used for detecting series or parallel arcs under the
sampling frequency of 250 kHz. A detailed design of a DC series AFD based on a well-
known cost-effective TMS320F28335 digital signal processor can be found in [69]. The
hardware structure in [69] is illustrated in Figure 2.6.
CPU Timer 0
32 bit CPU
16 bit
Peripheral Bus
GPIO MUX SRAM JTAG SPI
DSP: Implement series arc fault identification technique
16 bit
ADC
8
th
Order Active
Band-pass Filter
Current Sensor
Trip Signal Output
Current Signal Input
Power SupplyDC 7 V~12 V
CT: PA3655NL
Op Amp: PA3655NL
SM73307 SM73201
TMS320F28335
L78L05A for 5 V Power Supply
LD1117A25 for 2.5 V Power Supply
LD1117A33 for 3.3 V Power Supply
Figure 2.6 An example of hardware structure of the DC series AFD [72]
28
Similarly, the 40-80 kHz and 30-100 kHz frequency band of CT current is used
under the sampling frequency of 200 kHz in [70] and [71], respectively. However, those
kinds of methods may fail when other electromagnetic interference shows up [37], [38].
In [72], instead of comparing the whole spectrum, it separates the frequency
spectrum into different sub-bands. In this way, the accuracy increases substantially. In
[73], Kanemaru et al. apply a sorting method to mitigate the impact from switching
noise associated with power electronic devices. The amplitude of each frequency bin
within 10-100 kHz frequency band in the frequency spectrum of the CT current (500
kHz sampling frequency) is sorted in the ascending order. It is known that the overall
noise intensity in signals increases when an arc occurs. Furthermore, it is reported in
some case studies that the amplitude of switching noise is bigger than that of arc noise
in certain frequency band [62]. Therefore, when switching noise and its harmonics have
a greater noise intensity than arc noise, the switching noise components concentrate
toward the higher sort number. By only considering those frequency contents in the
region of small sort number, the switching noise can be eliminated effectively. Under
the experimental condition, the proposed method dramatically increases the ratio of
signal intensity (the maximum integrated value of arc current before arc occurrence to
the maximum integrated value of arc current after arcoccurrence) from 2.2 to 907.
However, the proposed algorithm is not validated under other normal transient
conditions such as the sudden drop of current due to partial shading, inverter start-up,
etc. It requires further study to validate its effectiveness to withstand nuisance tripping.
The predefined threshold value is a limitation as it can be different for different
systems. In [74], a series arc fault detection method based on relative magnitude
comparison has been proposed, and adaptive threshold values for frequency domain are
calculated statistically. The sampling frequency is 250 kHz, the length of the window is
29
0.002 s, and the feature frequency band is 10-50 kHz. After using time domain analysis
to lock the potential arc instant point in the loop current (the candidate point), the
threshold value for each frequency component is determined by calculating the mean
and standard deviation of each frequency content in several consecutive windows
before the candidate points. When most of the magnitudes of frequency contents exceed
the threshold value for a certain number of consecutive periods, a series arc fault event
can be confirmed. This method can effectively differentiate arc state from many normal
operations such as inverter disconnected and load switching in DC microgrid, and the
detection speed is quite high (less than 16 ms). Most importantly, the dynamic threshold
value adjustment makes the algorithm flexible for different systems and operating
environment. However, this algorithm may not work properly in the scenarios that
inverter or converter start-up (this is often the case in PV systems). In this case, there
will be a steep change in the time domain and the magnitude of high-frequency contents
will increase in the frequency domain just like an arc fault, which may cause unwanted
trips.
Miao et al. proposed DC series arc fault detection in PV systems using pink noise
characteristics of the loop current with a sampling frequency of 50 kHz [75], [76].
Besides the algorithms they proposed, another important contribution of their work is
the use of tunnel magnetoresistance sensor to measure the PV loop current. Based on
their study, tunnel magnetoresistance sensor has smaller size, relatively high bandwidth
(~1 MHz), and relatively low cost (few USDs) compared to other types of current
sensors such as current probe [52], current transformer [77], [78], or magnetic sensors
such as Hall-effect sensors [79] used in some existing literatures.
A summary of FFT based detection methods is presented in Table 2.2.
30
Table 2.2 Summary of FFT based detection methods
Ref.
Verified by
experiment?
Microcontroller or
product
Sampling
Frequency
Test accuracy Detection time
[61] Yes RD-195 250 kHz Not mentionedrate (13.42%) with 0% rejection rate. The three
mentioned features would form a new mixed criterion and it significantly decreases
both malfunction rate (0%) and rejection rate (0.875%). The drawback of the proposed
method is the high sampling rate. However, mixing criteria greatly makes up the
inadequacies of the single criterion method.
In [83], WPD with coif4 has been applied to extract the energy of different sub-
bands under the sampling frequency of 100 kHz. After 6-level decomposition, both
series and parallel arc faults can be detected once the ratio of the sum of the square of
reconstruction coefficients of the high-frequency band (781.25 Hz-50 kHz) to that of
the lowest frequency band (0-781.25 Hz) exceeds the predefined threshold value for
several consecutive analysis periods.
In [52] and [85], 2-level WPD with db8 wavelet has been applied for
normalisation purpose under the sampling frequency of 200 kHz. The RMS value of
coefficients in the frequency band 0-25 kHz (including DC offset) is normalised by that
of in 25-50 kHz in a certain length of the time window (i.e. 10 ms). The peak RMS
value goes up from 6% to 15% after normalisation, which means the signature induced
by the series arc fault has been enlarged.
The performance of such algorithms significantly relies on the mother wavelet. Db
wavelets are proven to be suitable for arc fault diagnosis in resistive systems [86] but
exhibit detection limitations when applied to inverter-based systems. To address this
34
issue, Chen et al. choose rbio3.1-based mother wavelet to extract more distinguishable
arc features in various grid-tied PV systems [91].
The main difference between DWT and WPD is shown in Figure 2.8. Besides
decomposing the lower frequency band at each level, WPD can also decompose the
higher frequency band of the signal. Thus, WPD offers more information than DWT at
the expense of almost double computation burden.
A summary of DWT and WPD based detection methods is presented in Table 2.4.
Original signal
A1
h(n) 2
g(n) 2
H: Low-pass filtering and decimation
G: High-pass filtering and decimation
D1
A2 D2
A3 D3
Original signal
A1 D1
AA2 DA2 AD2 DD2
AAA3 AAD3 ADA3 DDA3 AAD3 DAD3 ADD3 DDD3
H G
H G
H G H G H G H G H G
H G H G
H G
Discrete Wavelet Transform Wavelet Packet Decomposition
π/8 π/4 3π/8 π/2 5π/8 3π/4 7π/8 π
Frequency
D1D2D3A3
π/8 π/4 3π/8 π/2 5π/8 3π/4 7π/8 π
Frequency
DDD3DDA3AAD3
Figure 2.8 Comparison of DWT and WPD analysis (3-level as an example)
35
Table 2.4 Summary of wavelet transform based detection methods
Ref. Technique
Verified by
experiment?
Microcontroller
or product
Sampling
Frequency
Test accuracy
Detection
time
[26] DWT Yes Not implemented 1 MHz Not mentioned
Not
mentioned
[88] DWT Yes Not implemented 200 kHz Not mentioned
Not
mentioned
[89] DWT Yes RD-195 200 kHz Not mentioned 0.1 s
[52]
[85]
WPD and
Statistics
Yes TMS320F28335 200 kHz
100% at voltage
below 300V and
current below 25A;
60% at 240V/25A;
40% at 300V/25A
~0.1 s
2.4.5. Statistical Analysis
There are various statistical features that can be used for DC arc faults detection,
such as mean, standard deviation, RMS value, entropy, and the extreme value of the
input signal.
The methods based on time domain analysis can be found in [33], [52], [58], [92]–
[98]. In [92], the change rate of the loop current in the time domain is proposed as an
indicator to determine the arc fault event. However, it is easily affected by random spike
disturbance. In [52] and [93], the difference between the maximum and minimum
current values over a certain length of the time window is defined as indicator. This
method is simple but quite effective to recognize arc fault, especially in the initial stage.
Although it can substantially eliminate the random disturbance from the noise, its
performance may be affected by other factors such as MPPT operation from inverter
when the irradiation level changes quickly. In [94] and [95], an outlier analysis based
detection method achieves 98% PV series arc fault detection rate at a false alarm rate of
36
0.01% in a single PV module through the simulation. Minimum covariance determinant
estimator is used to optimise the performance of the algorithm. In this outlier analysis,
the operating voltage and current of different PV modules at the same time instant are
fed into the minimum covariance determinant estimator. Then, the distance in I-V
characteristic curve between each PV module and the centre of the PV module’s
distribution is calculated by the estimator and used for detection.
In [96], a finite impulse response estimator is used to calculate the variance of the
input voltage signal under the sampling frequency of 50 kHz. The input signal is first
passed through a band-pass filter with cut-off frequency of 1 kHz and 7.5 kHz, and then
fed into the estimator, and the estimator then compares the current value with the
previous value. When the estimation is perfect, the variance is supposed to be 0. Then,
once the variance exceeds a pre-defined threshold value, an arc fault event can be
detected. The proposed algorithm is easier to be implemented and much cheaper, but at
the expense of lower accuracy compared to others.
In [97], multiple detection criteria are used for jointly detecting arc faults in PV
systems. In the time domain, statistical features including the mean and variance of the
loop current are calculated. In the frequency domain, the ratio of frequency contents of
the loop current in 1 Hz - 4 kHz to DC components and to AC components is calculated.
When at least one of time domain features and frequency domain feature exceed the
pre-defined threshold value respectively, an arc event can be determined. It should be
highlighted that the multiple detection criteria increase the accuracy of the algorithm
significantly.
Recently, a statistical detection approach based on arc current entropy has been
introduced in [58]. This method can effectively differentiate arc faults from the normal
events (non-arc states) such as MPPT operation and switch-on of the inverter. It
37
calculates the modified Tsallis entropy of loop current twice. Tsallis entropy can reveal
the degree of disorder and signal intrinsic behaviour:
𝑀 = ∑ 𝑝(𝑥𝑘)𝑞
𝐾
𝑘=1
= 1 − (𝑞 − 1)𝐸𝑇𝑠𝑎𝑙𝑙𝑖𝑠,𝑀 (2.15)
𝑝(𝑥 = 𝑥𝑘|𝑡 = 𝑘𝑇) = 𝑝(𝑥𝑘) ≝
∥ 𝑥𝑘 ∥2
∑ ∥ 𝑥𝑘 ∥2𝐾
𝑖=1
(2.16)
𝑀′ = ∑ 𝑝(𝑥𝑘,𝑀)
𝑞′
𝐾′
𝑘=1
= 1 − (𝑞′ − 1)𝐸𝑇𝑠𝑎𝑙𝑙𝑖𝑠,𝑀′ (2.17)
𝑝(𝑥 = 𝑥𝑘,𝑀|𝑡 = 𝑘𝑇) = 𝑝(𝑥𝑘,𝑀) ≝
∥ 𝑥𝑘,𝑀 ∥2
∑ ∥ 𝑥𝑘,𝑀 ∥2𝐾′
𝑖=1
(2.18)
where 𝑀 is the modified Tsallis entropy, 𝑞 and 𝑞′ > 0, 𝑥𝑘 denotes the samples of the
signal of interest, 𝐾 is the sliding window size in the first calculation, 𝐾’ is the sliding
window size in the second calculation, and in each sliding window both the sum of
𝑝(𝑥𝑘) and 𝑝(𝑥𝑘,𝑀) is 1. In the first stage of modified Tsallis entropy evaluation (𝑀)
where the captured current is the input signal, MPPT algorithm (due to fast-moving and
mechanical vibration induced by wind) may introduce variance in the value of 𝑀 with
certain patterns. Then, with proper value of 𝐾 and 𝐾’ , those disturbances can be
eliminated in the second stage of modified Tsallis entropy evaluation (𝑀′) where 𝑀 is
the input signal. The 𝑀′ will pass a first order infinite impulse response filter with 16
Hz cutoff frequency to remove the DC offset of 𝑀′ to get the detection feature 𝑀𝑧𝑜.
Then, the threshold value can be calculated based on standard deviation of 𝑀𝑧𝑜 and it
will be updated every 0.5 seconds to keep up withthe changing operating condition of
the system. The sampling frequency is only 10 kHz, and the computation load is just
24𝑁 flops compared to 5𝑁𝑙𝑜𝑔2𝑁 flops of FFT per sliding window, which is very cost-
effective. However, this method will be less effective in a noisier environment.
38
In [98], besides using the line current, Lu et al. also extracted useful information
from the PV supply voltage. Based on extensive experimental study on DC series arc
faults under different conditions in an experimental PV system, the change rate of the
line current, the change rate of the average line current, and the standard deviation of
the line current and AC components of the supply voltage are selected for detection.
However, there are more threshold parameters to set. It would take more efforts to fine-
tune those threshold parameters in a different PV system. Furthermore, although the
proposed algorithm has relatively low calculation complexity, it requires an additional
sensor (a current sensor and a voltage sensor) as compared to other methods (typically
only require a current sensor or a voltage sensor). The case study is carried out under
200 kHz sampling frequency. It is worthwhile to investigate the feasibility and
effectiveness of the proposed method with lower sampling frequency (e.g. 20 kHz).
Overall, the statistics-based fault diagnosis methods generally require less sampling
frequency and computation effort, but their performance will be severely affected by the
noise level of the surrounding environment. A summary of statistical analysis-based
detection methods is presented in Table 2.5.
Table 2.5 Summary of statistical analysis based detection methods
Ref.
Verified by
experiment?
Microcontroller or
product
Sampling
Frequency
Test accuracy Detection time
[58] Yes TMS320F28335 10 kHz
Pass all test
cases under
experimental
condition
~0.511 s
[94] Yes Not mentioned Not mentioned 98% Not mentioned
[95] Yes Not mentioned Not mentioned 98% Not mentioned
[96] Yes TMS320F2808 50 kHz Not mentionedof 99.4%.
SVM is believed to be better than ANN because of the following reasons. Firstly,
SVM does not suffer from the over-fitting problem in the training process. Secondly,
SVM can always find the global minimum during the training process, while ANN may
converge on local minimum; in other words, ANN often provides locally optimal
solutions and loses the big picture. Finally, SVM could get high accuracy by using
fewer data sets than ANN. The performance of SVM dramatically depends on the
quality of the training data. Both ANN and SVM are heuristic techniques, and thus their
reliability is difficult to be proved.
FL can also be applied for DC arc fault detection. In [112], Grichting et al.
proposed a PV series arc fault detection method using two-level FL and electrical
parameters calculated from the loop current and input voltage of the inverter. The rules
in the fuzzy system are designed according to fault pattern and mechanism. The input
signal will be firstly fuzzified as input of the fuzzy system, and then the arc faults and
normal operation can be classified with predefined rules. The performance of FL-based
methods highly depends on the expert knowledge. Incomplete or incorrect knowledge
could severely affect the detection accuracy. However, FL-based methods provide
economical solutions since they require relatively low computation efforts for real-time
implementation.
Other ML techniques, such as decision tree (DT) learning and learning classifier
system, have been used for fault detection (not including arc faults) and classification in
PV systems a few years ago [113], [114]. In [113], a semi-supervised graph-based
model has been applied to detect and classify line-to-line fault, open circuit fault, and
normal condition. In [114], a supervised DT-based model has been used for similar
function. Most recently, in [91], Chen et al. proposed a random forest (RF) based
43
protection strategy using rbio3.1-based DWT features for series arc fault detection in
PV systems. RF demonstrates better performance when the input size is greater as
compared to SVM. With the help of more generalised features extract by rbio3.1-based
wavelet compared to the features extract by db-based wavelet, the RF classifier can
achieve 90% accuracy level in different grid-connected PV systems even without
adjusting its parameters. The Hidden Markov model (HMM), which is superb to the
application related to non-stationary and highly transient signals, has been applied to
detect DC series arc faults [103]. A DC network model combined with series arc model
in [53] has been established in Matlab/Simulink with a dataset of different conditions
(series arc fault condition, nominal steady-state condition, and nominal transient
condition). The DWT (the mother wavelet is db2) level 1-3 approximation coefficient
and level 1-2 detail coefficient, and moving average of the loop current of different
conditions in a 50-ms window (6 features in total) are chosen as the features for series
arc fault detection under the sampling frequency of 20 kHz. Those features are fed into
the HMM for training purpose. The HMM will output the log-likelihood metric to
quantify the probability of the presence of an arc. With proper selection of the threshold
values for log-likelihood, a series arc fault event can be detected and discriminated from
other conditions. HMM is one of the probabilistic models, and there are several
limitations to probabilistic model-based methods. Firstly, the accuracy highly relies on
the quality of the training data, which can be collected from real systems or simulation.
Capturing data in real world is very costly, and it is difficult to cover all the conditions,
while data obtained from simulation highly depends on the accuracy of system
modelling. Secondly, compared to other methods, the probabilistic models-based
method needs more computation effort. For instance, although HMMs has the
advantage of the minimal computation load for calculating the log-likelihood, its order
44
of computation complexity is still very high: 𝒪 (𝑁2) compared to 𝒪 (𝑁𝑙𝑜𝑔𝑁) of FFT,
and 𝒪 (𝑁) of one dimensional db2 DWT.
Instead of using only one ML classifier to make the decision, the ensemble ML
combines multiple ML classifiers using three different ensemble learning techniques:
bagging, boosting, and stacking as illustrated in Figure 2.9.
ML Classifier 1
ML Classifier 2
ML Classifier n
Training dataset
Training dataset
Training dataset
Meta ML
Classifier
Decision
ML Classifier 1
ML Classifier 2
ML Classifier n
Decision
Weight 1
Weight 2
Weight n
Ensembled
Model
ML Classifier 1
ML Classifier 2
ML Classifier n
Sub Training dataset 1
Model Average
Decision
Bagging
Boosting
Stacking
Sub Training dataset 2
Sub Training dataset n
Sub Training dataset 1
Sub Training dataset 2
Sub Training dataset n
(a)
(b)
(c)
Figure 2.9 Ensemble ML techniques: (a) Bagging; (b) Boosting; (c) Stacking
45
In [115], Le et al. comprehensively investigated various ensemble ML learning
algorithms (using the load current as input signal) with different conventional ML
learning methods, such as DT, RF, k-Nearest Neighbours (kNN), Gaussian Naïve Bayes
(NB), and SVM. Based on extensive and rigorous analysis performed by the authors, an
input vector is formed by five time-domain features extracted from the load current
signals including the average, median, variance, RMS, and the difference between the
maximum and minimum value. It is found that the stacking ensemble algorithm formed
by kNN, RF, Gaussian NB, and logistic regression (LR) as a meta-classifier
demonstrates the best performance. The same authors further investigated the
effectiveness of semi-supervised ensemble ML learning (SVM or DT) under the
condition when there are a significant number of unlabelled samples and limited
labelled samples [116].
The advantages of ensemble ML are:
• it usually can improve the performance over any single ML model;
• it has less probability to overfit and is more stable.
The drawbacks of ensemble ML are:
• it does not perform well on simple dataset;
• it is usually computationally expensive, and therefore, it adds additional training
time and more memory constrains to the applications;
• it reduces the model interpretability.
A summary of ML based detection methods is presented in Table 2.7.
46
Table 2.7 Summary of ML detection methods
Ref. Methodology
Microcontroller
or product
Sampling
Frequency
Test accuracy
Detection
time
[91] DWT + RF Not implemented
Not
reported
> 90% 0.55 s
[103] DWT + HMM Not implemented 20 kHz
98.3% (simulation)
100% (experiment)
57.1 ms
[104] FFT + BPNN Not implemented 5 MHzThe computed
value represents the propagation of radio-frequency signal generated by arc, which will
be different after arc occurrence. In this application, it requires a very high sampling
frequency above 1 MHz.
In [118], Ahmadi et al. proposed a hybrid method that combines the cross-
correlation and signal-to-noise ratio to detect series arc fault in PV systems (sampling
frequency is 10 kHz). The measure DC terminal voltage 𝑉𝐷𝐶 is firstly normalised by the
47
open circuit voltage of the system 𝑉𝑂𝐶 to get the normalised DC terminal voltage 𝑉𝑎 =
𝑉𝐷𝐶/𝑉𝑂𝐶. After that, 𝑉𝑎 is filtered by a high-pass filter with cutoff frequency of 1 kHz to
get the noise components 𝑉𝑏. Both signals are framed with a frame size of 𝐿. Then, the
𝑛𝑡ℎ and 𝑛𝑡ℎ − 10 data-frame of 𝑉𝑏 are selected to calculate the cross-correlation
between |𝑉𝑏
𝑛| and |𝑉𝑏
𝑛−10| in order to find out the index of the most similar parts of
these two signals, 𝐿𝑎𝑔_𝑚𝑎𝑥. After that, 𝑉𝑏
𝑛 and 𝑉𝑏
𝑛−10 are resized using the following
equation:
{
𝑉𝑏
𝑛 = 𝑉𝑏(𝑛, |𝐿𝑎𝑔_ max |: 𝐿)
𝑉𝑏
𝑛−10 = 𝑉𝑏(𝑛 − 10, 1: 𝐿 − |𝐿𝑎𝑔_ max |)
if Lag_max ≥ 0
{
𝑉𝑏
𝑛 = 𝑉𝑏(𝑛, 1: 𝐿 − |𝐿𝑎𝑔_ max |)
𝑉𝑏
𝑛−10 = 𝑉𝑏(𝑛 − 10, |𝐿𝑎𝑔_ max |: 𝐿 )
if Lag_max2.9 Comparison of detection methods for DC arc fault detection
Detection
method
Domain
Frequency
resolution
Time
resolution
Sampling
Frequency
Computation
Effort
Popularity
Trend
FFT Frequency High N/A Medium/High Medium Stable
STFT Both Medium Medium Medium/High Medium/High Stable
DWT Both Medium Medium Medium/High Low/Medium Higher
WPD Both Medium Medium Medium/High Medium Stable
Statistics-
based
Both Depends Depends Depends Low Higher
Shallow
ANN
Both Depends Depends Depends Medium/High Lower
SVM Both Depends Depends Depends Medium/High Higher
HMM Both Depends Depends Depends High Higher
FL Both Depends Depends Depends Low/Medium Lower
Model-
based
Both Depends Depends Depends Medium/High Higher
EMR Time N/A High High Low Higher
SSTDR Time N/A High High Medium Stable
Cross
correlation
Time N/A High High Medium Stable
Kalman
filter
Time N/A Depends Low Low Higher
52
2.5. Discussion and Conclusion
Among those methods mentioned in the previous section, most of the fault
signatures are extracted from the 1-100 kHz frequency band, where it has the least
impact on the arc fault diagnosis from the majority of disturbance except
inverter/converter noise. Therefore, eliminating the influence of power electronics noise
can significantly increase the reliability and accuracy of the detection algorithms.
Recently, many detection methods with good switching noise immunity have been
introduced. For example, inverter noise is eliminated by calculating the Tsallis entropy
twice [58] or by the cross-correlation function [118], switching noise is suppressed by
performing a denoising algorithm [72], and normal transient events can be classified by
using intelligent detection systems [103]. More noise suppression techniques should be
developed and incorporated into detection algorithms to reduce the impact from
different noise sources to avoid unwanted tripping [37]. Up to now, common fault
signatures are still yet to be discovered, and more thorough DC arc fault characteristic
studies and development of good feature extraction techniques should be carried out.
Threshold comparison is the main approach to perform fault diagnosis, and the
threshold values are critical to most of the detection methods. The fixed threshold
values are often the main limitation for most of the methods because the behaviours of
arc faults may vary with different system structures and environment (i.e. different
background noise levels). Recently, adaptive threshold values have been used to
mitigate this problem, and all the threshold values are determined statistically [39], [58],
[74], [81]. Furthermore, most of the traditional algorithms rely on a single detection
criterion, which can be easily influenced by various disturbances. Therefore, many
recently proposed algorithms use several criteria extracted from hybrid domains or
multiple signals (e.g. source voltage and load current) to improve arc fault diagnosis
53
performance [39], [52], [74], [81], [90], [97], [98], [121].
With the advancement of computation technology and recent success progress,
ML-based techniques become increasingly popular in many fields including arc fault
diagnosis and prognosis [91], [103]–[116]. Based on the comprehensive reviews
presented above, DL has not yet been applied to this field, which remains a research gap
to be filled. One of prominent challenges for most of these techniques is shortage of
training dataset; this problem can be potentially mitigated by accurate system and arc
fault modelling [99], [100], [103], or data augmentation. However, existing literatures
mainly focus on simulation of the static V-I characteristics of arc faults. Although
several high-frequency arc fault models have been proposed recently, they can only
demonstrate similar patterns of real arc faults in the frequency domain. Therefore, more
precise high-frequency arc fault models or more advanced data augmentation
techniques need to be developed. There are other challenges when applying ML for
practical applications, and they will be discussed in detail in Section 4.6.
Besides the detection algorithms using electrical signals, high-frequency EMR
signals are also considered for DC arc fault detection in PV systems [124], [127], [128].
As the detection range is usually limited, this type of method might be a good candidate
for small household PV systems.
In conclusion, this chapter has comprehensively reviewed DC arc fault modelling
methods and the detection methods developed recently that can be used for PV systems.
The advantages and disadvantages of different detection methods have been discussed
and compared in detail. Better arc fault detection methods for PV systems with good
self-adjustability, robustness, and cost-effectiveness are still yet to be realised. More
efforts are still needed to improve the detection accuracy and reliability.
54
Some of the work described in this chapter has been published in:
Shibo Lu, B. T. Phung and Daming Zhang, “A Comprehensive Review on DC Arc
Faults and Their Diagnosis Methods in Photovoltaic Systems,” Renewable and
Sustainable Energy Reviews, vol. 89, pp.88-98, June 2018.
55
3. Characteristics Study on DC Series Arc Fault
3.1. Introduction
Voltage-current characteristics of arcing fault have been studied over past few
decades, and many empirical equations have been proposed to describe the static V-I
characteristics [43]. However, behaviours of the high frequency components of the arc
current and their dependency on many other factors are often ignored. Yao et al.
proposed a DWT based algorithm and it achieves 100% accuracy at lower current and
voltage level, while the accuracy decreases significantly at higher current and voltage
level [52]. In many applications the sum of power spectrum components within the 40-
100 kHz frequency band has been used as an indicator, which increases after arc fault
occurrence [61], [68], [69]. However, the algorithm may fail at higher current level: it
has been shown that, under fixed source voltage and gap distance, the probability
distribution of arcing signal (power density spectrum signal strength vs. probability) can
extend into the area of normal signal at higher load current level [129]. This can affect
the detection accuracy and reliability of the detection algorithm. Furthermore, Zhen et
al. found that fault indicator with satisfactory performance for 28V DC systems may
fail to work for 270V DC systems under the same load current [33]. Therefore, high
frequency noise characteristics and their determinants play an important role in the
development of arc fault detection algorithms. All these factors call for a more in-depth
understanding of DC arc fault phenomenon to achieve better detection. As arc fault
shows different behaviour at different working condition, new methods are also needed
to extract the consistent features.
The high-frequency variation is related to cathode spot activities [130]. The reason
is cathode spot motion can cause localised breakdown in the plasma sheath, which can
56
generate broadband radio frequency signals. Focusing on the radio-frequency current
with frequency up to tens of MHz requires high-performance sensors and measurement
equipment to precisely capture the high frequency components. It also requires a more
powerful and costly microcontroller to process large amount of data under the higher
sampling rate. Furthermore, the behaviours of radio-frequency current, especially above
100 kHz, can be affected by various factors in PV systems [131]. Therefore, this thesis
focuses on the frequency contents below 100 kHz.
3.2. Experimental Setup
The power source used is LAB/HP 15600 and it can supply DC voltage and current
up to 500 volts and 30 amps, respectively. An adjustable resistiveload bank with up to
85.5 Ω is used. Since this chapter mainly focuses on exploring arc features, power
electronics load is not considered. Additional inductive components and capacitive
components are not included in the experiment circuit in this session.
The DC information of the loop current and arc voltage is captured by PROSyS-
CP35 differential current probe and SI-9000 differential voltage probe, respectively. In
addition, the high frequency variation of the loop current is captured by a Pearson-4688
current transformer with 600 Hz cutoff frequency (-3dB attenuation). The data
acquisition (DAQ) system consists of a National Instruments PXIe-1073 and a PXIe-
4300. The PXIe-4300 can stream the raw data to a personal computer at sampling
frequency of 200 kHz with 16-bit analog-to-digital converters. The experimental setup
for this chapter is shown in Figure 3.1.
57
Resistor
load bank
PC
NI PXIe
4300 DAQ
Arc
generator
DC Power
source
Voltage
sensor
Current
sensors
AC V
AA
CT Probe+
-
PC
DAQ
Arc generator
DC power supply
Load
1
Figure 3.1 Experimental setup for characteristics study of DC series arc fault
The DC arc fault generation method has been introduced in PV DC arc fault circuit
protection outline, UL-1699B, in 2011 [10], [14]. It uses “steel wool” method to initiate
an arc fault: the moving electrode is firstly inserted into a polymer sheath tube
containing some steel wool, and then adjusted to a given distance from a stationary
electrode. When voltage is applied across the electrodes, the steel wools burn and melt
quickly, which ignite an arc in the electrode gap. However, arc generation with the
“steel wool” method takes more time and brings more complexity with less repeatability.
More importantly, it is difficult to facilitate low power arc fault such as 100 W arc
58
faults [34]. Furthermore, as the arc length and arc gap of parallel arc in the real world is
much longer than series arc, the steel-wool method is a better way to initiate and create
sustainable parallel arc. The higher power and longer air gap of parallel arc easily
destroys the steel wool and thus it would vaporise quickly for establishment of arc. The
“pull-apart” method is better for creating a series arc, whereas the “steel wool” method
with tube is easier to develop a parallel arc.
The first edition of the formal UL-1699B Standard became available from August
2018 [15]. Steel wools and polymer sheath tube are removed from the guideline of arc
fault detection tests. Therefore, as recommended in [14], [15], an arc generator without
steel wools and polymer sheath tube is used in this chapter to reduce the complexity and
increase the repeatability of the arc generation experiment. The diagram of the arc
generator is illustrated in Figure 3.2. Copper is the material used for electrodes in the
test as it is commonly used for connecting in DC systems. Copper electrodes of 6.35
mm (1/4 inch) diameter with flat tip are used. The experimental conditions are listed in
Table 3.1.
Figure 3.2 Diagram of arc fault generator in UL-1699B Standard
The same data collection procedures are followed for all experimental conditions:
59
1. Wait until electrodes and load resistor cool down to the ambient temperature;
2. Polish electrodes with sandpaper, mount electrodes onto the arc generator, and
connect them together;
3. Adjust the load resistance;
4. Start the DC power source and adjust the voltage and current to the desired level;
5. Start the data acquisition (DAQ) system in the personal computer;
6. Initiate an arc by separating the copper electrodes to the pre-defined gap distance;
7. Shut down the DC power source to extinguish the arc and stop the measurement.
All the datasets are saved locally. Then, the on-site data sets are visualised,
analysed, and processed by Matlab.
Table 3.1 Experimental conditions for characteristics study of DC series arc fault
Case Condition (air gap width = 1, 2, 3 mm)
Fixed DC voltage (200V) 4.1, 7.9, 11, 13.6 A
Fixed DC current (6.5A) 87, 111, 134, 158 V
3.3. Static Characteristics
3.3.1. V-I Characteristics
Arcing is a very complex physical phenomenon. Nowadays, most arc studies are
based on observation of experiments and analysis of acquired data, and scholars mainly
use the V-I curve to characterise this phenomenon [43], [131], [132]. The arc can be
then treated as a non-linear resistance. The experimental data is fitted to the Nottingham
arc model, which is suitable for gap distance from 1-10 mm, as shown in (3.1):
𝑉𝑎𝑟𝑐 = 𝐴 +
𝐵
𝐼𝑎𝑟𝑐
𝑛 (3.1)
60
where A, B, and n are arc constants depending on the gap distance. The result is plotted
in Figure 3.3, and it is found consistent with the results in [43], [132].
Figure 3.3 V-I characteristic under different gap distance
It should be noted that the arc gap distance is not equal to the actual arc length, and
the additional impedance injected into the circuit is contributed by the arc length;
however, it could be considered they are the same when the air gap distance is small,
which is the case in this research. When the gap distance is fixed, in the lower arc
current region, the smaller the arc current, the larger the arc voltage, where the arc
power (𝑃𝑎𝑟𝑐 = 𝑉𝑎𝑟𝑐𝐼𝑎𝑟𝑐) tends to remain the same; whilst in the higher arc current region,
with the increasing arc current, the arc voltage remains approximately unchanged.
When the arc current is fixed, the larger the gap distance is, the larger the arc voltage is,
and thus the larger the arc resistance is. If assuming the electrical conductivity and the
effective cross-section area of the arc column remain the same in the quasi-stationary
state, the arc resistance is proportional to distance according to (3.2):
61
𝑅𝑎𝑟𝑐 =
1
𝜎
𝑙
𝑆
(3.2)
where σ is electrical conductivity of the arc column, l is arc length (it equals gap
distance for short arc), and S is the effective cross-section area of the arc column.
3.3.2. Stable Operating Point
In a typical DC circuit, the following equation can be obtained:
𝑉𝑠𝑜𝑢𝑟𝑐𝑒 = 𝐿𝑙𝑜𝑎𝑑
𝑑𝐼
𝑑𝑡
+ 𝑅𝑙𝑜𝑎𝑑𝐼 + 𝑉𝑎𝑟𝑐 (3.3)
where 𝑉𝑠𝑜𝑢𝑟𝑐𝑒 is DC source voltage, 𝑅𝑙𝑜𝑎𝑑 and 𝐿𝑙𝑜𝑎𝑑 are circuit parameters, and 𝑉𝑎𝑟𝑐 is
the arc voltage, which equals to 0 before arc fault instant. (3.3) can be rewritten as:
𝑑𝐼
𝑑𝑡
=
1
𝐿𝑙𝑜𝑎𝑑
[(𝑉𝑠𝑜𝑢𝑟𝑐𝑒 − 𝑅𝑙𝑜𝑎𝑑𝐼) − 𝑉𝑎𝑟𝑐] (3.4)
When an arc fault occurs, dI/dt is negative, and then the current starts to decrease.
When the gap distance reaches the final value, a stable arcing is formed where 𝑉𝑠𝑜𝑢𝑟𝑐𝑒 −
𝑅𝑙𝑜𝑎𝑑𝐼 = 𝑉𝑎𝑟𝑐 and dI/dt = 0 as shown in Figure 3.4. The stable operating point is point
A instead of point B. For example, consider the case when the current experiences a
small disturbance at point A, which causes a decrease. Because dI/dt > 0 and where
𝑉𝑠𝑜𝑢𝑟𝑐𝑒 − 𝑅𝑙𝑜𝑎𝑑𝐼 > 𝑉𝑎𝑟𝑐, the input power is greater than the power dissipated by arc.
Therefore, the arc tends to burn more steadily, and the loop current increases back to
point A. Similarly, when there is a small disturbance causing an increase in the current
at point A, dI/dt 0
dI/dt > 0
Vsource
Varc2
Varc1
Vsource-IRload
A
B
Figure 3.4 The condition for a stable arcing point
The typical waveforms of a free burning DC series arc fault in DC networks are
shown in Figure 3.5. The arcvoltage remains approximately the same and then keeps
decreasing after a few seconds. This is because the electrical conductivity of the electric
arc increases with temperature once the temperature is high enough to cause thermal
ionisation. Specifically, due to increasing temperature, the degree of dissociation and
thermal ionisation of the inter-electrode medium and metal vapor keep increasing,
leading to increasing number of charge carriers and electrical conductivity. Therefore, it
effectively decreases the arc resistance and the arc voltage [133]. This suggests that
under the same load condition (same source voltage and load current) and gap distance,
the quasi-stationary V-I characteristic of DC arc may vary with temperature as the arc
keeps burning continuously. Therefore, the stable operating point of the arc would
change when the temperature is high enough to cause thermal ionisation.
63
(a)
(b)
Figure 3.5 Waveforms of a typical DC series arc fault: (a) arc current; (b) arc voltage
3.3.3. Load Current Effect and Source Voltage Effect
Because arc is a very complex physical phenomenon, it is difficult to analyse its
behaviour point by point because it cannot be exactly treated as a constant resistor.
Therefore, the average arc resistance is calculated by averaging the arc resistance of
each sampled data point in the first 2 seconds duration after the moving electrode has
reached the predefined distance point. In Figure 3.6 (a), the average arc resistance and
the amount of change in average resistance decrease as the load current level increases
when the source voltage is fixed. On the other hand, it is fair to say that the DC voltage
level has fewer impacts on the arc resistance compared to the load current level as
shown in Figure 3.6 (b). The average arc resistance decreases slightly with increasing
source voltage level as expected. Note that this condition is only satisfied when
generating sustainable arcing. Also, arc resistance increases with increasing gap
64
distance, which agrees with (3.2).
(b)
(a)
Figure 3.6 Average arc resistance under (a) fixed source voltage; (b) fixed load current
3.4. High-frequency Variation in Arc Current
In this part, Fourier transform and wavelet packet entropy are applied to analyse arc
fault current signals under different conditions.
3.4.1. Wavelet Packet Entropy
WPD is a multi-resolution analysis that can divide a signal into different frequency
65
bands as illustrated in Figure 2.8. Mathematically, discrete WPD can be achieved
through a series of convolutions with a pair of high-pass and low-pass filters, 𝑔(𝑛) and
ℎ(𝑛) [134]. The representations of these filters are shown as follows:
𝑔(𝑛) =
1
√2
〈𝜓(𝑡), 𝜑(2𝑡 − 𝑛)〉 (3.5)
ℎ(𝑛) =
1
√2
〈𝜑(𝑡), 𝜑(2𝑡 − 𝑛)〉 (3.6)
𝑔(𝑛) = (−1)𝑛ℎ(1 − 𝑛) (3.7)
where 〈∙,∙〉 is the inner produce operator, 𝑡 and 𝑛 are variables, and 𝜑(∙) and 𝜓(∙)
denotes the scale function and its corresponding wavelet function, respectively.
For a discrete time-series signal, the wavelet coefficients for different frequency
bands at different decomposition levels can be calculated iteratively using the following
functions:
𝑊2𝑚−1
𝑗+1
(𝑘) = ∑ ℎ(𝑛 − 2𝑘)𝑊𝑚
𝑗
(𝑛)
𝑘
(3.8)
𝑊2𝑚
𝑗+1
(𝑘) = ∑ 𝑔(𝑛 − 2𝑘)𝑊𝑚
𝑗
(𝑛)
𝑘
(3.9)
where 𝑊1
0 denotes the original discrete time-series signal with length of L, 𝑊𝑚
𝑗
(𝑛)
represents the wavelet coefficients at 𝑗𝑡ℎ level and 𝑚𝑡ℎ frequency band, 𝑛 =
1, 2, … , 𝐿 2𝑗⁄ , 𝑊2𝑚−1
𝑗+1
(𝑘) is the wavelet coefficients at (𝑗 + 1)𝑡ℎ level and (2𝑚 − 1)𝑡ℎ
frequency band, 𝑊2𝑚
𝑗+1
(𝑘) is the wavelet coefficients at (𝑗 + 1)𝑡ℎ level and 2𝑚𝑡ℎ
frequency band, and 𝑘 = 1, 2, … , 𝐿 2𝑗+1⁄ . It should also be noted that 𝑚 = 1, 2, … , 2𝑗
at 𝑗𝑡ℎ level.
Entropy can be used to measure the degree of disorder. Then, the wavelet packet
66
entropy is achieved by combining WPD and entropy theory together. Specifically, after
WPD analysis, the set of wavelet coefficients at 𝑗𝑡ℎ level and 𝑚𝑡ℎ frequency band, 𝑊𝑚
𝑗
,
with values of {𝑤𝑚1
𝑗
, 𝑤𝑚2
𝑗
, ⋯ , 𝑤𝑚𝑛,
𝑗
⋯ , 𝑤𝑚𝑁,
𝑗
} can be obtained, where 𝑚 = 1, 2, ⋯ , 2𝑗
and, 𝑁 denotes the number of coefficients. Then, the entropy 𝐻𝑚
𝑗
can be calculated as
follows:
𝐻𝑚
𝑗
= − ∑ 𝑝(𝑤𝑚𝑛
𝑗
)𝑙𝑜𝑔𝑝(𝑤𝑚𝑛
𝑗
)
𝑁
𝑛=1
(3.10)
𝑝(𝑤𝑚𝑛
𝑗
) =
|𝑤𝑚𝑛
𝑗
|2
∑ |𝑤𝑚𝑛
𝑗
|2𝑁
𝑛=1
(3.11)
where 𝑝(𝑤𝑚𝑛
𝑗
) is the probability of 𝑤𝑚𝑛
𝑗
, and ∑ 𝑝(𝑤𝑚𝑛
𝑗
) = 1𝑁
𝑛=1 .
In the case study of this chapter, the sampling frequency is 200 kHz, the WPD level
is 𝑗 = 3, and the analysed window length is 𝑇 = 40 ms. Accordingly, there are 8 sub-
bands with 𝑁 = 1000 wavelet coefficients for each band. Furthermore, db9 is chosen as
the mother wavelet because of its excellent performance in arc fault signal analysis in
DC systems with resistive loads [26].
3.4.2. Effect of Arc Phase
Based on experimental observations, after initiation of arc fault, the arc behaviour
can be characterised as three different arc phases under the experimental conditions.
The waveforms of DC series arc fault generated at 11A/200V with a gap distance of 1
mm is used as an example as shown in Figure 3.7:
• Arc phase 1: Drastic variation can be observed in arc current signal.
• Arc phase 2: Less variation can be observed in arc current signal. Also, some
67
spikes can be observed.
• Arc phase 3: The arc tends to be steady. There are less variation and spikes
observed in arc current signal.
Drastic variation Less variation,
some spikes
Less variation,
less spikes
Arc inception Arc quenching
Power off
Figure 3.7 High frequency variation of DC series arc fault at 11A/200V
This phenomenon is concluded based on experimental observations, and similar
phenomenon was observed in [77]. More comprehensive analysis related to its physics
is not carried out since it is not within the scope of this thesis.
Because of natural convection caused by air heated by ionisation and upward
convective flow caused by lower density of hot air, the convection force makes arc fault
have a bow shape [43]. In addition, the arc root tends to be stable at the edge of the
electrodes. The frequency spectrum is shown in Figure 3.8. The arc noise intensity in
the current spectrum increases significantly after its occurrence. Furthermore, as
described above, the spectrum level decreases when the phase shifts with time. The arc
current spectrum also exhibits pink noise characteristics.
68
Figure 3.8 Average frequency spectrum of the 2-second data at non-arc state and arc
state at difference arc phase (FFT analysis window is 0.2 seconds)
The wavelet packet entropy calculation is shown in Figure 3.9. Right after the
instant of arc initiation, the entropy for each band decreases significantly towards zero.
This is mainly caused by the large spike introduced in arc current because of the arc
ignition. After the initiation, the entropy level of bands 2-4 and 6-8 shows some
increase with little fluctuations, while that of band 1 and band 5 remains approximately
the same but with larger fluctuations. The entropy level for each band at non-arc state
and different arc phases are shown in Table 3.2. Unlike the frequency spectrum, which
changes significantly with different arc phases, the entropy of bands 2-4 and 6-8
remains approximately unchanged.
69
Figure 3.9 Wavelet-packet entropy of DC series arc fault at 11A/200V
Table 3.2 Wavelet-packet entropy level for different arc phases (11A/200V)
Non-arc Phase1 Phase2 Phase3
Band8 (87.5-100 kHz) 5.805 8.814 8.780 8.749
Band7 (75-87.5 kHz) 6.329 8.824 8.843 8.840
Band6 (62.5-75 kHz) 5.896 8.620 8.644 8.622
Band5 (50-62.5 kHz) 7.476 7.097 8.203 8.359
Band4 (37.5-50 kHz) 6.069 8.797 8.807 8.834
Band3 (25-37.5 kHz) 6.120I also extend my sincere gratitude to my secondary supervisor, Dr. Daming Zhang,
for providing valuable feedback and timely discussions to my research work.
In addition, I would like to acknowledge Mr. Zhenyu Liu, technical support staff,
for his detailed instructions and fruitful discussions of the experiments throughout my
research study. Besides, I would like to thank all my friends and colleagues in my life,
in particular Tharmakulasingam Sirojan, Hua Chai, Miao Li, Dr. Muhammad Tariq
Nazir all from the UNSW Energy Systems research group, and Rui Ma from University
of Miami, USA. Thank you all for the friendship, help, and support.
I would also like to thank the Tyree Foundation and UNSW for their financial
support.
I am so grateful for the unconditional love and unwavering support from my family
members throughout my life, in particular my parents, Xuejun Lu and Hui Yin, without
whom I would never have enjoyed so many opportunities. Last but certainly not the
least, I would like to specially thank my wife, Zhi Chen. Zhi, I could not be able to
finish this work without your love and support. Thank you for stepping into my life and
being with me.
V
Abstract
Grid integration of renewable sources including solar energy is growing faster than
ever before. Nowadays, solar power development is increasing throughout the world,
and solar photovoltaic (PV) systems play an important role to support the main loads
and micro-grids. However, one needs to consider the long-term performance of PV
components. Their deterioration can be caused by various factors such as ageing,
weathering, the higher DC operating voltage level, improper installation, inadequate
maintenance, etc. The consequence is a growing potential of electrical arcing incidents
especially the series arc fault in PV systems. Without timely detection and interruption,
such dangerous events can cause catastrophic fires, posing a severe threat to human
safety and properties.
In this thesis, a comprehensive review of DC arc fault and their diagnosis methods
in PV systems is presented. Experimental study of DC series arc fault characteristics is
carried out. The feasibility of applying deep learning (DL) in series arc fault detection in
PV systems is systematically investigated. Specifically, convolutional neural networks
(CNN) are successfully applied and demonstrate superior diagnosis performance over
conventional machine learning algorithms and other popular DL algorithms. For cost-
effective real-time deployment, a lightweight CNN structure is designed to achieve a
good balance between model complexity and detection accuracy. Moreover, novel
frameworks, including domain adaptation and deep convolutional generative adversarial
network (DA-DCGAN) and lightweight transfer convolutional neural network with
adversarial data augmentation (LTCNN-ADA), are proposed. They aim to address the
challenges when applying DL to practical applications, including lack of fault data from
VI
the field, data inconsistency between laboratory and field, and limited computation
resources in edge devices. The proposed methods are validated through comprehensive
offline analysis using pre-recorded data. In addition, the trained DL classification
models are deployed in an embedded system and tested in single-phase and three-phase
PV systems in real-time under different test conditions. Both offline and online
experimental results show that the proposed methods can accurately and reliably detect
series arc fault in PV systems.
VII
INCLUSION OF PUBLICATIONS STATEMENT
UNSW is supportive of candidates publishing their research results during their candidature as
detailed in the UNSW Thesis Examination Procedure.
Publications can be used in their thesis in lieu of a Chapter if:
• The student contributed greater than 50% of the content in the publication and is the
“primary author”, i.e. the student was responsible primarily for the planning, execution and
preparation of the work for publication
• The student has approval to include the publication in their thesis in lieu of a Chapter from
their supervisor and Postgraduate Coordinator.
• The publication is not subject to any obligations or contractual agreements with a third party
that would constrain its inclusion in the thesis
Please indicate whether this thesis contains published material or not.
☐ This thesis contains no publications, either published or submitted for publication
☒
Some of the work described in this thesis has been published and it has been documented
in the relevant Chapters with acknowledgement
☐
This thesis has publications (either published or submitted for publication) incorporated
into it in lieu of a chapter and the details are presented below
CANDIDATE’S DECLARATION
I declare that:
• I have complied with the Thesis Examination Procedure
• where I have used a publication in lieu of a Chapter, the listed publication(s) below
meet(s) the requirements to be included in the thesis.
Name
Shibo LU
Signature Date (dd/mm/yy)
VIII
Table of contents
ACKNOWLEDGEMENT ................................................................................................. IV
ABSTRACT .......................................................................................................................... V
TABLE OF CONTENTS ................................................................................................ VIII
LIST OF ACRONYMS .................................................................................................... XV
LIST OF FIGURES ....................................................................................................... XVII
LIST OF TABLES ....................................................................................................... XXIII
1. INTRODUCTION ........................................................................................................ 1
1.1. BACKGROUND AND RESEARCH MOTIVATION.......................................................... 1
1.2. SUMMARY OF RESEARCH CONTRIBUTIONS ............................................................. 5
1.3. THESIS ORGANISATION ........................................................................................... 7
1.4. LIST OF PUBLICATIONS ............................................................................................ 8
2. LITERATURE REVIEW .......................................................................................... 11
2.1. INTRODUCTION...................................................................................................... 11
2.2. DC ARC IN PHOTOVOLTAIC SYSTEMS ................................................................... 11
2.2.1. Photovoltaic Systems Structure and Arc Hazards ............................................ 11
2.2.2. Challenges to Detect DC Arc Faults ................................................................ 15
2.3. DC ARC MODELS .................................................................................................. 17
2.3.1. Physics-based Arc Model ................................................................................. 19
IX
2.3.2. V-I Characteristic-based Arc Model ................................................................ 20
2.3.2.1. Nottingham Arc Model ........................................................................ 20
2.3.2.2. Hall, Myer, and Viicheck Arc Model .................................................. 20
2.3.2.3. Stokes and Oppenlander Arc Model .................................................... 21
2.3.2.4. Paukert Arc Model ............................................................................... 21
2.3.2.5. Modified Paukert Arc Model ............................................................... 22
2.3.3. Heuristic8.840 8.856 8.851
Band2 (12.5-25 kHz) 6.534 8.772 8.817 8.834
Band1 (0-12.5 kHz) 8.961 8.283 8.700 8.743
From the experiment, one can see that the higher the current, the faster the
temperature rise in the arc column, and then the sooner arc phase 3 is reached as shown
70
in Figure 3.10. It also reveals the temperature dependency and preference. At higher
temperature, the degree of thermal ionisation of the metal vapor and mixed gases is
higher. For example, copper vapor starts ionisation above 4000 K [133]. The additional
electrons enable the continued process of arc burning.
Arc phase 1 Arc phase 2
Arc phase 1 Arc phase 2 Arc phase 3
Arc phase 1 Arc phase 3AP 2
Figure 3.10 DC current dependency for different arc phases
3.4.3. Load Current Effect
Based on experiment, the noise level of the low-current arc fault shows its
dependency on circuit parameters, such as source voltage and load current. As shown in
Figure 3.11, under fixed voltage and fixed gap distance, there are less high frequency
variation at higher load current level. In addition, for the same V-I characteristic curve,
when the load current level is higher, the stable operating point is much far away from
the interrupted point (the circuit characteristic curve is tangent to the V-I characteristic
71
curve) as shown in Figure 3.12. It gives the stable operating point more margins, which
accordingly increase the stability of the arc and thus less variation in the arc current
signal. The wavelet packet entropy calculation for different load current under fixed
source voltage is shown in Figure 3.13, Table 3.3, and Table 3.4. C1-C8 represents
different series arc fault cases generated at different load current and source voltage
levels (arc gap distance is 1 mm): C1: 4.1A/200V; C2: 7.9A/200V; C3: 11A/200V; C4:
13.7A/200V; C5: 6.5A/87V; C6: 6.5A/111V; C7: 6.5A/134V; C8: 6.5A/158V. The
results are similar as discussed in the previous section. More importantly, the entropy
level of arc-state of bands 2-4 and 6-8 remains approximately unchanged when the load
current changes.
Figure 3.11 DC load current dependent arc spectrogram under fixed source voltage
72
Voltage
Current
Fix gap distance
Interrupted point Stable operating point
Figure 3.12 V-I curve for fixed source voltage
Table 3.3 Wavelet-packet entropy level (Non-arc state) for different load current and
source voltage
Non-arc state
C1 C2 C3 C4 C5 C6 C7 C8
Band8 (87.5-100 kHz) 6.140 5.813 5.774 5.838 6.208 6.171 6.767 6.347
Band7 (75-87.5 kHz) 7.108 6.386 6.499 6.325 6.883 6.908 7.320 6.742
Band6 (62.5-75 kHz) 6.189 5.780 5.807 5.657 6.522 6.476 7.030 6.576
Band5 (50-62.5 kHz) 9.369 7.926 7.453 7.502 6.8985 7.449 7.845 8.734
Band4 (37.5-50 kHz) 6.095 6.111 6.223 5.988 6.562 6.373 7.061 6.949
Band3 (25-37.5 kHz) 6.661 6.108 6.313 6.174 6.456 6.305 7.048 6.638
Band2 (12.5-25 kHz) 6.739 6.607 6.997 6.612 7.071 6.951 7.683 6.776
Band1 (0-12.5 kHz) 9.130 9.164 8.855 8.921 9.525 9.280 9.168 9.379
Table 3.4 Wavelet-packet entropy level (Arc state) for different load current and source
voltage
Arc state
C1 C2 C3 C4 C5 C6 C7 C8
Band8 (87.5-100 kHz) 8.692 8.695 8.745 8.845 8.718 8.539 8.539 8.541
Band7 (75-87.5 kHz) 8.678 8.745 8.789 8.835 8.709 8.543 8.626 8.637
Band6 (62.5-75 kHz) 8.712 8.606 8.493 8.791 8.703 8.427 8.418 8.305
Band5 (50-62.5 kHz) 8.706 7.501 7.316 8.241 8.388 7.454 8.095 7.634
Band4 (37.5-50 kHz) 8.643 8.650 8.773 8.743 8.692 8.604 8.671 8.603
Band3 (25-37.5 kHz) 8.677 8.707 8.708 8.822 8.707 8.549 8.689 8.674
Band2 (12.5-25 kHz) 8.861 8.789 8.787 8.685 8.710 8.650 8.730 8.640
Band1 (0-12.5 kHz) 8.562 8.725 8.682 8.238 8.669 8.679 8.403 8.489
73
Figure 3.13 Wavelet-packet entropy under fixed DC source voltage
3.4.4. Source Voltage Effect
Under fixed load current and gap distance, there are less high frequency variation at
higher source voltage level as shown in Figure 3.14-15. Similarly, the higher source
voltage gives arcing more margins to the interrupted point. Therefore, the arc is more
stable and generates fewer variation at higher source voltage level. On the other hand,
as shown in Figure 3.16, Table 3.3, and Table 3.4, the entropy level of arc-state of band
2-4 and band 6-8 remains approximately same with the changing DC source voltage.
74
Figure 3.14 DC source voltage dependent arc spectrogram under fixed load current
Voltage
Current
Fix gap distance
Interrupted point
Stable operating point
Figure 3.15 V-I curve for fixed load current
75
Figure 3.16 Wavelet-packet entropy under fixed load current
3.4.5. Gap Distance Effect
Similarly, as shown in Figure 3.4, with gap distance increasing, the V-I
characteristic changes and finally two curves intercept at the interrupted point. When
the gap distance exceeds the critical gap distance, the arc becomes unstable and then
extinguished. Therefore, it is expected that the high-frequency variation will be more
significant at longer gap distance as the stable operating point is closer to the interrupted
point. As shown in Figure 3.17, the gap distance dependency has been revealed and the
results are the same as expected. Although the gap distance is different, the entropy
level remains roughly the same as shown in Table 3.5. The case of series arc faults
generated at 6.5A/158V are used as examples.
76
Figure 3.17 Average frequency spectrum of the first 2 seconds data after the gap
distance reached the desired value (FFT analysis window is 0.2 seconds)
Table 3.5 Wavelet-packet entropy level for different gap distance (6.5A/158V)
1mm gap 2mm gap 3mm gap
Band8 (87.5-100 kHz) 8.851 8.547 8.717
Band7 (75-87.5 kHz) 8.637 8.529 8.685
Band6 (62.5-75 kHz) 8.404 8.347 8.660
Band5 (50-62.5 kHz) 7.634 7.418 7.664
Band4 (37.5-50 kHz) 8.603 8.658 8.622
Band3 (25-37.5 kHz) 8.674 8.506 8.687
Band2 (12.5-25 kHz) 8.630 8.687 8.627
Band1 (0-12.5 kHz) 8.489 8.505 8.412
3.5. Discussion and Conclusion
Based on experimental analysis, the arc current and its spectrum show dependency
on the source voltage, load current, and gap distance. The arc fault tends to produce less
arcing noise when the stable operating point is further away from the interrupted point:
the high frequency variation induced by the arcing increase with decreasing source
77
voltage, decreasing load current, and increasing gap distance between the two
electrodes. Therefore, when testing the effectiveness of AFD and AFCI in DC networks,
the arc fault should be generated at high voltage and current level at a small gap
distance to obtain the lowest level of arc noise, i.e. the worst-case scenario. In addition,
the minimum threshold value for traditional detection methods can be determined under
the worst-case scenario. In UL-1699B Outline, it does not properly take these important
parameters into consideration for series arc fault testing. In the latest UL-1699B
Standard, however, these parameters have been properly considered. For example, the
gap distance range is reduced from 1.6-6.4mm to 0.8-2.5mm.
For low-current arc fault, the time to reach the final arc phase is much longer
compared to the required detection time listed in UL-1699B, which is up to 2 seconds in
[14] and 2.5 seconds in [15]. Therefore, during the development stage of arc fault
detection, developers could focus on the characteristics in the initial stage of arc faults.
The results of this comprehensive experimental study also provide meaningful and
useful information for testing the effectiveness and robustness of AFD and AFCI.
Furthermore, by applying wavelet packet entropy analysis to all cases in the
experiment in this chapter, it is found that this method can extract a consistent entropy
pattern of series arc fault that is less sensitive to changes in source voltage, load current,gap distance, and arc phases. In addition, there are clear changes of entropy before and
after arc fault occurrence. Such entropy features of arc fault current could be adopted
for its more effective detection.
78
Some of the work described in this chapter has been published in:
1. Shibo Lu, B. T. Phung, Daming Zhang, and Hua Chai, “An Experimental Study of
Low-Current DC Series Arc Faults for Condition Monitoring Purpose,” International
Conference and Exhibition on Electricity and Distribution (CIRED), Madrid, Spain, 3-6
June 2019.
2. Shibo Lu, B. T. Phung, and Daming Zhang, “Study on DC Series Arc Fault in
Photovoltaic systems for Condition Monitoring Purpose,” Australasian Universities
Power Engineering Conference (AUPEC), Melbourne, Australia, Nov. 2017.
79
4. DC Series Arc Fault Detection in PV systems using Deep
Learning
4.1. Introduction
Nowadays, most of the traditional industries are transformed by digital Internet
technology. This paradigm shift requires data-driven intelligent decision-making to
enable automations with reduced operational risks. As an emerging field in industrial
applications and an effective solution for intelligent decision-making, ML is
increasingly being used and has demonstrated promising results in DC series arc fault
detection as reviewed in Section 2.4.7. The general procedures for ML-based DC arc
fault detection methods in PV systems are illustrated in Figure 4.1. To date, majority of
studies mainly focus on the conventional ML methods, while DL techniques are not
well investigated.
This chapter systematically investigates the feasibility of applying DL in series arc
fault detection in PV systems. A lightweight CNN structure is designed to improve
detection accuracy and reduce the computation burden, which makes it more suitable
for Internet-of-Things and edge computing applications. An experimental setup is
established to collect series arc fault data under different operating conditions. A
comparative study among different popular ML methods is then performed using the
same dataset to demonstrate the effectiveness and superior performance of the proposed
method.
Finally, barriers obstructing intelligent fault diagnostics from being applied in
practice are identified. Potential solutions in applying ML, especially DL, for practical
PV series arc fault detection are presented.
80
Feature extractions by experts
• Time domain features
• Frequency domain features
• Time-frequency domain features
Feature extractions by experts (optional)
Conventional machine learning methods Deep learning methods
Dimension reduction & feature selection
• Principal component analysis
• Sparse representation
• Random forest
• ...
Classification
• Shallow artificial neural network
• Support vector machine
• Fuzzy inference system
• ...
Classification
• Deep artificial neural network
• Convolutional neural network
• Deep belief network
• Stacked autoencoder
• Recurrent neural network
• ...
Automatic feature learning by deep learning models
Time (s)
C
T
C
u
rr
e
nt
(
A
)
Arc inception
Data acquisition:
• Current (I)
• Voltage (V)
• V-I Curve
• Radio-frequency
• ...
Measurement system/Edge device
...
Voltage (V)
C
u
rr
en
t
(A
) Normal
MPPNew MPP
Arc
inception
Figure 4.1 General procedures for ML-based DC arc fault detection methods
4.2. Classical Machine Learning
4.2.1. Artificial Neural Network
ANN is inspired by biological neural networks and has been used for different fault
diagnostics for decades. For example, multilayer perceptron (MLP) is a class of
feedforward neural networks consisting of at least three fully connected layers (one
input and one output with one or more hidden layers) of non-linearly activating nodes.
Given a dataset, {𝒙𝑖, 𝒚𝑖}𝑖=1
𝑛 , of n samples, the corresponding label vector, and a k-layer
81
MLP (the number of hidden layers is 𝑘 − 2), the mathematical representation of the
output for 𝑗𝑡ℎ layer, 𝑓𝑗(𝒙𝑖
𝑗
), is shown as follows:
𝑓𝑗(𝒙𝑖
𝑗
) = 𝜎𝑗(𝒘𝑗𝑇
𝒙𝑖
𝑗
+ 𝑏𝑗) (4.1)
where 𝜎𝑗 is the activation function, 𝒘𝑗 ∈ 𝒘 is the weight matrix, 𝑏𝑗 ∈ 𝒃 is the bias
coefficient, 𝒙𝑖
𝑗
is the input of 𝑗𝑡ℎ layer, and 𝑗 = 2, … , 𝑘 . Note that the number of
neurons for hidden layers are flexible, while that of the input layer and output layer are
identical to the dimension of input data and label vector, respectively. The
backpropagation algorithm is widely used for training feedforward neural networks for
supervised learning. Therefore, any ANN trained using backpropagation algorithm is
also known as the BPNN. In general, BPNN is the most commonly selected type of
ANN for different applications, and the general structure (MLP as an example) is shown
in Figure 4.2.
Hidden layers
Input layer Output layer
Feature 1
Feature 2
Feature d
Type 1
Type l
d is the dimension of the input array and l is the number of categories for output
x1
x2
xd
y1
yl
Figure 4.2 Structure of a back propagation neural network (MLP)
82
Given a portion of the dataset {𝒙𝑖, 𝒚𝑖}𝑖=1
𝑚 ∈ {𝒙𝑖, 𝒚𝑖}𝑖=1
𝑛 for training, the main
objective of BPNN is to minimise the loss, L, between the predicted output and the label
vector. There are different kinds of loss functions that can be used for training. For
example, when the mean-square error, 𝐿𝑀𝑆𝐸 , is used, the objective function to be
minimised of the BPNN is:
min
𝒘,𝑏
𝐿𝑀𝑆𝐸(𝒘, 𝒃) =
1
𝑚
∑[𝑓𝑘(𝒙𝑖
𝑘) − 𝒚𝑖]
2
𝑚
𝑖=1
(4.2)
After feeding the training dataset, the calculated error will be propagated backward
all the way to the input layer to update the parameters of the BPNN using the gradient
descent with the learning rate of 𝛿 shown as follows:
𝒘 ← 𝒘 − 𝛿
𝜕𝐿(𝒘, 𝒃)
𝜕𝒘
，𝒃 ← 𝒃 − 𝛿
𝜕𝐿(𝒘, 𝒃)
𝜕𝒃
(4.3)
The principle of the backpropagation algorithm and gradient descent is shown in
Figure 4.3 and Figure 4.4, respectively.
Calculate L
მ f j-1
მ L მ L
=
Feed forward
Backpropagation
f j-1
მ f j
მ f j
მ f j-1
f j
მ w j
მ L მ L
=
მ f j
მ f j
მ w j
მ b j
მ L მ L
=
მ f j
მ f j
მ b j
მ f j-1
მ L
მ w j
მ L
მ b j
მ L
მ f j
მ L
b j
w j
Figure 4.3 Illustration of the backpropagation algorithm
83
Optimal learning rate Large learning rate Small learning rate
w
L(w)
Initial weight
Lmin(w)
Gradient
w
L(w)
Initial
weight
w
L(w)
Initial weight
Figure 4.4 Gradient descent with different learning rate
Gradient descent with a small learning rate requires significant number of
iterations before converging to the minimum. On the other hand, if the learning rate is
too large, gradient descent can overshoot the minimum, which can cause non-
convergence or even divergence. Therefore, the learning rate needs to be carefully
selected through trial and error in order to achieve a desirable performance.
ANNs offer several advantages, such as easily solving multi-classification
problems and able to manage large amount of data and input variables. However, ANNs
have low interpretability because of their black-box nature. In addition, for relatively
large ANN models, they cannot always find the global optimum which makes them
prone to be overfitted.
4.2.2. Support Vector Machine
SVM is a supervised learning algorithm, which is widely employed in many fault
diagnostics.
Take the non-kernel SVM (or linear SVM) as an example: consider a binary
classification problem, given a dataset, {𝒙𝑖, 𝑦𝑖}𝑖=1
𝑛 , of n samples and the corresponding
84
label, where 𝒙𝑖 ∈ 𝒙 and 𝑦𝑖 ∈ {−1, 1}, a hyperplane 𝑓(𝒙) = 0 is chosen to separate the
data into a positive and a negative group, shown as follows:
𝒘𝑓(𝒙𝑖) = 𝒘𝑇𝒙𝑖 + 𝑏 = 0 (4.4)
where 𝒘 and b are the parameters to determine the hyperplane. Then, a positive
boundary and a negative boundary can be determined bythe closest point from the
hyperplane in each group. Any point above the positive boundary is of one class with
label 1, while any point below the negative boundary is of one class with label -1. After
rescaling the distance of the closest point from the hyperplane in each group to be 1, the
chosen hyperplane is subject to the following condition to separate the dataset:
𝑦𝑖𝑓(𝒙i) = 𝑦𝑖(𝒘𝑇𝒙𝑖 + 𝑏) ≥ 1, 𝑖 = 1, 2, … , 𝑛 (4.5)
As shown in Figure 4.5, in order to find the optimal hyperplane to achieve perfect
separation, the margin 𝛾 = 2/‖𝒘‖ (the distance between the positive boundary and
negative boundary) is expected to be maximised, where ‖∙‖ is the norm operator. As a
result, the following optimisation problem is formulated for non-kernel SVM [135]:
min
𝒘,𝑏
𝐶𝑜𝑠𝑡(𝒘, 𝑏) =
‖𝒘‖2
2
𝑠. 𝑡. 𝑦𝑖(𝒘𝑇𝒙i + 𝑏) ≥ 1, 𝑖 = 1, 2, … , 𝑛
(4.6)
Note that the square term in (4.6) is for computation optimisation purpose. For a
dataset that is not linearly separable, a hyperplane with a soft margin can be found by
adding regularisation terms in the cost function in (4.6).
SVMs have a solid mathematical foundation in statistical learning theory. The
solution for a typical SVM is a convex optimisation problem, which can always find the
global optimum. Therefore, SVMs are less prone to overfitting problems compared to
85
ANNs. However, the main disadvantage is that SVMs are sensitive to the optimal
choice of kernel. Additionally, they are computationally inefficient with a large dataset
and does not work well when the number of input features is greater than the number of
samples. Consequently, it requires more works in feature engineering for SVMs.
Best margin
Optimal
hyperplane
Class A Class B
Feature 1
F
ea
tu
re
2
Figure 4.5 A simple linear SVM for classification
4.2.3. Decision Tree and Random Forest
DT is a supervised learning algorithm widely used in classification problems,
which breaks the input space into regions with separate parameters for each region. It is
a decision-making process by establishing the relationship between the attributes and
the classes using a flowchart-like structure. A simple example of series arc fault
detection using DT is visualised in Figure 4.6.
DTs are usually unstable, and a tiny change in the data can cause a significant
change in the optimal structure of DT. On the other hand, DTs are simple to interpret
and easy to achieve good performance with simple input data, which are suitable for
some specific applications.
86
RF is one particular tree-based model that mitigates overfitting in DT by
integrating multiple DT-based classifiers. It is found to demonstrate good property of
generalisation in series arc fault detection in different grid-connected PV systems
recently [91].
Feature 1 satisfies
condition 1?
Feature 2 satisfies
condition 2?
Feature 3 satisfies
condition 3?
Non-Arc
Series Arc Fault
Yes No
Yes
Yes No
No
A decision tree
Non-Arc
Non-Arc
Figure 4.6 A simple illustration of decision tree for series arc fault detection
4.2.4. k-Nearest Neighbours
Consider a training dataset, {𝒙𝑖, 𝑦𝑖}𝑖=1
𝑛 , of n samples with corresponding label, and
an unlabelled testing dataset {𝒙𝑗}𝑗=1
𝑚 . The kNN classification algorithm firstly computes
distance 𝑑(𝒙𝑖 , 𝒙𝑗) to every training example 𝒙𝑖 . One of the most common distances,
Euclidean distance, is given in (4.7):
𝑑(𝒙𝑖, 𝒙𝑗) = ‖𝒙𝑖 − 𝒙𝑗‖
2
(4.7)
Then, it selects k closest instances {𝒙𝑖1, … , 𝒙𝑖𝑘} and their labels {𝑦𝑖1, … , 𝑦𝑖𝑘} .
Finally, it can output the most frequent label in {𝑦𝑖1, … , 𝑦𝑖𝑘} as 𝑦𝑝𝑟𝑒𝑑𝑖𝑐𝑡 for the input
sample. For the kNN regression algorithm, the last step is modified: it computes the
mean of {𝑦𝑖1, … , 𝑦𝑖𝑘} as follows:
87
𝑦𝑝𝑟𝑒𝑑𝑖𝑐𝑡 =
1
𝑘
∑ 𝑦𝑖𝑛
𝑘
𝑛=1
(4.8)
There are two main advantages for kNN: firstly, it is a very simple ML model and
easy to implement; secondly, it has fewer hyperparameters to tune. However, the
computation cost increases significantly when the sample size is large, and the value of
k is difficult to select. Additionally, Goodfellow et al. report that the output of kNN on
small training sets will essentially be random [136]. All those negative factors make the
kNN method less popular.
4.2.5. Others
Besides the abovementioned algorithms, there are other conventional ML methods
such as NB, FL, kNN, HMM, etc.
4.3. Deep Learning
Conventional ML classifiers with shallow structures require a powerful feature
extractor that solves the selectivity-invariance dilemma [137], and they cannot be
continually improved by increasing the size of the training data. DL is the most active
development in the area of ML nowadays. The DL architecture has more similar
working principle as biological nervous system of human. DL enables systems to
discover complex features through learning many simple features, and the simple
features are represented in terms of each other. To be specific, in the hidden layers, the
output from the previous layer will be the input of the next layer, and finally many
layers are cascaded. In this way, it discards the useless information and disturbance
(disentangle the factor of variation) through deep structure, which leads to a better
performance of the system.
88
DL is getting more attention because it requires less need for feature engineering
and can achieve higher performance with the help of big data, and revolutions in
algorithms and hardware.
For example, the batch normalisation (BN) layer is developed to mitigate the
problems that arise because of poor coefficient initialisation and it helps gradient flow
in deeper models during training of the deep neural network using mini batches [138].
The mathematical representation of BN is as follows:
𝑓𝐵𝑁(𝒙) = 𝛾 (
𝒙 − 𝜇𝑏
√𝜎𝑏
2 + 𝜀
) + 𝛽 (4.9)
where x, µb, σb
2, γ, β, and ε denote input, mini-batch mean, mini-batch variance, scale
factor, offset and stability parameter. Initially, the BN layer transforms the input to a
mapping with zero mean and unit variance. After that, it shifts and scales that mapping
with the learnable parameters, γ and β, to make it optimal for the successive layers in
the deep neural network. In this way, the effects of internal covariate shift can be
mitigated, and the training of the neural network can be stabilised and accelerated.
Another good example of major algorithmic change is usage of ReLU activation
function. Conventionally, sigmoid and tanh activation functions are widely used in
shallow ANN. They tend to cause severe vanishing gradient problem because of
saturation. Once saturation occurs, it becomes challenging for the learning algorithm to
continue adjusting the weights to improve the performance of the model. The ReLU
activation function overcomes the vanishing gradient problem because it is nearly linear
and does not cause saturation. The models are also easier to optimise since their
behaviour is closer to linear when using ReLU [136]. The mathematical representation
of different activation functions used in this research are summarised in Table 4.1.
89
Table 4.1 Activation functions used in this thesis
Name Representation Plot
Sigmoid 𝑓(𝑥) =
1
1 + 𝑒−𝑥
Tanh 𝑓(𝑥) =
𝑒𝑥 − 𝑒−𝑥
𝑒𝑥 + 𝑒−𝑥
ReLU 𝑓(𝑥) = {
0, 𝑥 ≤ 0
𝑥, 𝑥 > 0
Leaky
ReLU
𝑓(𝑥) = {
𝛼𝑥, 𝑥 ≤ 0
𝑥, 𝑥 > 0
𝛼 = 0.2 as an example
Softmax
𝑓(𝒙) =
𝑒𝑥𝑖
∑ 𝑒𝑗𝐽
𝑗=1
for 𝑖 = 1, … , 𝐽
N/A
90
4.3.1. Deep Fully-Connected Neural Network
The simplest DL structure is the deep MLP, which consists of multiple fully-
connected layers for hierarchical feature extraction and a decision layer for
classification (generally the softmax layer). The mathematical representations are
similar to shallow MLP introducedin Section 4.2.1, while the number of hidden layers
is generally greater than three.
4.3.2. Autoencoder
A simple auto-encoder consists of two parts: an encoder and a decoder. Given the
input dataset {𝒙𝑖, 𝑦𝑖}𝑖=1
𝑛 with n samples, the encoding process and decoding process can
be represented as follows:
𝑓𝑒(𝒙𝑖) = 𝒉𝒊 = 𝜎𝑒(𝒘𝒆
𝑇𝒙𝑖 + 𝒃𝑒) (4.10)
𝑓𝑑(𝒉𝒊) = 𝒙𝑖
′ = 𝜎𝑑(𝒘𝒅
𝑇𝒉𝒊 + 𝒃𝑑) (4.11)
where the subscripts e and d represent the encoder and decoder; 𝒉𝒊 denotes the
features extracted by encoder; 𝒙𝑖
′ is the reconstructed sample by decoder; 𝜃 = {𝒘, 𝒃}
and 𝜎 are the network parameters and activation function, respectively. The
optimisation objective of the auto-encoder is to minimise the reconstruction error of the
input samples. Therefore, it can be optimised using the following cost function:
min
𝜃𝑒,𝜃𝑑
𝐶𝑜𝑠𝑡(𝜃𝑒 , 𝜃𝑑) =
1
𝑛
∑‖𝒙 − 𝒙𝑖
′‖2
𝑛
𝑖=1
(4.12)
There are typically two ways to construct an auto-encoder with deep structure. The
first way is achieved by stacking multiple auto-encoders to form a stacked autoencoder
(SAE). The output from the encoder part of the first auto-encoder is used as the input to
the second auto-encoder. Then, greedy layer-wise pre-training is performed to establish
91
the SAE. The typical process to construct a SAE is visualised in Figure 4.7.
Add classification layer and fine-tuning
Type 1
Type l
xi
y1
yl
hi,1 xi
~
Encoder
Decoder
Train the first AE Train the second AE
hi,1 hi,2
xi
hi,1
hi,1
~
hi,2
Train the last AE
hi,n-1 hi,n hi,n-1
~
hi,n-1 hi,n
Stack all the encoder parts of AEs
Figure 4.7 Diagram of a SAE for series arc fault detection
After establishing the SAE with pre-training parameters, a decision layer (i.e. a
softmax layer) is connected, and labels can be used to fine-tune the whole structure in a
supervised way to achieve classification. The other method to construct an auto-encoder
with deep structure is replacing a single layer with multiple layers in both encoder and
decoder parts.
92
4.3.3. Convolutional Neural Network
As shown in Figure 4.8, a CNN model typically consists of a convolution
operators-based feature extractor, fully-connected layers for higher level reasoning, and
a classification layer. The computation complexity of convolution layers is less
compared to the fully-connected layers in terms of the required matrix multiplication
operations because of the configuration method, where each neuron in a convolution
layer is only connected to a small set of neurons in the following layer. Also, such a
configuration makes CNNs excellent in extracting regional characteristics of the input
sample. For an input matrix 𝒙𝑖
𝑗−1
with P channels from the previous layer, K filters
with size of 𝐻𝑓 × 𝐿𝑓, and step size 𝑠=1, the convolution operation of the 𝑘𝑡ℎ filter at the
𝑗𝑡ℎ layer can be represented as follows:
(𝒙𝑖
𝑗
)
ℎ𝑜,𝑙𝑜,𝑘
= 𝜎(𝒙𝑖
𝑗−1
∗ 𝒘𝑘
𝑗
+ 𝑏𝑘
𝑗
)
= 𝜎(∑ ∑ ∑ (𝒙𝑖
𝑗−1
) 𝑠×ℎ𝑜+ℎ𝑓,𝑠×𝑙𝑜+𝑙𝑓,𝑝 × (𝒘𝑘
𝑗
)ℎ𝑓,𝑙𝑓,𝑝 + 𝑏𝑘
𝑗
𝐿𝑓−1
𝑙𝑓=0
𝐻𝑓−1
ℎ𝑓=0
𝑃−1
𝑝=0
)
(4.13)
where 𝒙𝑖
𝑗
denotes the output feature map at the 𝑗𝑡ℎ layer with the size of 𝐻𝑜 × 𝐿𝑜 ×
𝐾 ; ℎ𝑜 = {1, … 𝐻𝑜} , 𝑙𝑖𝑝 = {1, … 𝐿𝑜} , 𝑘 = {1, … 𝐾} represent the row, column, depth
index of the output feature map, respectively; 𝒘𝑘
𝑗
and 𝑏𝑘
𝑗
are the weight matrix and bias
coefficient of 𝑘𝑡ℎ filter in 𝑗𝑡ℎ layer, respectively; 𝜎 denotes the activation function,
which is typically ReLU for DNN since it can mitigate the gradient vanishing problem
[139]. After each CNN layer, a pooling layer is usually used to achieve dimension
reduction. Maxpooling layer is the most common pooling layer, which can be
represented as follows:
(𝒙𝑖
𝑗
)
ℎ𝑜,𝑙𝑜,𝑘
= 𝑚𝑎𝑥((𝒙𝑖
𝑗−1
)
ℎ𝑜:ℎ𝑜+𝐻MaxP−1,𝑙𝑜:𝑙𝑜+𝐿MaxP−1,𝑘
)) (4.14)
93
where the size of the max operator is 𝐻MaxP × 𝐿MaxP , the operation step size is 1, the
size of the output feature is 𝐻𝑜 × 𝐿𝑜 × 𝐾. Then, the hierarchical features can be found
by stacking several CNN layers and pooling layers. Next, the last pooling layer is
flattened to 1D vector and connected to fully-connected layers for further reasoning.
Finally, a classification layer (i.e. softmax layer) is connected to map the input sample
into the target class.
Convolutional
kernel
Convolutional layer Pooling layer
Convolutional layer
Pooling layer
Pooling kernel
Fully connected
layers
D
e
c
isio
n
la
y
er
Feature extraction Reasoning
& Decision
Figure 4.8 General structure for a convolutional neural network
4.3.4. Recurrent Neural Network
The links of a recurrent neural network (RNN) between the nodes form a directed
graph along a temporal sequence, which makes RNN capable of exploring the dynamic
behaviour of time-series data. Long short-term memory (LSTM) model, consisting of
many LSTM blocks, is one of the best-performing and most popular RNNs. The most
beneficial part of LSTM is that it introduces an internal recurrence besides the outer
recurrence in traditional RNNs. With such a modification, the LSTM model is easier to
train since the gradient can flow for long durations [136]. An LSTM block contains
94
several units to control the flow of information, including a state unit 𝒔𝑖
𝑡, a forget gate
unit 𝒇𝑖
𝑡, an external input gate unit 𝒈𝑖
𝑡, an output gate unit 𝒒𝑖
𝑡 for a time step t, layer
index i, and the current input vector 𝒙𝑖
𝑡. Then, the mathematical representation of an
LSTM block can be formulated as follows:
𝒇𝑖
𝑡 = 𝜎𝑠𝑚(𝒖𝑖
𝑓𝑇
𝒙𝑖
𝑡 + 𝒘𝑖
𝑓𝑇
𝒉𝑖
𝑡−1 + 𝒃𝑖
𝑓
) (4.15)
𝒔𝑖
𝑡 = 𝒇𝑖
𝑡𝒔𝑖
𝑡−1 + 𝒈𝑖
𝑡𝜎𝑡𝑎𝑛ℎ(𝒖𝑖
𝑠𝑇
𝒙𝑖
𝑡 + 𝒘𝑖
𝑠𝑇
𝒉𝑖
𝑡−1 + 𝒃𝑖
𝑠) (4.16)
𝒈𝑖
𝑡 = 𝜎𝑠𝑚(𝒖𝑖
𝑔𝑇
𝒙𝑖
𝑡 + 𝒘𝑖
𝑔𝑇
𝒉𝑖
𝑡−1 + 𝒃𝑖
𝑔
) (4.17)
𝒉𝑖
𝑡 = 𝜎𝑡𝑎𝑛ℎ(𝒔𝑖
𝑡)𝒒𝑖
𝑡 (4.18)
𝒒𝑖
𝑡 = 𝜎𝑠𝑚(𝒖𝑖
𝑞𝑇
𝒙𝑖
𝑡 + 𝒘𝑖
𝑞𝑇
𝒉𝑖
𝑡−1 + 𝒃𝑖
𝑞) (4.19)
where 𝒉𝒊, 𝒘𝒊, 𝒖𝒊, 𝒃𝒊 are the current hidden layer vector (it is also the output for the
current LSTM unit), recurrent weights, input weights, and biases for the 𝑖th layer of
LSTM block, respectively; the superscripts s, f, g, and, q indicate the correspondence of
the parameters to different units; 𝜎𝑠𝑚 and 𝜎𝑡𝑎𝑛ℎ are the sigmoid and tanh activation
function, respectively. An illustration of an LSTM unit is shown in Figure 4.9. After the
information over time is obtained by the LSTM model (consisting of several LSTM
modules), fully-connected layers are used to reason the output of the LSTM model in
many-to-one mode or in many-to-many mode. Finally, a softmax layer is connected at
the end to achieve classification.
There is an extended LSTM model called bidirectional LSTM (Bi-LSTM).
Basically, Bi-LSTM combines two independent LSTMs together, which allows the
whole network to have both backward and forward information about the input
sequence at every time step [140]. This type of model generally gives better
performance when the context of the input is needed.
95
× +
+
σ σ σ
tanh
tanh
st-1,i
st,i
ht-1,i
wf,i uf,i
bf,i +
wg,i ug,i
bg,i +
ws,i us,i
bs,i +
wq,i uq,i
bq,i
×
×
LSTM block for time step t and layer i
ht,i
yt,i
xt,i
L
S
T
M
a
t
ti
m
e
st
ep
t
-1
L
S
T
M
a
t
ti
m
e
st
ep
t
+
1
LSTM at time step t and layer i+1
Figure 4.9 Diagram of a LSTM block in a LSTM model
4.4. Experimental Setup
Arc fault experiments are performed using the experimental system shown in
Figure 4.10. Different from the experimental setup established in Chapter 3, the DC
power source is replaced by a Magna power TSD-1000-20/415 programmable DC
power supply (PV emulator) and the resistive load is replaced by a Sunny Boy 1.5
single-phase inverter. Thus, a 1.5-kW emulator-based grid-tied PV system is
established. Thepurpose of using a PV emulator is to achieve different operating
condition (e.g. different irradiance and PV cell temperature) in a more controllable
environment [141]. Similarly, sensors comprising a PROSys-CP35 differential current
probe, an SI-9000 differential voltage probe, and a Pearson-4688 CT are used to sense
the loop current, arc voltage, and high-frequency information of the loop current,
respectively. All the electrical signals are streamed to the personal computer using a
DAQ system which consists of a National Instruments PXIe-1073 and a PXIe-4300
module with 200-kHz sampling rate and 16-bit analog-to-digital conversion resolution.
96
Flat-tip cylindrical copper rod electrodes with 6.35 mm (1/4 in.) diameter are used in
the arc generator, and a Nema-42 servo motor with a controller based on an Arduino
Uno and A4988 motor driver are employed to accurately control the electrode
separation speed and distance.
Figure 4.10 Schematic diagram of the experimental setup
The series arc faults are generated between the positive terminal of the PV emulator
and the positive terminal of the solar inverter with a separation rate of 5 mm/s and a
separation distance of 0.5 mm. To obtain series arc fault signals at different conditions,
the same experiment is repeated with different combination of irradiance level from
400 𝑊/𝑚2 to 1000 𝑊/𝑚2 and temperature from 0 𝐶° to 45 𝐶°.
97
4.5. Proposed Deep Learning based Series Arc Fault Detection Method using
Convolutional Neural Network
Traditional fault diagnosis techniques need to extract the important statistical
parameters from the raw data manually to achieve fault identification. This process is
known as feature extraction. Researchers have proposed several feature extraction
techniques to identify series arc faults based on domain-specific knowledge as reviewed
in Chapter 2. In the conventional ML based series arc faults detection approaches, the
extracted features are fed into ML classifiers for decision making as illustrated in Figure
4.1. The main drawback of these approaches is that the decision accuracy heavily
depends on the input features [91]. In addition to that, the domain experts need to invest
considerable time on feature learning for each fault condition. Even so, there is no
guarantee that the extracted features can fully represent the unique characteristics of
series arc faults under different operating conditions.
DL enables automated hierarchical feature learning from data and it proves its
success in industrial applications from various domains such as image processing [142],
speech recognition [143], time series sensor data analytics [144] etc. Recently,
researchers have started to apply DL based data analysis techniques to enhance the
operations in industrial applications. Zeng et al. [145] proposed a two-stream multi-rate
RNN for pedestrian identification with the aid of two CNNs to extract the spatial and
motion features from raw video frames. Sun et al. [146] came up with a DL based
sparse deep stacking network that can eliminate the overfitting of DL models in motor
fault diagnosis applications. Guo et al. [147] exploited the CNNs combined with
continuous wavelet transformation to detect earth faults via transforming the fault
current signal into time-frequency grey-scale images. Jiang et al. [148] proposed
multiscale CNN based DL architecture to automate the fault feature extraction from raw
98
vibration signals of wind turbine gearbox. Zhao et al. [149] developed deep residual
networks with dynamically weighted wavelet coefficients to enhance the fault diagnosis
of planetary gearboxes. Zhao et al. [150] combined time domain hand-engineered
features with bidirectional gated recurrent unit networks to monitor machine health
conditions. Different DL architectures and input features are used in the aforementioned
researches based on their application context. The choice of architecture and the feature
set are key requirements to design a viable solution.
Inspired by these successes of DL, particularly the successes of CNN, CNN based
intelligent detection is proposed for series arc fault detection in PV systems without any
hand-crafted features. The proposed method requires little prior knowledge on signal
characteristics. Some application-specific requirements and constraints are addressed in
the rest of thesis.
4.5.1. Dataset Preparation
Different types of disturbance exist in different frequency ranges in PV systems as
summarised in Section 2.2.2. Above 100-kHz, the signals are mainly affected by radio-
frequency noise [57], [131]. In the lower frequency range, switching noise is one of the
main concerns that can potentially cause nuisance tripping of AFD/AFCI. For example,
solar inverters can generate switching noise and its harmonics from 1 kHz to above 100
kHz [57], [58], [108], [151], [152]. Based on extensive experimental studies carried out
by Sandia National Labs, and considering the pink noise nature of arc faults (the power
spectral density is inversely proportional to the frequency of the signal) and severe
radio-frequency interference in the higher frequency range, the frequency band of 0.1-
100 kHz is recommended for detection purpose even though the signal is inevitably
affected by the switching noise [57].
99
Although a higher sampling rate offers more information about the signals, it
significantly increases the computation load of the algorithm, which is not practical for
real-time implementation. Many recent studies have also obtained good results using
frequencies less than 10 kHz (at sampling frequency of 20 kHz) for DC arc fault
detection [58], [103], [152]. Therefore, the PV loop current signals from CT are down-
sampled to 20 kHz and used for series arc fault detection. Using such a CT can
eliminate low-frequency external disturbances (typically less than 1 kHz) caused by
sand, mechanical vibrations, etc. [58]. Further, a case study is presented in Chapter 5,
which confirms that the choice of using 20 kHz as sampling frequency is appropriate.
All the raw CT time-series signals are divided into windowed samples with 𝑙2
points and scaled into [-1, 1] for standardisation. After that, two-dimensional (2D)
arrangement is performed to arrange each windowed sample to a 2D matrix with size of
𝑙 × 𝑙 [153], [154]. The standardisation and 2D arrangement are achieved using (4.20)
and (4.21).
𝓍(𝑖, 𝑗) =
𝓍𝑟𝑎𝑤(𝑙(𝑗 − 1) + 𝑖) − min (𝓍𝑟𝑎𝑤)
max(𝓍𝑟𝑎𝑤) − min (𝓍𝑟𝑎𝑤)
(4.20)
𝓍(𝑖, 𝑗) = 2 × 𝓍(𝑖, 𝑗) − 1 (4.21)
where 𝓍𝑟𝑎𝑤 is the raw time-series segment and 𝓍 is the standardised 2D sample.
Similarly, with increasing window sizes, more useful information can be extracted from
the input signal by the algorithms at the expense of increasing computation complexity
and detection latency. The value of l, which represents the height/width of the squared
2D matrix, is 20 based on considerations of the work from other researches [39], [58],
and a case study with different value of l is presented in Section 5.4. Ultimately, 20,000
normal samples and 20,000 arcing samples are extracted to form the dataset for the case
study. Each sample consists 400 data points, corresponding to 20 ms data duration. Note
100
that these samples are prepared by partitioning many time-series signals of different
cases into several 20 ms-duration sections without any overlapping.
4.5.2. Hyperparameters Setting and Offline Validation Results
There are several well-known CNN architectures such as, LeNet 5 [155], AlexNet
[156], VGG 16 [157], etc. [158]. LeNet 5 is the first popular CNN architecture proposed
in 1998, and it is a relatively small CNN compared to today’s standards. Because of its
limited capability of representation learning (aka feature learning), it is less popularin
some complex tasks such as large image recognition nowadays. However, it is quite
suitable for Internet-of-Things applications because of its relatively small number of
parameters. AlexNet is similar to LeNet 5 but with a deeper structure. It is the first CNN
architecture that stack convolution layers directly on top of each other. After that, CNNs
start to become deeper in order to improve the performance. Both AlexNet and VGG 16
demonstrate good results for many complex problems such as image classification and
localisation. However, the main drawback associated with deep CNN, such as AlexNet
and VGG 16, is the use of millions of parameters (e.g. 60 million and 138 million
parameters in the original applications of AlexNet and VGG 16, respectively).
Therefore, it is computationally expensive and difficult to be deployed in real-time in a
resource-constraint edge device. The architectures of LeNet 5, AlexNet, and VGG 16
are illustrated in Figure 4.11.
Firstly, it is worthwhile to investigate the effectiveness of these deep CNNs in
series arc fault detection in PV systems. The original input size for LeNet 5, AlexNet,
and VGG 16 are 32 × 32 × 1, 227 × 227 × 3, and 224 × 224 × 3, respectively, while
their minimum required input size are 9 × 9 × 1 , 37 × 37 × 3 , and 32 × 32 × 3 ,
respectively.
101
Input
224×224×3
3×3 Conv.
224×224×64
3×3 Conv.
224×224×64
2×2 MaxPool
112×112×64
3×3 Conv.
112×112×128
3×3 Conv.
112×112×128
2×2 MaxPool
56×56×128
3×3 Conv.
56×56×256
3×3 Conv.
56×56×256
2×2 MaxPool
28×28×256
3×3 Conv.
56×56×256
3×3 Conv.
28×28×512
3×3 Conv.
28×28×512
2×2 MaxPool
14×14×512
3×3 Conv.
28×28×512
3×3 Conv.
14×14×512
3×3 Conv.
14×14×512
2×2 MaxPool
7×7×512
3×3 Conv.
14×14×512
Flatten
Dense
25088
Dense
4096
Dense
4096
1000 (Class)
Original input 64 Filters 128 Filters 256 Filters 256 Filters 512 Filters
Input
32×32×1
5×5 Conv.
28×28×6
2×2 AveragePool
14×14×6
Original input
Zero-padding applied
5×5 Conv.
10×10×16
2×2 AveragePool
5×5×6
Flatten
Dense
400
Dense
120
Dense
84
10 (Class)
6 Filters 16 Filters
No zero-padding
(c) VGG 16
(a) LeNet 5
Input
227×227×3
11×11 Conv.
55×55×96
3×3 MaxPool
27×27×96
3×3 Conv.
13×13×384
3×3 Conv.
13×13×384
3×3 Conv.
13×13×256
Flatten
Dense
43264
Dense
4096
Dense
4096
1000 (Class)
Original input 96 Filters 256 Filters 384 Filters 256 Filters 512 Filters
Zero-padding applied
(b) AlexNet
5×5 Conv.
27×27×256
3×3 MaxPool
13×13×256
Stride=4
MaxPool Stride = 2
AveragePool Stride = 2
MaxPool Stride = 2
Figure 4.11 Original CNN architecture: (a) LeNet 5; (b) AlexNet; (c) VGG 16
102
In order to apply LeNet 5, AlexNet, and VGG 16 to this application, where the size
of input sample is 20 × 20 × 1, the following modifications to the CNN models and
input samples are carried out:
• For LeNet 5, it can be easily implemented by only replacing the original
classification layer by a 1-neuron fully-connected classification layer, because
series arc fault detection is a binary classification problem. No modifications are
needed for the input samples.
• For AlexNet, the filter size and stride in the first convolution layer are changed
from 11 × 11 to 9 × 9 and from 4 to 1, respectively. Similarly, the output layer is
replaced by a 1-neuron fully-connected classification layer. Each original sample is
transformed from 20 × 20 × 1 to 20 × 20 × 3 by replicating itself three times
along the channel axis.
• For VGG 16, the strides in the last 3 max pooling layers (there are 5 max pooling
layers in total) are changed from 2 × 2 to 1 × 1 . Similarly, the output layer is
replaced by a 1-neuron fully-connected classification layer and each original
sample is transformed from 20 × 20 × 1 to 20 × 20 × 3 by replicating itself three
times along the channel axis.
Besides the modifications mentioned above, BN is applied to each convolution
layer and fully-connected layer (except the classification layer) in order to improve the
speed, performance, and stability of ANN models especially for deep ANN models
[138]. The results are shown in Section 4.5.2.5.
There are two main requirements for real-time DC series arc fault detection in PV
systems:
• It requires time-sensitive data processing,
103
• It can be deployed at the edge: the overall computation burden of the designed
algorithm cannot exceed the computation capability of the edge device.
As a result, to address these application-specific requirements and constraints
mentioned above, it is necessary to design an optimal CNN structure that can achieve a
balance between required computation efforts and performance on series arc fault
detection. There are several tunable hyperparameters in a typical CNN structure,
including the number of convolution layers, the number of filters in each convolution
layer, the size of filter, the number of fully connected layers, and the number of neurons
in each fully connected layer. The optimal lightweight CNN structure is shown in
Figure 4.12 and Table 4.2.
N2
data points window
Minmax Normalisation
&
2D Arrangement
C
o
n
v
o
lu
ti
o
n
L
ay
er
(
5
b
y
5
)
B
at
ch
N
o
rm
al
is
at
io
n
L
ay
er
+
R
eL
U
L
ay
er
+
M
ax
P
o
o
li
n
g
L
ay
er
(
2
b
y
2
)
F
u
ll
y
C
o
n
n
ec
te
d
L
ay
er
+
B
at
ch
N
o
rm
al
is
at
io
n
L
ay
er
+
R
eL
U
L
ay
er
F
u
ll
y
C
o
n
n
ec
te
d
L
ay
er
+
C
la
ss
if
ic
at
io
n
L
ay
er
Real-time data captured by CT
F
u
ll
y
C
o
n
n
ec
te
d
L
ay
er
+
B
at
ch
N
o
rm
al
is
at
io
n
L
ay
er
+
R
eL
U
L
ay
er
-0.5
-1
0
0.
5
1
Point 1
j=1,...,N
i=1,..., N
Point (j-1)N+i
Feature visualization in
convolution layer
Feature visualization in
max pooling layer
Figure 4.12 The optimal lightweight CNN structure and feature visualisation
104
Table 4.2 Structure and parameters of the optimal lightweight CNN
L
ay
er
N
o
.
L
ay
er
t
y
p
e
K
er
n
el
s
iz
e
N
o
.
o
f
k
er
n
el
S
tr
id
e
Z
er
o
-p
ad
d
in
g
B
N
A
ct
iv
at
io
n
O
u
tp
u
t
sh
ap
e
1 2D Convolution 5×5 3 1 No Yes ReLU 16×16×3
2 Maxpooling 2×2 1 2×2 No No - 8×8×3
3 Dense 8 1 - - Yes ReLU 8
4 Dense 5 1 - - Yes ReLU 5
5 Dense 1 1 - - No Sigmoid 1
The optimal CNN structure is determined by hyperparameter tuning through trial
and error based on following extensive studies. For training, the dataset is divided into a
training dataset and a testing dataset with a ratio of 50%:50%. Some other training
parameters are fixed for all case studies in this chapter: the learning rate is 0.001, the
maximum number of epochs is 100, the batch size is 64, the optimiser is stochastic
gradient decent (SGD), and categorical cross-entropy is used as the loss function. The
training process and offline validation are performed using a CentOS 7 Linux operating
system with a Tesla P100 GPU and an Intel (R) Xeon (R) Gold 6126 CPU in the rest of
chapters. The following metrics are used to evaluate the performance of the algorithms:
𝐴𝑟𝑐𝑖𝑛𝑔 =
𝑇𝑃
𝑇𝑃 + 𝐹𝑃
× 100% (4.22)
𝑁𝑜𝑟𝑚𝑎𝑙 =
𝑇𝑁
𝑇𝑁 + 𝐹𝑁
× 100% (4.23)
𝑆𝑒𝑛𝑠𝑖𝑏𝑖𝑙𝑖𝑡𝑦 =
𝑇𝑃
𝑇𝑃 + 𝐹𝑁
× 100% (4.24)
𝑆𝑎𝑓𝑒𝑡𝑦 =
𝑇𝑁
𝑇𝑁 + 𝐹𝑃
× 100% (4.25)
𝑂𝑣𝑒𝑟𝑎𝑙𝑙 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃 + 𝑇𝑁
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁
× 100% (4.26)
where TP, TN, FP, and FN represent the number of correct detections of fault, correct
105
detections of normal conditions, undetected faults, and nuisance tripping, respectively.
Sensibility, which is also known as precision, is an indicator to measure the system
sensitivity related to normal operating conditions and normal transient events (the rate
to avoid false alarm). Safety, which is also known as recall, is an indicator to measure
the system sensitivity to DC series arc fault (the rate toavoid missing alarm).
4.5.2.1. Size of Filter
To demonstrate the effect of filter size on CNN performance, other settings of the
CNN in Table 4.2 are kept the same. As shown in Table 4.3, the filter size is set to
3 × 3, 5 × 5, 7 × 7, and 9 × 9. As the filter size increases from 3 × 3 to 5 × 5, the
testing accuracy is greatly improved. Also, it can be seen that a larger filter size
introduces fewer parameters in the CNN structure. However, it is found that increasing
the filter size further would lead to a decline in the accuracy, especially arcing accuracy.
Thus, the optimal filter size is 5 × 5.
Table 4.3 Influence of filter size on CNN performance
Filter size Arcing Normal Sensibility Safety
Overall
Accuracy
Number of
parameters
3 × 3 99.10% 99.36% 99.36% 99.10% 99.23% 2097
𝟓 × 𝟓 99.46% 99.99% 99.99% 99.46% 99.72% 1737
7 × 7 99.20% 99.99% 99.99% 99.21% 99.60% 1449
9 × 9 99.13% 99.96% 99.96% 99.14% 99.54% 1223
4.5.2.2. Number of Filters in the Convolution Layer
Similarly, to investigate the impact of the number of filters in the convolution layer,
the other parts in the optimal CNN structure are fixed. As presented in Table 4.4, the
overall accuracy of CNN improves substantially by 18.51% when the number of filters
increases from 1 to 2. When it rises from 2 to 3, there is still a noticeable enhancement
106
in arcing accuracy from 98.78% to 99.46%. As can be seen, a slight improvement can
be obtained when the number of filters is further increased. However, the trade-off is
significant growth in the total number of parameters, which is not cost-effective.
Therefore, 3 filters are selected in the first CNN layer.
Table 4.4 Influence of number of filters on CNN performance
Number of filters
(5 × 5)
Arcing Normal Sensibility Safety
Overall
Accuracy
Number of
parameters
1 81.73% 80.04% 80.37% 81.42% 80.88% 653
2 98.78% 99.99% 99.99% 98.79% 99.39% 1195
3 99.46% 99.99% 99.99% 99.46% 99.72% 1737
5 99.54% 99.98% 99.98% 99.54% 99.76% 2821
7 99.49% 99.99% 99.99% 99.49% 99.74% 3905
4.5.2.3. Number of Convolution Layers
In general, when there is adequate training data, more convolution layers can lead
to better performance since higher level and more generalised features can be learned
by the CNN [136]. As expected in this case study, CNN with two convolution layers
can improve the arcing accuracy from 99.46% to 99.57% as shown in Table 4.5.
However, the computation burden increases significantly: the CNN with two
convolution layers has about 4.4 times as many parameters as that with one convolution
layer. Therefore, one convolution layer is sufficient to achieve high performance with
low computation requirements.
Table 4.5 Influence of number of convolution layers on CNN performance
Convolution layer
settings
Arcing Normal Sensibility Safety
Overall
Accuracy
Number of
parameters
3-MP 99.46% 99.99% 99.99% 99.46% 99.72% 1737
3-MP-6 99.57% 99.99% 99.99% 99.57% 99.78% 7593
3-MP-6-12 99.62% 99.99% 99.99% 99.62% 99.80% 8685
107
4.5.2.4. Number of Fully Connected Layers and Number of Neurons in Each Layer
Similarly, CNNs with two fully connected layers followed by a classification layer
(the arcing accuracies are above 99%) is better than CNNs with one fully connected
layer followed by a classification layer (the arcing accuracies are less than 99%) as
shown in Table 4.6. When the number of fully connected layer is fixed, as the number
of neurons in each fully connected layer increases, CNNs tend to demonstrate slightly
better performance at the expense of substantially increasing number of parameters.
Therefore, the proposed setting of fully connected layer is optimal.
Table 4.6 Influence of fully connected layer settings on CNN performance
Fully connected
layer settings
Arcing Normal Sensibility Safety
Overall
Accuracy
Number of
parameters
8-1 98.74% 99.96% 99.96% 98.76% 99.35% 1675
16-1 98.73% 99.99% 99.99% 98.75% 99.36% 3259
32-1 98.81% 99.99% 99.99% 99.82% 99.40% 6427
8-5-1 99.46% 99.99% 99.99% 99.46% 99.72% 1737
16-10-1 99.48% 99.99% 99.99% 99.48% 99.73% 3463
32-20-1 99.53% 99.97% 99.97% 99.53% 99.75% 7155
64-40-1 99.50% 99.96% 99.96% 99.50% 99.73% 15499
4.5.2.5. Comparison with Very Deep CNNs
The proposed lightweight CNN is also compared to deeper CNNs. The settings and
modifications of these deep CNNs are discussed in the beginning of Section 4.5.2, and
the results are presented in Table 4.7. It is worth mentioning that very deep CNNs are
difficult to train without BN, generally caused by gradient vanishing. Modified AlexNet
and modified VGG 16 without BN fail to converge most of the times. Also, the
performance of LetNet 5 without BN (overall accuracy of 99.37%) is worse than
LeNet5-BN (overall accuracy of 99.84%). Although deep CNNs can achieve near-
perfect overall accuracy, it is not possible to deploy such complex models in real-time
108
in a cost-effective manner. For example, the overall accuracy for modified VGG 16-BN
is 99.95%, while the number of parameters is about 40 million.
Table 4.7 Performance comparison of different CNNs
Filter size Arcing Normal Sensibility Safety
Overall
Accuracy
Number of
parameters
LetNet 5 99.15% 99.58% 99.58% 99.15% 99.37% 28589
LeNet 5-BN 99.74% 99.94% 99.94% 99.74% 99.84% 29493
Modified AlexNet-BN 99.84% 99.99% 99.99% 99.84% 99.91% 24757761
Modified VGG 16-BN 99.91% 99.99% 99.99% 99.91% 99.95% 39941313
Proposed CNN 99.46% 99.99% 99.99% 99.46% 99.72% 1737
In summary, the proposed lightweight CNN is proven to be optimal, striking a good
balance between accuracy and computation burden, and it is well suited for real-time
deployment in resource-constraint edge devices.
4.5.3. Evaluation of Different ML Classifiers
Although several ML algorithms demonstrate promising results in recent literature,
the performance of these ML classifiers cannot be directly compared because the
datasets used in such investigations are different. Therefore, a comparative study among
different popular ML algorithms using the same datasets is performed and presented in
this section. Their effectiveness in DC series arc fault detection in PV systems is
examined.
4.5.3.1. Datasets Preparation
One of the datasets is the same as the one described in Section 4.5.1, where no
manual feature extraction is performed on the raw CT signals.
In addition, to investigate the influence of feature extraction, a 2D feature map is
designed based on the wavelet packet entropy as described in Chapter 3. The fault
109
signal is analysed frame-by-frame to extract the common patterns. After performing 3-
level wavelet packet entropy calculation to a frame of data, a feature vector with the
size of 8 × 1 is formulated. To standardise the data, self-normalisation is performed
with a cumulative value of 1. The key point to note during the feature extraction process
is that the extracted features should be both sufficient to identify the faults and immune
to false tripping. Since the fault current nature includes random variation as its
properties, the feature needs to be more reliable against false tripping. To include more
temporal information and improve the reliability of the proposed feature, 6 adjacent
frames with 50% overlapping are merged into a 2D feature map. Each feature map is
calculated using 0.07-second CT signal under 20 kHz sampling frequency, and the
frame size is 400 (corresponding to 0.02 s duration).
The detailed formulation process of 2D feature map based on normalised wavelet
packet entropy is shown in Figure 4.13. Some examples of normal and arcing 2D
feature map are also visualised. It can be seen that high entropy values tend to be
concentrated in the intermediate frequency bands (band 3-7) in the normal state, whilst
theytend to be concentrated in the lower frequency bands (band 1-4) in the arcing state.
This gives an intuitive indication that they have some distinctive features from each
other. Therefore, the designed feature map is suitable as input to different ML classifier
to obtain higher-level features for classification.
Ultimately, 20,000 arcing feature maps and 20,000 normal feature maps are
prepared as the second dataset.
110
Frame size
(400 points)
Band 8
Band 7
Band 6
Band 5
Band 4
Band 3
Band 2
Band 1N
o
rm
al
is
ed
w
av
el
et
p
ac
k
et
e
n
tr
o
p
y
v
ec
to
r
Band 8
Band 7
Band 6
Band 5
Band 4
Band 3
Band 2
Band 1
Band 8
Band 7
Band 6
Band 5
Band 4
Band 3
Band 2
Band 1
50% overlapping
8 by 6 2D feature map
Visualisation of normal 2D feature map
Visualisation of arcing 2D feature map
Figure 4.13 Feature extraction process and visualisation of normal/arcing feature map
111
4.5.3.2. Settings for Different ML Classifiers
For comparative study, conventional ML methods including the shallow BPNN,
NB, SVM, and RF, are selected. In order to obtain the optimal performance, grid search
and five-fold validation is performed during the training of each conventional ML
classifier [115]. Note that the 2D samples are flattened before being fed into these ML
classifiers. The settings of conventional ML classifiers for both datasets are as follows:
• For SVM: Gaussian, linear, polynomial (cubic), sigmoid kernels are considered;
regularisation parameter is set to 0.01, 0.1, 1, and 10; the gamma parameter is set to
auto, which is 1 divided by the number of features.
• For NB: Gaussian kernel is selected.
• For shallow BPNN: ReLU and sigmoid are selected as activation functions. Two
hidden layers are used, and the number of each hidden layers is determined by trial
and error. Generally, the number of neurons in the first hidden layer is identical to
the number of features in the input layer, and the number of neurons in the second
hidden layer is reduced by half.
• For RF: The maximum depth of the tree is set to 10, 30, 50, 80, and 100; the
maximum number of features to consider when looking for the best split is set to 2,
3, log2(number of input features), and square root of the number of input features;
minimum samples required to be at a leaf node is set to 3, 5, and 10; the minimum
number of samples required to split an internal node is set to 6, 10, and 20; the
number of trees in the forest is set to 100, 200, and 400.
Furthermore, DL methods, including deep BPNN, SAE, LSTM, Bi-LSTM, and
CNN, are chosen. The common settings for the training of neural networks are the same
as described in Section 4.5.2. Similarly, 2D samples are flattened before being fed into
the deep BPNN. Also, a BN-ReLU layer is applied after each fully connected (except
112
the final classification layer), convolution, LSTM or Bi-LSTM layer in this case study.
The neural network structures are also determined by extensive trial and error.
The optimal settings of different DL models for the first dataset are as follows:
• For deep BPNN: it consists of four hidden layers. The number of neurons in the
first hidden layer is identical to the number of features in the input layer, and the
number of neurons in the second hidden layer is reduced by half. The last two
hidden layers and the classification layer are the same as the designed optimal
lightweight CNN in Table 4.2, which is the 8-5-1 structure.
• For SAE: in the encoder part, it consists of three fully connected layer with number
of neurons of 200, 100, and 50, respectively, i.e. for each additional fully connected
layer, the number of neurons is reduced by half.
• For LSTM: it consists of two LSTM layers followed by the 8-5-1 structure for
fully-connected layers. The first LSTM layer has 20 units, and the type of
connection between the first LSTM layer and the second LSTM layer is many-to-
many. The second LSTM layer has 10 units, and the type of connection between
the second LSTM layer and the next fully-connected layer is many-to-one.
• For Bi-LSTM: it has the same structure as the LSTM model except the two LSTM
layers are replaced by two Bi-LSTM layers.
• For CNN: it has the same structure as shown in Table 4.2.
For the second dataset, the settings in first few layers are modified based on the
input since the input size changes from 20 × 20 to 8 × 6 (e.g. for LSTM, the number of
units in the first and the second layers are changed from 20 to 10, and 10 to 5,
respectively). In order to maintain the number of parameters approximately the same,
the 8-5-1 structure is changed to 16-10-1 for the second dataset.
113
4.5.3.3. Results of Comparative Study
The results are presented in detail in Table 4.8. The same metrics, as presented in
(4.22) to (4.26), are used to evaluate the performance of different ML algorithms.
When raw data (the first dataset) is used as input, DL algorithms, including LSTM,
Bi-LSTM, and CNN demonstrate superior diagnosis performance compared to other
ML methods. The shallow BPNN with sigmoid activation function only reaches 72.24%
accuracy, while the situation becomes much better when ReLU activation function is
used. This is because ReLU introduces sparsity into the network for regularisation, and
redundant features can be effectively discarded. The deep BPNN further increases the
overall accuracy since hierarchical and more robust features have been learnt by the
deep structure. Conventional methods such as NB, SVM, and RF show mediocre
performance as expected, because the input dimension is too large. The CNN with the
proposed lightweight structure achieves the best detection accuracy with an overall
accuracy of 99.72%.
When effective feature extraction is applied (the second dataset), all conventional
ML methods can be significantly improved. Among them, SVM demonstrates the best
overall detection accuracy of 97.55%. However, the performance of DL method is
degraded due to possibility of losing some useful information during the manual feature
extraction process. Likewise, CNN achieves the best detection accuracy with an overall
accuracy of 97.95%.
Among all the tested ML methods, CNN achieves the best overall classification
accuracy regardless of the presence of feature extraction or not. Furthermore, DL
models with raw data as input can achieve best-in-class overall accuracy.
114
Table 4.8 Evaluation of different popular ML methods
Method Arcing Normal Sensibility Safety
Overall
Accuracy
Using raw data as input (0.02 seconds, 20-kHz sampling rate)
Shallow BPNN (Sigmoid) 62.30% 82.18% 77.76% 68.55% 72.24%
Shallow BPNN (ReLU) 89.75% 97.46% 97.25% 90.48% 93.61%
NB 72.34% 71.42% 71.68% 72.08% 71.88%
SVM 80.89% 87.96% 87.04% 82.15% 84.42%
RF 85.81% 84.84% 84.99% 85.67% 85.32%
Deep BPNN 92.56% 98.44% 98.34% 92.97% 95.50%
SAE 93.95% 98.21% 98.13% 94.20% 96.08%
LSTM 98.18% 99.48% 99.47% 98.20% 98.83%
Bi-LSTM 98.95% 99.31% 99.31% 98.95% 99.13%
CNN 99.46% 99.99% 99.99% 99.46% 99.72%
Using 2D feature map based on normalised wavelet packet entropy as input
(0.02*3.5 seconds, 20-kHz sampling rate)
Shallow BPNN (Sigmoid) 93.97% 88.36% 88.98% 93.61% 91.16%
Shallow BPNN (ReLU) 95.23% 95.37% 95.36% 95.24% 95.30%
NB 94.07% 88.93% 89.47% 93.75% 91.50%
SVM 96.87% 98.23% 98.21% 96.91% 97.55%
RF 96.64% 96.38% 96.39% 96.63% 96.51%
Deep BPNN 96.99% 98.31% 98.29% 97.03% 97.65%
SAE 97.27% 98.50% 98.48% 97.30% 97.88%
LSTM 92.81% 97.82% 97.71% 93.15% 95.32%
Bi-LSTM 96.52% 96.35% 96.36% 96.51% 96.44%
CNN 97.85% 98.05% 98.05% 97.85% 97.95%
4.5.4. Real-time Implementation and Validation Results
For real-time validation experiments, the PV emulator is programmed to simulate a
1.5-kW grid-tied PV system with open-circuit voltage (𝑉𝑜𝑐) of 207.2 V and short-circuit
current ( 𝐼𝑠𝑐 ) of 7.95 A at standardtest condition (STC). Additionally, a 21 μH
inductance is connected between the PV emulator and the 1.5-kW single-phase solar
inverter. The reason to include this inductance is to simulate the filtering effect of the
PV connection cables [14], which can also test the ability of generalisation of the
proposed method.
For real-time implementation, a final decision operator is typically introduced to
115
strike a balance among reliability, accuracy, and response speed [52], [159]. For
instance, when the detection algorithm outputs 4 successive arcing events, a final
decision will be made [52]. Although the final detection time is 4 times longer than the
original detection time, it can greatly reduce the false tripping rate.
In this thesis, multiple-window strategy based on binomial distribution is adopted.
A flow chart of the overall real-time series arc fault detection is illustrated in Figure
4.14. Here, 𝛼 and 𝛽, assumed to be binomial distributed, are the mis-operation rate and
mal-function rate of every output from the LTCNN classifier. The improved mis-
operation rate and mal-function rate, 𝑃𝑚𝑖𝑠 and 𝑃𝑚𝑎𝑙 , can be calculated by (4.27) and
(4.28):
𝑃𝑚𝑖𝑠 = ∑ 𝐶𝑘
𝑖
𝑚−1
𝑖=0
(1 − 𝛼)𝑖𝛼𝑘−𝑖 (4.27)
𝑃𝑚𝑎𝑙 = 1 − ∑ 𝐶𝑘
𝑖
𝑚−1
𝑖=0
𝛽𝑖(1 − 𝛽)𝑘−𝑖 (4.28)
where 𝐶𝑘
𝑖 is the binomial coefficient, and 𝑘 ≥ 𝑖 ≥ 0. By varying the values of m and k,
one can adjust the false tripping rate and the malfunction rate. k is chosen based on the
worst case detection time, which is quite flexible to the users as long as the time limit
satisfies the absolute time limit indicated in UL-1699B Standard, which is 𝑇𝑅 =
max( 750/(𝐼𝑎𝑟𝑐 × 𝑉𝑎𝑟𝑐) , 2.5) seconds [15]. Considering the fact that the detection
time of existing arc fault detectors is in the order of hundreds of ms, k=10 is chosen to
get the worst-case detection time of 200 ms. After choosing the value of k, the value of
m can be manually selected to achieve satisfactory performance. For example, given the
accuracies of the proposed CNN in Table 4.7, 𝛼 = 0.54% and 𝛽 = 0.01% can be
obtained. Then, 𝑃𝑚𝑖𝑠 and 𝑃𝑚𝑎𝑙 are both almost 0 using (4.25) and (4.26) with m=3 and
k=10. Similar results can be obtained with m=2 and k=10, which indicates that m=2 is
116
sufficiently accurate; however, m=3 is chosen to make the detection algorithm more
immune to false tripping. Therefore, the final alarm signal will be issued when at least
m=3 samples are determined as arcing in a sliding window consisting of k=10 samples
with a step size of 1. Thus, great improvements in reliability and accuracy can be
achieved. Based on the tests using pre-recorded time-series data without shuffling, it is
verified that no wrong decisions occurred.
Real-time current data captured by CT
Minmax normalisation (i)
Output of the LTCNN classifier, O(i)
Raw Input (i), 20ms sliding window
(i is the window index)
2D Arrangement (i)
FlagCount(i)=count1([O(i),
O(i-1), O(i-k+1)])
If FlagCount(i) > (m-1)
Series arc fault detection
i++
If i > k-1
No
Yes
Yes
No
DL based classification
Accuracy & reliability
improvement
FlagCount(i)=count1([O(i),
O(i-1), , O(1)])
Figure 4.14 Flowchart of real-time series arc fault detection
Then, the trained CNN is deployed in a prototype based on an NI-CompactRIO-
9030 real-time embedded system and tested in various conditions. NI-CompactRIO is a
general purpose real-time embedded industrial controller (with NI Linux real-time
operating system), which is convenient and flexible for developing prototype and proof-
of-concept. The overall resource utilisation of NI-CompactRIO for the deployment of
117
the proposed CNN is less than 10%. For mass production and commercialisation, the
algorithm could be implemented in field programmable gate arrays or application-
specific integrated circuits, which could result in significant cost reduction. For
example, low-cost devices such as Kendryte K210 and Zynq-7020 based development
systems, which cost in range of $ 30-200 AUD, have the capabilities of operating more
complex CNNs as compared to the proposed CNN. Therefore, low-cost real-time
implementation can be easily achieved. The real-time test results are presented in the
rest of this section. The oscilloscope screen display shows 4 different signals:
• Internal output digital signal of CNN (yellow trace, CH1)
• Loop current capture by the current probe with a ratio of 1:10 (green trace, CH2)
• Loop current captured by the CT, which is the input signal (blue trace, CH3)
• Output digital signal of the final decision with the multiple-window strategy (pink
trace, CH4).
The quick sudden changes in irradiance can cause some AFCI/AFD to trip because
they determine the series arc faults by monitoring the rapid changes in current signal
[37], [58]. In Figure 4.15, the irradiance level is reduced from 1000 W/m2 to 500 W/
m2 with a step of 125 W/m2 using PV Power Profile Emulation software to simulate
some small disturbances. In Figure 4.16, a relatively large disturbances with a large
current drop (approximately 50%) is produced by changing the irradiance level from
1000 W/m2 to 600 W/m2 . The proposed algorithm does not trip on these step
changes. Throughout the experiment, although a mis-operation event is generated from
CNN during the normal conditions as shown in Figure 4.17, the final decision of the
algorithm is still correct due to the multiple-window strategy.
In Figure 4.18, an inrush current is obtained caused by the inverter initialisation
operation. After about 30 seconds, the inverter starts to adjust the operating point to the
118
maximum power point, and the signature of inverter noise varies dramatically as
illustrated on the blue trace. During this start-up period, the proposed algorithm does
not experience any unwanted tripping.
Step changes caused by irradiance level changes
No response
Figure 4.15 Response to small step changes induced by irradiance level changes
A large step change caused by irradiance level change
No response
Figure 4.16 Response to a relatively large step change
119
A mis-operation
during the normal states
Final decision is correct
Step changes caused by irradiance level changes
Figure 4.17 A mis-operation is experienced during normal conditions
Inverter initialisation
transient
Inverter start-up and
MPPT operation
Figure 4.18 Response to start-up transients and MPPT operation of the inverter
120
Not only the proposed algorithm is able to respond to the series arc fault generated
during the inverter start-up period in 56 ms as shown in Figure 4.19, but it can also
detect short-term intermittent series arc faults accurately within 60 ms as shown in
Figure 4.20 and Figure 4.21. In Figure 4.22, even though a mal-function event is
experienced from CNN during the arcing conditions, the final decision of the algorithm
is correct. This event might be because of the arc noise reduction caused by long-term
arc burning. The higher temperature makes arcing more stable and generates less
variation in the loop current signal [77], thus hiding the arc signatures.
A series arc fault during
inverter start-up
54 ms
Figure 4.19 Response to series arc fault during inverter start-up and MPPT operation
121
Several intermittent
series arc faults
54 ms
58 ms
Sustained
arcing
No response
Figure 4.20 Response to several intermittent series arc faults followed by a sustained
arcing
53 ms 58 ms
Several intermittent
series arc faultsStep change due to sudden
irradiance level change
Arc extinguish
No response
Figure 4.21 Response to several intermittent series arc faults followed by an arcing
with increasing gap distance
122
Step change due to sudden
irradiance level change Sustained arc fault
A mal-operationduring the arc states
Final decision is correct
Figure 4.22 A malfunction experienced during arc conditions
4.6. Discussion and Recommendations
As demonstrated in Table 4.8, ML methods, especially DL methods, can achieve
excellent DC series arc fault diagnosis accuracy when sufficient labelled data is
available for training. However, this is generally not the case for many practical
applications [160], [161]. In the following, the practical challenges are discussed, and
some potential solutions are presented.
4.6.1. Imbalanced Dataset or Small Dataset
Training ML classifiers with an imbalanced or small dataset can severely affect the
classification performance. Researchers in [115] report on the impacts of the number of
training samples on classification accuracy. It is found that the diagnosis accuracy
reduces with fewer training samples for all types of ML methods investigated in [115].
ML classifiers trained with imbalanced dataset tend to concentrate on classifying the
majority class (consist of sufficient samples) while neglecting the minority class
123
(consist of insufficient samples). This is the case in reality; the fault data in the target
system are usually rare or even not available. To mitigate this problem, a common way
is to carry out simulation or superimpose the normal signals with Gaussian/Pink noise
to enlarge the fault dataset [103], [131], [162]. However, it is still unclear whether the
artificial fault samples superimposed with such noises share the same features with real
fault signals or not. For example, Andrea et al. in [163] have shown that the Gaussian
noise used in their arc fault simulation is not a good choice, especially for time-domain
fluctuation patterns. Using a high-quality dataset is a pre-requisite for ML methods,
especially for end-to-end DL algorithms without feature extraction. As a result,
advanced data augmentation techniques need to be developed and investigated to create
high-quality synthetic samples to enlarge the dataset.
Another possible way is to use transfer learning (e.g. domain adaptation). Transfer
learning aims to leverage the knowledge of a well-defined domain (e.g. laboratory
system where sufficient data can be obtained) to enhance the performance of the ML
models on a target domain task with less required training samples (e.g. in-service
system where few labelled fault data can be obtained). Although good results have been
achieved through transfer learning in other fields of fault diagnosis [161], it has not
been utilised for DC series arc fault detection in PV systems.
4.6.2. Inconsistency between Training and Testing Dataset
In the majority of studies with ML methods, the dataset is divided into a training
dataset and a testing dataset with certain ratios (e.g. 80% to 20% or 50% to 50%) under
the following important assumptions: (1) the training dataset and testing dataset have
different data distribution; (2) the testing dataset can represent the data encountered in
field operation. However, this is often not the case, especially for practical applications.
124
For example, Xia et al. observed the same phenomena for DC series arc fault detection
in electrical vehicle systems caused by various reasons, such as different types of loads
[110], different voltage/current levels [109], and different arc gap distance [109]. This
problem could be mitigated by establishing a classification model trained on a complex
dataset with all required cases. In this case, the testing data is drawn from the same
distribution with the training data. However, it is extremely time-consuming and even
unrealistic to gather enough data from a large number of cases. Furthermore, the ML
classifier trained with a complex dataset generally performs worse than that trained with
a simple dataset [146]. Therefore, transfer learning may provide a more feasible
solution to tackle this problem effectively.
The investigation of such phenomenon is yet to be conducted in PV applications of
DC arc fault detection, thus the motivation for this research as detailed in the following
chapter.
4.6.3. Unlabelled Dataset
The majority of researchers are focusing on supervised learning, which indicates
that the datasets need to be fully annotated. Compared with labelled datasets which
require manual annotations from domain experts, unlabelled datasets are much easier to
obtain [116]. Moreover, the use of transfer learning has demonstrated promising results
in massive unlabelled datasets in fault diagnostics fields [161]. Consequently, using
unlabelled data to develop effective DL methods can be a worthy research area in the
future.
4.6.4. ML Model Complexity and Real-time Capability
The computation complexity of ML algorithms should be considered for cost-
125
effective real-time deployment in resource-constraint edge computing devices (e.g.
distributed on-line monitoring system or intelligent arc fault detector). The computation
complexity and the accuracy of ML models mainly depend on the architecture of the
model, the sampling rate for data acquisition, the size of the input samples, etc.
Although DL approaches can outperform conventional ML methods in fault diagnostics
in many cases, accuracy improvements are at the expense of exponentially increasing
computation resources as well as energy consumption. Therefore, the complexity of the
problem as well as resource constraints of the specific application need to be taken into
account in the development of the ML models, especially for DL models. As a result, it
is recommended to report essential information such as training/testing time and
sensitivity to hyperparameters of the models, which can enable direct comparison across
different models [164]. On the other hand, researches on efficient hardware
implementation in low-cost devices, such as field programmable gate arrays and
application-specific integrated circuits, can also pave the way to realize cost-effective
solutions based on DL methods for on-line intelligent fault diagnostics.
4.6.5. Model Interpretability
Even though DL methods can achieve top-rated accuracy compared to conventional
ML methods, they are less amenable to interpretation because of their deeper structures
[136]. The lack of core understanding restrains the development of DL on a
fundamental level. The DL models are selected based on trial and error instead of
rigorous logical theories. To address this, some prior works in [165], [166] carried out
comprehensive theoretical analysis about interpreting and understanding DL models.
These studies shed some light on selecting the optimal sets of hyper-parameters for DL
models, which can help to achieve more reliable and robust DC series arc fault
126
detection performance.
4.7. Conclusion
This chapter presents a DL-based approach to detect DC series arc faults in PV
systems, utilising a CNN with a 2D raw data matrix as input. To the best of the author’s
knowledge, this is the first time DL is applied in the context of DC series arc faults
detection. A lightweight CNN structure is designed through hyperparameter-tuning to
facilitate cost-effective real-time implementation in practice. Based on the off-line
validation results, with sufficient training data, the proposed CNN is capable of
detecting series arc faults under different experimental conditions with an overall
accuracy exceeding 99%. A comparative study on different popular ML classifiers is
carried out using the same dataset, and the CNN demonstrates the best-in-class
detection accuracy. Furthermore, a higher level of accuracy can be achieved if a more
complex CNN model is used.
The proposed algorithm is deployed in a prototype monitoring system to verify the
real-time operations under different conditions with a PV emulator based single-phase
grid-tied system. ThereArc Model ......................................................................................... 22
2.3.4. High-Frequency Variation Caused by Arc Faults ............................................ 24
2.4. DC ARC FAULTS DETECTION METHODS IN PV SYSTEMS...................................... 25
2.4.1. Sensors for Measurement ................................................................................. 26
2.4.2. Fast Fourier Transform .................................................................................... 26
2.4.3. Short Time Fourier Transform ......................................................................... 30
2.4.4. Wavelet Transform ........................................................................................... 31
2.4.5. Statistical Analysis ........................................................................................... 35
2.4.6. Model-based Methods ...................................................................................... 39
2.4.7. Machine Learning based Methods ................................................................... 40
2.4.8. Other Types of Methods ................................................................................... 46
2.5. DISCUSSION AND CONCLUSION ............................................................................. 52
3. CHARACTERISTICS STUDY ON DC SERIES ARC FAULT ........................... 55
3.1. INTRODUCTION...................................................................................................... 55
3.2. EXPERIMENTAL SETUP .......................................................................................... 56
X
3.3. STATIC CHARACTERISTICS .................................................................................... 59
3.3.1. V-I Characteristics ........................................................................................... 59
3.3.2. Stable Operating Point ..................................................................................... 61
3.3.3. Load Current Effect and Source Voltage Effect .............................................. 63
3.4. HIGH-FREQUENCY VARIATION IN ARC CURRENT .................................................. 64
3.4.1. Wavelet Packet Entropy ................................................................................... 64
3.4.2. Effect of Arc Phase .......................................................................................... 66
3.4.3. Load Current Effect .......................................................................................... 70
3.4.4. Source Voltage Effect ...................................................................................... 73
3.4.5. Gap Distance Effect ......................................................................................... 75
3.5. DISCUSSION AND CONCLUSION ............................................................................. 76
4. DC SERIES ARC FAULT DETECTION IN PV SYSTEMS USING DEEP
LEARNING ........................................................................................................................ 79
4.1. INTRODUCTION...................................................................................................... 79
4.2. CLASSICAL MACHINE LEARNING .......................................................................... 80
4.2.1. Artificial Neural Network ................................................................................ 80
4.2.2. Support Vector Machine .................................................................................. 83
4.2.3. Decision Tree and Random Forest ................................................................... 85
4.2.4. k-Nearest Neighbours ....................................................................................... 86
4.2.5. Others ............................................................................................................... 87
4.3. DEEP LEARNING .................................................................................................... 87
XI
4.3.1. Deep Fully-Connected Neural Network ........................................................... 90
4.3.2. Autoencoder ..................................................................................................... 90
4.3.3. Convolutional Neural Network ........................................................................ 92
4.3.4. Recurrent Neural Network ............................................................................... 93
4.4. EXPERIMENTAL SETUP .......................................................................................... 95
4.5. PROPOSED DEEP LEARNING BASED SERIES ARC FAULT DETECTION METHOD
USING CONVOLUTIONAL NEURAL NETWORK .................................................................... 97
4.5.1. Dataset Preparation .......................................................................................... 98
4.5.2. Hyperparameters Setting and Offline Validation Results .............................. 100
4.5.2.1. Size of Filter ....................................................................................... 105
4.5.2.2. Number of Filters in the Convolution Layer ...................................... 105
4.5.2.3. Number of Convolution Layers ......................................................... 106
4.5.2.4. Number of Fully Connected Layers and Number of Neurons in Each
Layer ............................................................................................................................ 107
4.5.2.5. Comparison with Very Deep CNNs .................................................. 107
4.5.3. Evaluation of Different ML Classifiers .......................................................... 108
4.5.3.1. Datasets Preparation........................................................................... 108
4.5.3.2. Settings for Different ML Classifiers ................................................ 111
4.5.3.3. Results of Comparative Study ........................................................... 113
4.5.4. Real-time Implementation and Validation Results ........................................ 114
4.6. DISCUSSION AND RECOMMENDATIONS ............................................................... 122
XII
4.6.1. Imbalanced Dataset or Small Dataset ............................................................ 122
4.6.2. Inconsistency between Training and Testing Dataset .................................... 123
4.6.3. Unlabelled Dataset ......................................................................................... 124
4.6.4. ML Model Complexity and Real-time Capability ......................................... 124
4.6.5. Model Interpretability .................................................................................... 125
4.7. CONCLUSION ....................................................................................................... 126
5. INTELLIGENT DC SERIES ARC FAULT DETECTION IN PV SYSTEMS
USING DA-DCGAN WITHOUT TARGET-DOMAIN FAULT DATA .................... 128
5.1. INTRODUCTION.................................................................................................... 128
5.2. EXPERIMENTAL SETUP ........................................................................................ 130
5.2.1. Experimental Setup and Conditions in Source Domain ................................. 130
5.2.2. Experimental Setup and Conditions in Target Domain ................................. 130
5.3. PROPOSED DA-DCGAN ..................................................................................... 132
5.3.1. Generative Adversarial Networks .................................................................. 132
5.3.2. Optimisation Procedures and Deep Learning Model Structures .................... 133
5.4. CASE STUDY 1: OFFLINE VALIDATION RESULTSis no unwanted tripping experienced during the inverter start-up,
MPPT operation, or current step changes. Furthermore, the proposed algorithm can
accurately detect both short-term intermittent and sustained series arc faults in a timely
manner. The sensitivity, effectiveness, and robustness of the proposed algorithm are
confirmed through both offline and real-time validations.
Finally, detailed recommendations and potential solutions are provided to address
several problems that prevent intelligent DC series arc fault detection from being
applied to real-world engineering applications.
127
Some of the work described in this chapter has been published in:
1. Shibo Lu, Animesh Sahoo, Rui Ma, and B. T. Phung, “DC Series Arc Fault
Detection using Machine Learning in Photovoltaic Systems: Recent Developments and
Challenges,” International Conference on Condition Monitoring (CMD), Phuket,
Thailand, 25-28 Oct. 2020.
2. Shibo Lu, Hua Chai, Animesh Sahoo, and B. T. Phung, “Condition Monitoring
based on Partial Discharge Diagnostics using Machine Learning Methods: A
Comprehensive State-of-the-Art Review,” accepted for publication in IEEE
Transactions on Dielectrics and Electrical Insulation, in press, 3 July 2020.
128
5. Intelligent DC Series Arc Fault Detection in PV Systems
using DA-DCGAN without Target-Domain Fault Data
5.1. Introduction
With the recent advances in computing methodologies and information technology,
data-driven ML-based methods become increasingly popular and demonstrate
promising results in fault diagnosis task in many fields such as high impedance fault
detection in medium voltage networks [167], failure detection in electrical machines
[154], and track circuit fault in railway systems [168]. A number of recent studies have
achieved good results for DC series arc fault detection using conventional ML methods
as reviewed in Section 2.4.7. In Chapter 4, the effectiveness of various DL algorithms
has been examined, and they demonstrate superior detection accuracy over conventional
ML methods. Although these ML methods, especially DL methods, can achieve
excellent accuracy, one of the main issues is domain-switching, which can cause severe
performance degradation as discussed in Section 4.6.2. For example, some algorithms
are developed using data collected from power-electronics based DC power sources (i.e.
PV emulators). Even though such power sources can reproduce similar DC output
characteristics to that of real PV systems, the specific features of signal may vary
depending on the types of power source as well as the structure of the experimental
setup [37], [169]. In addition to that, the electrical arc signals can be masked and
modified because of the parasitic capacitance and inductance contributed by real PV
modules and circuit cables [131], [169]. Different domain adaptations as well as transfer
learning techniques have been developed in other fields to address domain-shifting
problems, such as handwriting-digit recognition [170] and object recognition [171],
[172]. Recently, they have been applied to fault diagnosis, such as gear fault detection
129
in gearbox systems using relatively small datasets with deep CNN-based transfer
learning [173], machine fault diagnosis using deep one-dimensional CNN-based
transfer learning network with unlabelled data [174], and cross-domain fault diagnosis
using sparse auto-encoder and fine-tuning with target-domain data [175].
However, most of these investigations assume the target-domain fault data is
available and sufficient, and this is a major challenge. In practice, obtaining sufficient
series arc fault data in a real PV system is costly and time-consuming. Furthermore,
even though supervisory control and data acquisition systems can provide information
about PV systems, they are not much useful due to low-sampling frequency and
intensive efforts for extracting useful arc fault signals from huge amount of unlabelled
data. Because of these reasons, the performance of data-driven ML algorithms often
degrades dramatically when they are applied to a different domain (i.e. from laboratory
to field).
In this chapter, an effective methodology, DA-DCGAN, is proposed to address the
performance degradation of the DL based detection algorithm in a different domain.
Target-domain fault data can be artificially generated based on normal signals, and they
can be applied for domain adaptation to achieve reliable and accurate diagnosis of
cross-domain DC series arc faults. It is more applicable to the practical situations in
some real industries and applications, where only the source-domain data (both normal
and faults data) and the target-domain normal data are available for detection algorithm
development. It could significantly reduce the effort (i.e. collecting arc fault signal)
during the algorithm development stage. Furthermore, a lightweight CNN classifier is
employed for cross-domain fault diagnosis instead of using a very deep neural network
as in other papers, which could be an appropriate solution to enable real-time DC series
arc fault diagnosis in different practical environments.
130
The proposed methodology is formulated under the following conditions:
1. Sufficient normal data and arcing data can be obtained in the source domain
(laboratory setup based on PV emulator), while only sufficient normal data can be
obtained in the target domain for training (real grid-connected PV systems). Target-
domain arcing data is collected for testing only.
2. The source-domain data and target-domain data have different distributions
because the system parameters, characteristics, and operating conditions are
different.
5.2. Experimental Setup
5.2.1. Experimental Setup and Conditions in Source Domain
For the preparation of the source-domain data, the same experimental setup as in
Section 4.4 is used: a 1.5-kW emulator-based grid-connected PV system. The same
dataset, which has been presented in Section 4.5.1, is used as the source-domain dataset
for DA-DCGAN case study. It consists of 20,000 normal samples and 20,000 arcing
samples with a total size of 40,000. Each sample consists 400 data points (corresponds
to 0.02-second).
5.2.2. Experimental Setup and Conditions in Target Domain
For target-domain data collection and real-time testing, the PV emulator is replaced
by a rooftop PV string consists of four JINKO JKM350M-72 mono-crystalline PV
panels. The key information of PV panels can be found in Table 5.1. Additionally, more
than 82 meters of 6 𝑚𝑚2 regular solar cables are wired for connections from the
rooftop to the PV connection box. The overall experimental setup and schematic
representation are illustrated in Figure 5.1. Series arcing experiments are performed at
131
different times on a sunny day from 11 am to 5 pm to obtain target-domain arcing and
normal sample at different current levels. It should be emphasised that series arc faults
are generated at the start, middle, and end of the PV string. These fault locations are
recommended for series arc fault detection tests by UL-1699 B [3]. Ultimately, 25,000
normal samples and 5,000 arcing samples are extracted to form the target-domain
dataset, where 20% of randomly chosen normal samples are reserved for offline testing
and the rest of them are used for training and generating dummy target-domain arcing
samples.
Figure 5.1 Experimental setup and schematic representation for target-domain data
collection and real-time validation tests
132
Table 5.1 Specification of PV model JINKO (JKM350M-72) at STC
Rated Power 𝐼𝑀𝑃𝑃 𝑉𝑀𝑃𝑃 𝐼𝑆𝐶 𝑉𝑂𝐶
350 W 8.94 A 39.10 V 9.38 A 47.5 V
Note that every time-sequence sample in both source domain and target domain is
pre-processed using [-1, 1] minmax normalisation and arranged into an image-basedsignal of 20×20 size before being fed into any neural network in DA-DCGAN.
5.3. Proposed DA-DCGAN
5.3.1. Generative Adversarial Networks
Generative adversarial network (GAN) was initially proposed by I. Goodfellow et
al. in 2014 [176], which provides an alternative solution to maximum likelihood
estimations. An illustration of a general framework of GAN is shown in Figure 5.2.
GANs are deep neural nets comprising two parts: a generator, G, and a discriminator, D.
G generates synthetic samples and tries to fool D, while D tries to distinguish these
generated samples from the real ones. Therefore, G and D compete in a two-player
minimax game as expressed in (5.1):
min
𝐺
max
𝐷
𝐿𝐺𝐴𝑁(𝐷, 𝐺) = 𝔼𝓍~𝑃𝑟𝑒𝑎𝑙(𝓍)[log 𝐷(𝓍)] + 𝔼𝓏~𝑃𝒵(𝓏)[log(1 − 𝐷(𝐺(𝓏)))] (5.1)
where 𝔼(∙) denotes the expectation of the input ∙; 𝓍 and 𝓏 are the data sampled from
real data distribution 𝑃𝑟𝑒𝑎𝑙(𝓍) and noise distribution 𝑃𝒵(𝓏) , respectively; D is the
output of the discriminator, which is the probability that came from the real data 𝑥
instead of from the generated data.
For a perfect D, 𝐷(𝓍) should be 1 when a real sample is input, and 𝐷(𝓍) should be
0 when a fake sample is input. However, original GANs sometimes can generate noisy
133
and incomprehensible fake samples. Due to this, GANs are known to be difficult to
train. Therefore, to enhance the performance and stabilise the training process of GANs,
a class of CNNs called DCGAN is introduced to mitigate the existing problems [177].
Generator
(G)
Discriminator
(D)
Real
or
Fake?
Real samples
Fake samples
Latent space
Back-propagation
Figure 5.2 General framework of a generative adversarial network
5.3.2. Optimisation Procedures and Deep Learning Model Structures
The process of the proposed DA-DCGAN methodology is visualised in Figure 5.3,
and it comprises two stages.
Figure 5.3 Overview of DA-DCGAN for DC series arc fault diagnosis in PV systems
134
Let 𝓍𝑠,𝑎 and 𝓍𝑠,𝑛 be samples of arcing and normal category from the source-
domain data distributions 𝑃𝑠; 𝓍𝑡,𝑎 and 𝓍𝑡,𝑛 be samples of arcing and normal category
from the target-domain data distributions 𝑃𝑡 ; 𝑦𝑎 = 1 and 𝑦𝑛 = 0 be labels for arcing
and normal data from both domains, respectively.
At the first stage, the DCGAN is trained through an adversarial training process,
where D is trained to distinguish real arcing data 𝓍𝑠,𝑎 from artificially generated arcing
data 𝐺(𝓍𝑠,𝑛), and G is trained to transform normal data 𝓍𝑠,𝑛 into high-quality fake
arcing data 𝐺(𝓍𝑠,𝑛) to fool D. Therefore, the optimisation objectives can be defined as
follows: the network D tries to minimise the classification loss 𝐿𝐷 between fake arcing
𝐺(𝓍𝑠,𝑛) and real arcing 𝓍𝑠,𝑎 using the source-domain dataset:
𝐿𝐷 = −𝔼𝓍𝑠,𝑎~𝑃𝑠(𝓍𝑠,𝑎)[log 𝐷(𝓍𝑠,𝑎)]
− 𝔼𝓍𝑠,𝑛~𝑃𝑠(𝓍𝑠,𝑛)[log(1 − 𝐷(𝐺(𝓍𝑠,𝑛)))] (5.2)
while the transformative network G tries to minimise the following loss 𝐿𝐺:
𝐿𝐺 = 𝔼𝓍𝑠,𝑛~𝑃𝑠(𝓍𝑠,𝑛)[log(1 − 𝐷(𝐺(𝓍𝑠,𝑛)))] (5.3)
In the second stage, a lightweight CNN with the optimal CNN structure as designed
in Chapter 4 is trained to minimise the classification error using labelled source-domain
dataset 𝒟𝑠(𝓍𝑠, 𝑦) = { {(𝓍𝑖
𝑠,𝑎, 𝑦𝑖
𝑎)}𝑖=1
𝑁𝑠,𝑎 , {(𝓍𝑖
𝑠,𝑛, 𝑦𝑖
𝑛)}𝑖=1
𝑁𝑠,𝑛} of 𝑁𝑠,𝑎 arcing samples and 𝑁𝑠,𝑛
normal samples, and target-domain dataset 𝒟𝑡(𝓍𝑡, 𝑦) =
{ {(𝐺(𝓍𝑖
𝑡,𝑛), 𝑦𝑖
𝑎)}𝑖=1
𝑁𝑡,𝑛 , {(𝓍𝑖
𝑡,𝑛, 𝑦𝑖
𝑛)}𝑖=1
𝑁𝑡,𝑛} of 𝑁𝑡,𝑛 normal samples and 𝑁𝑡,𝑛 dummy arcing
samples. The binary categorical cross-entropy loss 𝐿𝐵𝐶𝐸 based on a batch of data with
size of m can be defined as follows:
135
𝐿𝐵𝐶𝐸 = −
1
𝑚
∑ 𝑦𝑖 log
1
1 + ℯ−((𝑤𝑙𝑓)𝑇𝑓(𝓍𝑖)+𝑏𝑙𝑓)
𝑚
𝑖=1
+ (1 − 𝑦𝑖) log(1 −
1
1 + ℯ−((𝑤𝑙𝑓)𝑇𝑓(𝓍𝑖)+𝑏𝑙𝑓)
)
(5.4)
where 𝑦𝑖 is the ground truth category label of 𝑖𝑡ℎ sample (arcing for 1 and normal for
0); 𝑓(∙) is the high-level feature in the lightweight CNN classifier before classification
layer; 𝑤𝑙𝑓 and 𝑏𝑙𝑓 denote the weight coefficient matrix and bias coefficient of the last
fully-connected layer. Since the convolution layer, pooling layer, and fully-connected
layers in the lightweight CNN classifier work together as a high-level feature extractor,
it is responsible for extracting useful features from the input signal. Maximum mean
discrepancy (MMD) is employed to assist the optimisation process in the lightweight
CNN classifier to enable domain invariant feature learning in order to achieve inter-
domain condition classification. MMD is a metric to estimate the distribution
discrepancy distance between data drawn from different distributions [178]. The square-
MMD loss in DA-DCGAN can be expressed as follows:
𝐿𝑀𝑀𝐷 = ‖
1
𝑁𝑠,𝑎
∑ 𝑓(𝓍𝑖
𝑠,𝑎) −
𝑁𝑠,𝑎
𝑖=1
1
𝑁𝑡,𝑎
∑ 𝑓(𝓍𝑗
𝑡,𝑎)
𝑁𝑡,𝑎
𝑗=1
‖
ℋ
2
+ ‖
1
𝑁𝑠,𝑛
∑ 𝑓(𝓍𝑖
𝑠,𝑛) −
𝑁𝑠,𝑛
𝑖=1
1
𝑁𝑡,𝑛
∑ 𝑓(𝓍𝑗
𝑡,𝑛)
𝑁𝑡,𝑛
𝑗=1
‖
ℋ
2
(5.5)
where ‖∙‖ℋ is any reproducing kernel Hilbert space. To reduce 𝐿𝑀𝑀𝐷 , high-level
features from different domains can be closer in ℋ . Therefore, combining both
optimisation objects of the second stage, the overall optimisation objective for
lightweight CNN classifier is to minimise (5.6):
𝐿𝐶 = 𝐿𝐵𝐶𝐸 + 𝜆𝐿𝑀𝑀𝐷 (5.6)
136
where 𝜆 is the punishment coefficient for 𝐿𝑀𝑀𝐷 term to control how strong the domain
adaptation is. The detailed architectures and parameters of different neural networks in
DA-DCGAN are listed in Table 5.2. In the Table, Conv. Denotes the two-dimensional
convolution layer, and Conv. (UpSam) denotes the up-sampling two-dimensional
convolution layer. For the leaky ReLU activation function used in DA-DCGAN,
parameter α is set to 0.2 based on the recommendations in [177].
Table 5.2 The architecture of different neural networks in DA-DCGAN
N
et
w
o
rk
N
o
.
L
ay
er
t
y
p
e
K
er
n
el
s
iz
e
N
o
.
o
f
K
er
n
el
S
tr
id
e
P
ad
d
in
g
B
N
A
ct
iv
at
io
n
D
ro
p
o
u
t
O
u
tp
u
t
G 1 Dense 3200 1 - - Yes ReLU No 3200
2 Conv. (UpSam) 3×3 128 2×2 Yes Yes ReLU No 10×10×128
3 Conv. (UpSam) 3×3 64 2×2 Yes Yes ReLU No 20×20×64
4 Conv. 5×5 1 1 Yes No Tanh No 20×20×1
D 1 Conv. 5×5 32 1 No No Leaky ReLU 0.6 16×16×32
2 Conv. 3×3 64 1 No Yes Leaky ReLU 0.6 14×14×64
3 Conv. 3×3 128 1 No Yes Leaky ReLU 0.6 12×12×128
4 Conv. 3×3 256 1 No Yes Leaky ReLU 0.6 10×10×256
5 Dense 1 1 - - No Sigmoid No 1
CNN 1 Conv. 5×5 3 1 No Yes ReLU No 16×16×3
2 Maxpooling 2×2 1 2×2 No No - No 8×8×3
3 Dense 8 1 - - Yes ReLU No 8
4 Dense 5 1 - - Yes ReLU No 5
5 Dense 1 1 - - No Sigmoid No 1
137
Furthermore, the overall detailed optimisation procedure as well as value of
different key parameters of DA-DCGAN are presented in Algorithm 5.1
Algorithm 5.1: Proposed DA-DCGAN
Parameters: N𝑒𝑝𝑜𝑐ℎ,𝐷𝐶𝐺𝐴𝑁 = 8000, N𝑘 = 1, 𝑚 = 32, 𝑙𝑟 = 0.0001, 𝛽1 = 0.5, 𝛽1 = 0.9 ,
𝑁𝑒𝑝𝑜𝑐ℎ,𝐶 =
4𝑁𝑠,𝑛
2𝑘
× 20, 𝑁𝑠,𝑛 = 𝑁𝑠,𝑎 = 𝑁𝑡,𝑛 = 20000, 𝑘 = 32
Step 1: DCGAN training using source domain data;
for num_epoch = 0, , 𝑁𝑒𝑝𝑜𝑐ℎ,𝐷𝐶𝐺𝐴𝑁 do
for num_train = 1, , 𝑁𝑘 do
Sample a batch of arcing data {𝓍𝑖
𝑠,𝑎}𝑖=1
𝑚 and a batch of normal data {𝓍𝑖
𝑠,𝑛}𝑖=1
𝑚
for i = 1, , m do
𝐿𝐷,𝑖 ⟵ −log 𝐷(𝓍𝑖
𝑠,𝑎) − log(1 − 𝐷(𝐺(𝓍𝑖
𝑠,𝑛)))
end for
Update D by descending its gradient using Adam:
θ𝐷 ⟵ Adam(∇θ𝐷
1
𝑚
∑ 𝐿𝐷,𝑖
𝑚
𝑖=1
, 𝑙𝑟, 𝛽1, 𝛽2)
end for
Sample a batch of normal data {𝓍𝑖
𝑠,𝑛}𝑖=1
𝑚
for i = 1, , m do
𝐿𝐺,𝑖 ⟵ log(1 − 𝐷(𝐺(𝓍𝑖
𝑠,𝑛)))
end for
Update G by descending its gradient using Adam:
θ𝐺 ⟵ Adam(θ𝐺 , ∇θ𝐺
1
𝑚
∑ 𝐿𝐺,𝑖
𝑚
𝑖=1
, 𝑙𝑟, 𝛽1, 𝛽2)
end for
for i = 0, , 𝑁𝑡,𝑛 do
𝓍𝑖
𝑡,𝑎 = 𝐺(𝓍𝑖
𝑡,𝑛)
end for
Step 2: Training thelightweight CNN classifier using source-domain data and target-
domain normal data;
for num_epoch = 0, , 𝑁𝑒𝑝𝑜𝑐ℎ,𝐶 do
Sample {(𝓍𝑗
𝑠, 𝑦𝑗)}𝑗=1
2𝑘 ={{(𝓍𝑖
𝑠,𝑎, 𝑦𝑖
𝑎)}𝑖=1
𝑘 ; {(𝓍𝑖
𝑠,𝑛, 𝑦𝑖
𝑛)}𝑖=𝑘+1
2𝑘 } from 𝒟𝑠(𝓍𝑠, 𝑦);
Sample {(𝓍𝑗
𝑡, 𝑦𝑗)}𝑗=1
2𝑘 ={{(𝓍𝑖
𝑡,𝑎, 𝑦𝑖
𝑎)}𝑖=2𝑘+1
3𝑘 ; {(𝓍𝑖
𝑡,𝑛, 𝑦𝑖
𝑛)}𝑖=3𝑘+1
4𝑘 } from 𝒟𝑡(𝓍𝑡, 𝑦);
for i = 1, , 4k do
Calculate 𝐿𝐵𝐶𝐸,𝑖 using (5.4)
end for
for j = 1, , 2k do
Calculate 𝐿𝑀𝑀𝐷,𝑗(𝓍𝑗
𝑠, 𝓍𝑗
𝑡) using (5.5)
end for
Update C by descending its gradient using Adam:
θ𝐶 ⟵ Adam(θ𝐶 , ∇θ𝐶
(
1
4𝑘
∑ 𝐿𝐵𝐶𝐸,𝑖
4𝑘
𝑖=1
+
𝜆
𝑘
∑ 𝐿𝑀𝑀𝐷,𝑗
2𝑘
𝑗=1
), 𝑙𝑟, 𝛽1, 𝛽2)
end for
138
5.4. Case Study 1: Offline Validation Results
To demonstrate the superior performance of the proposed DA-DCGAN
methodology, several methods are evaluated for comparison as follows:
1) Only the lightweight CNN with the same structure trained using source-domain
data;
2) Only the lightweight CNN with the same structure trained using source-domain
data and target-domain normal data;
3) Transfer component analysis with a SVM classifier trained using source-
domain data and target-domain normal data [179]. The input data is flattened;
4) Deep neural network for domain adaptation in fault diagnosis trained using
source-domain data and target-domain normal data. SVM is employed as the
classifier as suggested in [180]. The input data is flattened;
5) Proposed DA-DCGAN without MMD punishment term trained using the
source-domain data, normal data and dummy arcing data from the target-
domain.
The results using different methods on target-domain testing dataset are presented
in Table 5.3. DA-DCGAN demonstrates excellent performance compared to other
methods. The training detection accuracy of arcing and normal for DA-DCGAN is
98.80% and 99.56%, respectively. On testing dataset, DA-DCGAN dramatically
improves the arcing recognition accuracy (97.68%) while the classification accuracy of
normal state approximately remains unchanged (99.32%) in target-domain series arc
fault detection task as compared to methods 1)-5). Besides using classification accuracy
of arcing and normal category, two other important metrics, named sensibility and
safety, are used to evaluate the performance of the different methods. Sensibility is an
139
indicator to measure the system sensitivity related to normal operating conditions and
normal transient events (the rate to avoid false alarm). Safety is an indicator to measure
the system sensitivity to DC series arc fault (the rate to avoid missing alarm). These two
metrics can be calculated through (4.24) and (4.25) as presented in Section 4.5.2.
Table 5.3 Testing accuracy comparison for target domain series arc fault detection
Method Arcing Normal Sensibility Safety
Overall
accuracy
Computation
complexity
1) 99.98% 18.42% 55.07% 99.89% 59.20% 𝒪 (𝑁2)
2) 76.56% 99.48% 99.33% 80.93% 88.02% 𝒪 (𝑁2)
3) 70.32% 98.92% 98.49% 76.92% 84.62% 𝒪 (𝑁2) / 𝒪 (𝑁3)
4) 80.36% 99.22% 99.04% 83.48% 89.79% 𝒪 (𝑁2) / 𝒪 (𝑁3)
5) 92.12% 99.40% 99.35% 92.65% 95.76% 𝒪 (𝑁2)
Proposed 97.68% 99.32% 99.31% 97.72% 98.50% 𝓞 (𝑵𝟐)
The interactions between the solar inverter and the PV emulator introduce dramatic
variation into the current signal (DC side current is around 7A for both conditions) as
shown in Figure 5.4. Therefore, the fluctuation pattern and the magnitude of the signals
captured from the laboratory and real PV systems are significantly distinctive from each
other. This would degrade the performance of the detection algorithm if the classifier is
trained based solely on the source-domain data. As shown in method 1), Table 5.3, the
testing accuracy of normal state and the sensibility on target-domain dataset are only
18.42% and 55.07%, respectively, which could cause frequent false-tripping and thus
not viable for real-world deployment. Next, the target-domain normal data is included
as described in method 2). The testing accuracy of normal state and sensibility increases
at the expense of decrease in testing accuracy of arcing state and safety. The overall
performance is still not sufficient to be considered a reliable scheme.
140
DC current level = 7 A
Zoomed CT signal
Figure 5.4 Healthy signal capture by CT from source domain and target domain
To demonstrate the impacts of the MMD term on distribution discrepancy of high-
level features from different domains visually, the t-distributed stochastic neighbour
embedding (t-SNE) algorithm is adopted to transform high-level features into a 2D
space [181], [182]. When using this algorithm, 700 samples are randomly chosen from
each category (3500 samples in total). As shown in Figure 5.5 (a), there is some
overlapping of features in the middle that corresponds to different categories when
using the DA-DCGAN without MMD. It results in lower accuracy since it is not
possible to derive a clear separation line between normal and arcing for classification.
On the other hand, in Figure 5.5 (b), with the MMD punishment term, the distribution
discrepancy between the target-domain and source-domain features is minimised, which
helps the lightweight CNN to learn more domain-invariant/sharing features for
classification task with higher accuracy. As a result, the target-domain arcing
classification accuracy improves significantly from 92.12% to 97.68%.
141
(a) (b)
Figure 5.5 Visualisation of high-level features in the lightweight CNN before
classification layer using t-SNE method: (a) DA-DCGAN without MMD; (b) proposed
DA-DCGAN.
The comparisons of computation complexity between different methods are shown
in Table 5.3. SVM computation complexity can be 𝒪 (𝑁2) or 𝒪 (𝑁3) depending on the
kernel size [183]. Here, the computation complexity will be 𝒪 (𝑁2) for SVM in
method 3) and method 4). It is worthwhile to point out that very deep networks such as
GANs are only used in offline to generate fake arcing signal and a lightweight CNN is
used for online embedded application which have the computation complexity of
𝒪 (𝑁2) [184]. On the other hand, SVM works properly with a feature extractor as
shown in method 4) [180]. The feature extractor used in this case study for method 4) is
a SAE consisting of three fully connected layers. Therefore, the resultant computation
effort is similar to the proposed approach. However, the proposed approach could
achieve better outcomes.
142
The proposed method can achieve higher accuracy by varying the structures of the
lightweight CNN. As shown in Figure 5.6, the overall accuracy can improve from 98.5%
to 99.4% by increasing the number of kernels in the convolution layer of the lightweight
CNN from 3 to 8. This is often the case for CNN-based classifier because the kernels in
the convolution layer are responsible for extracting discriminative features for
classification. Thus, having a greater number of kernels can increase the prospect of
finding the optimal feature set but at the expense of increasing computation load.
Number of kernels (number of filters)
2 3 5 8
Sensibility
Figure 5.6 Impact of kernel numbers in the convolution layer of the lightweight CNN
on the performance of target-domain DC series arc fault diagnosis
The effect of the sampling frequency is also investigated. The original dataset is
down sampled to 5 kHz and 40 kHz, respectively, and fed into the DA-DCGAN
following the same procedures. For the 5 kHz dataset (the size of each input sample is
10 × 10 ), the accuracy of arcing and normal reduces to 86.20% and 96.08%,
143
respectively. This is normally the case due to lack of information that can be provided
by a 5 kHz-signal as compared to a 20 kHz-signal. Furthermore, the detection accuracy
of arcing and normal increases slightlyto 98.12% and 99.62% for 40 kHz scenario as
expected (the size of each input sample is 28 × 28). It can be seen that there is no
further significant improvement compared to 20 kHz scenario. Therefore, 20 kHz
sampling frequency (the size of each input sample is 20 × 20) is a relatively good
choice to demonstrate the proposed methodology, and it can be modified according to
the implementation requirements such as the capability of the microcontroller used.
5.5. Case Study 2: Real-time Implementation and Validation Results
Similar to real-time tests as described in Section 4.5.4, for real-time
implementation, a final decision operator is applied to strike a balance between
reliability, accuracy, and response speed. The flow chart of the overall real-time series
arc fault detection has been illustrated in Figure 4.14. m=3 and k=10 is used, and no
wrong decisions are witnessed based on the tests using pre-recorded time-series data
without shuffling.
The lightweight CNN classifier is then implemented in a prototype based on an NI-
CompactRIO-9030 embedded controller and validated under different experimental
conditions in real-time. Some examples of the results are presented in the rest of this
section.
The oscilloscope display shows four different signals:
• Voltage at the DC side of PV inverter – the PV voltage captured by the voltage
probe with ratio of 200:1 (yellow top trace, CH1);
• Output digital signal of the final decision from the monitoring unit (green bottom
trace, CH2);
144
• Arc voltage signal captured by the voltage probe with ratio of 200:1 (blue trace,
CH3);
• Loop current (arc current) captured by the current probe with ratio of 10:1 (pink
trace, CH4).
In Figure 5.7, two large current spikes can be seen before the inverter exporting
power to the main grid because of DC disconnect closing and initialisation operation.
After approximately 40 seconds, the inverter starts to deliver power and perform MPPT.
No unwanted tripping is experienced during the inverter start-up period. In Figure 5.8,
a series arc fault (about 220 W) is initiated at the start of the PV string shortly after a
fast-moving cloud. The fast-moving cloud introduces about 50% current drop in
approximately 5 s. In Figure 5.9, a sudden large current drop is observed caused by a
fast shading disturbance during the experiment (possibly caused by a flying bird). The
proposed algorithm can handle these step changes induced by changes of irradiance and
detect series arcing at 48 ms, which leaves a significant margin compared to the
required response time, 𝑇𝑅 (with an absolute limit of 2.5 s) as calculated in (2.1), listed
in UL-1699B Standard.
In Figure 5.10, series arc fault detection (about 80 W) at low-irradiance level is
performed, and the proposed algorithm responds to the event at about 60 ms. Multiple
intermittent series arc faults (170-200 W) diagnosis scenario is also presented as shown
in Figure 5.11, and the response time is fluctuating around 60 ms. Figure 5.12 shows an
experimental result that the proposed algorithm is able to accurately detect series arc
fault (210 W) generated at the middle of the PV string within 60 ms.
Real-time arc fault detection tests are repeated three times under applicable cases
used in UL-1699B [15]. The separation rate and distance of the arc generator is set to: (i)
2.5 mm/s and 0.8 mm for arcing test at low irradiance level (i.e. below intermediate
145
current levels); (ii) 5 mm/s and 0.8 mm & 2.5 mm for arcing test at high irradiance level
(e.g. near maximum allowable DC current). The arc power is ranging from about 80 W
to 350 W. The proposed diagnosis scheme achieves accurate decisions in all these
aforementioned experimental conditions and detects series arc event at around 60 ms,
meeting the required detection time with a significant margin. In addition, unwanted
tripping tests are carried out under three different loading conditions: single-phase
inverter, DC switch operation, and irradiance step changes. No false tripping is
encountered throughout the experiment.
Close DC disconnect
Inrush current during the initialisation stage
MPPT
No response
Figure 5.7 Response to DC disconnect switch closing, inrush current during
initialisation of inverter, start-up, and MPPT operation
146
Fast moving cloud
No response
Series arc fault
Detected rapidly
(a)
(b)
Figure 5.8 Response to fast moving cloud and a series arc fault at high irradiance level
(10A, full loading): (a) 5s per division; (b) 200ms per division (zoomed)
147
Fast disturbance
No response
Figure 5.9 Response to a fast shading disturbance
Series arc fault
at low irradiance level
Detected rapidly
Figure 5.10 Response to a series arc fault at low irradiance level in a cloudy day
148
Several intermittent
series arc faults
Sustained series arc fault
Arc extinguish
Detected rapidly
(a)
(b)
Figure 5.11 Response to several intermittent series arc faults followed by a sustained
arc fault: (a) 2s per division; (b) 200ms per division (zoomed)
149
Series arc fault at
the middle of array
Detected rapidly
Series arc fault at
the middle of array
Detected rapidly
Arc extinguish
(a)
(b)
Figure 5.12 Response to a series arc fault generated at middle of the PV string on a
sunny day: (a) 1s per division; (b) 100ms per division (zoomed)
150
5.6. Conclusion
This chapter presents a DL-based methodology, DA-DCGAN, for practical
domain-shifting series arc fault detection in PV systems without using target-domain
fault data during training. It provides an effective solution to address the challenges
when applied DL for practical series arc fault detection as discussed in Section 4.6.1,
Section 4.6.2, and Section 4.6.4.
Tests on pre-recorded PV loop current dataset and real-time experiments are carried
out to validate the effectiveness and robustness of the proposed methodology. Without
relying on fault data from real PV systems, which is aligned with practical situations,
the proposed method is able to achieve high detection accuracy without performance
degradation from domain switching.
Some of the work described in this chapter has been published in:
Shibo Lu, Tharmakulasingam Sirojan, B. T. Phung, Daming Zhang, and Eliathamby
Ambikairajah, “DA-DCGAN: An Effective Methodology for DC Series Arc Fault
Diagnosis in Photovoltaic systems,” IEEE Access, vol. 7, pp. 45831-45840, April 2019.
151
6. Intelligent DC Series Arc Fault Detection in PV Systems
using LTCNN-ADA with Limited Target-Domain Fault
Data
6.1. Introduction
Chapter 5 has addressed the domain-shifting problem under the extreme condition
that no fault data from the target domain is available. Sometimes, this is not necessarily
the case. For example, Sandia National Laboratories and many manufacturers do put a
lot of effort into collecting target-domain fault data in the field. These data collection
efforts may be somewhat inadequate but still of significant value. Therefore, it is also
important to develop strategies to optimise the performance of DL-based algorithms
with limited amount of target-domain fault data.
This chapter proposes a new framework, LTCNN-ADA, which aims to improve the
performance of intelligent DC series arc fault detection with limit amount of target-
domain fault data using transfer learning. Most of existing approaches use deep transfer
learning, employing very deep neural networks to enable complex cross-domain feature
learning and knowledge transfer [161], [171], [185], [186]. For example, Shao et al. in
[184] applied deep transfer learning in machine fault diagnosis using the well-known
VGG16 [156]. However, the major drawback of these approaches is that it is difficult to
cost-effectively deploy the trained deep models in real-time edge devices.In the
proposed LTCNN-ADA, a lightweight transfer learning-based strategy using
lightweight transfer network is adopted. To further boost the diagnosis performance and
stabilise the knowledge transfer process from source domain to target domain, ADA,
which can augment the fault dataset through adversarial learning, is also applied.
152
The proposed LTCNN-ADA framework is formulated under the following
conditions:
1. Sufficient normal data and arcing data can be obtained in the source domain (e.g.
laboratory setup based on PV emulator), while only sufficient normal data and
limited fault data can be obtained in the target domain for training (real grid-
connected PV systems).
2. The source-domain data and target-domain data have different distributions
because the system parameters, characteristics, and working conditions are
different.
6.2. Experimental Setup
6.2.1. Experimental Setup and Conditions in Source Domain
Source-domain experiments are performed using a programmable DC power
supply, an arc generated designed according to UL-1699B (2018) [7], a 1.5-kW single-
phase solar inverter, and a 5-kW three-phase solar inverter. The DC power supply is
programmed using PV Power Profile Emulation software to simulate PV systems with
different combination of temperature, irradiance levels, and system voltage and current
levels (at STC) to interface with single-phase or three-phase inverters. Then, DC series
arc faults are generated at a separation speed of 5 mm/s and a gap distance of 0.5 mm at
different experimental conditions. Detailed experimental conditions and setup in the
source domain are given in Table 6.1 and Figure 6.1. Finally, Dataset A and Dataset B
are prepared. Each dataset consists of 20,000 arcing samples and 20,000 normal
samples, and each sample corresponds to 400 points (20 ms duration under 20-kHz
sampling rate).
153
Figure 6.1 Experimental setup in: (a) source domain; (b) target domain
6.2.2. Experimental Setup and Conditions in Target Domain
Target-domain experiments are performed using four to twelve JINKO JKM350M-
72 mono-crystalline solar panels, more than 82-m of 6 mm2 regular solar cables,
together with the above-mentioned components except the DC power supply as shown
in Figure 6.1. DC series arc fault experiments are designed and carried out based on the
154
recommendations from UL-1699B (2018) at different current levels, separation rates,
gap distances, and arc fault locations (at the start, middle, and end of the PV array) as
presented in Table 6.1. According to the above configurations, Dataset C (16,000
normal samples and 10,500 arcing samples) and Dataset D (16,000 normal samples and
10,500 arcing samples) are prepared under 20-kHz sampling rate.
Table 6.1 Description of datasets for LTCNN-ADA case study
Dataset
index
Domain Inverter Power supply
System voltage and
current (STC)
A Source
SMA Sunny boy single
phase inverter (1.5kW)
Magna PV
emulator
𝑉𝑂𝐶: 190𝑉
𝐼𝑆𝐶: 9.38𝐴
B Source
SMA Sunny Tripower
three phase inverter (5kW)
Magna PV
emulator
𝑉𝑂𝐶: 380 − 665𝑉
𝐼𝑆𝐶: 9.38𝐴
C Target
SMA Sunny boy single
phase inverter (1.5kW)
JKM350M-72 PV
Panels
𝑉𝑂𝐶: 190𝑉
𝐼𝑆𝐶: 9.38𝐴
D Target
SMA Sunny Tripower
three phase inverter (5kW)
JKM350M-72 PV
Panels
𝑉𝑂𝐶: 238 − 570𝑉
𝐼𝑆𝐶: 9.38𝐴
Dataset
index
Domain
Temperature and
irradiance level
Separation distance
& separation rate
Number of extracted
real data
A Source
𝑇: 0 − 45 ℃
𝐼𝑟: 400 − 1000 𝑊/𝑚2
0.5 mm
5 mm/s
N:20,000
A:20,000
B Source
𝑇: 0 − 45 ℃
𝐼𝑟: 400 − 1000 𝑊/𝑚2
0.5 mm
5 mm/s
N:20,000
A:20,000
C Target
Sunny/Cloudy
Morning/Afternoon
0.8-2.5 mm
2.5-5 mm/s
N:16,000
A:10,500
D Target
Sunny/Cloudy
Morning/Afternoon
0.8-2.5 mm
2.5-5 mm/s
N:16,000
A:10,500
Similar to the case study in previous chapters, every time-sequence sample is pre-
processed using [-1, 1] minmax normalisation and arranged into a 2D sample of 20×20
size before being fed into any algorithms in this chapter.
6.3. Proposed LTCNN-ADA
For ease of reference, the symbols that are frequently used in this chapter are
summarised in Table. 6.2.
155
Table 6.2 Symbols and descriptions
Symbol Description Symbol Description
𝒟 Domain N Number
x A sample X Sample matrix
y Label of a sample Y Label vector
s, t (sub/sup) Source, target n, a (sub/sup) Normal, arcing
𝑃 Data distribution 𝐿 Loss function
𝔼(∙) Expectation of the input W, b Weight matrix and bias
𝛶 Label space 𝜒 Data space
6.3.1. Transfer Learning
Generally, the training data available must be adequate for ANN-based classifiers
to achieve acceptable fault diagnosis performance. However, for in-service systems in
real industry scenarios, only few labelled fault samples can be obtained since they are
not allowed to operate continuously under fault conditions, making it time-consuming
and difficult to obtain sufficient fault data. To overcome this problem, transfer learning
techniques have been introduced [161], [171]. Transfer learning allows ANN to make
use of the prior knowledge from the source domain, where training samples are
sufficient, then adapt the learned knowledge to assist the training process in the target
domain with limited data.
Given a source domain 𝒟𝑠 = {𝜒𝑠, 𝑃(𝑋𝑠)} and a target domain 𝒟𝑡 = {𝜒𝑡, 𝑃(𝑋𝑡)},
where 𝜒𝑠 and 𝜒𝑡 represent data spaces from the source and target domains, and 𝑃(𝑋𝑠)
and 𝑃(𝑋𝑡) represent the marginal probability distributions from the source and target
domains, 𝑋𝑠 ∈ 𝜒𝑠 , and 𝑋𝑡 ∈ 𝜒𝑡 . In this chapter, 𝑋𝑠 = {𝑥𝑖
𝑠,𝑎, 𝑥𝑗
𝑠,𝑛 }𝑖,𝑗
𝑁𝑠,𝑎,𝑁𝑠,𝑛 and 𝑋𝑡 =
{𝑥𝑖
𝑡,𝑎, 𝑥𝑗
𝑡,𝑛 }𝑖,𝑗
𝑁𝑡,𝑎,𝑁𝑡,𝑛 denote the source and target domain data, and the corresponding
labels are 𝑌𝑠 = {𝑦𝑖
𝑠,𝑎, 𝑦𝑗
𝑠,𝑛 }𝑖,𝑗
𝑁𝑠,𝑎,𝑁𝑠,𝑛 and 𝑌𝑡 = {𝑦𝑖
𝑡,𝑎, 𝑦𝑗
𝑡,𝑛 }𝑖,𝑗
𝑁𝑡,𝑎,𝑁𝑡,𝑛 , respectively. In real
scenarios, generally the number of arcing fault samples in target domain 𝑁𝑡,𝑎 is limited
156
(𝑁𝑡,𝑎 ≪ 𝑁𝑠,𝑎). Given a learning task 𝒯 = {𝛶, 𝑃(𝑌|𝑋)}, where 𝛶 is the label space, 𝑌 ∈
𝛶, and 𝑃(𝑌|𝑋) is the conditional probability distribution (prediction function), the main
objective of 𝒯 is to maximise the classification performance. When 𝒟𝑠 ≠ 𝒟𝑡, transfer
learning aims to accomplish 𝒯𝑡 in 𝒟𝑡 by leveraging the knowledge in 𝒟𝑠 and 𝒯𝑠. Such a
learning task in target-domain achieved by transfer learning can be denoted as 𝒯𝑡: 𝒟𝑠 →
𝒟𝑡 . Figure 6.2 gives a clearer illustration of difference between traditional machine
learning methods and transfer learning-based methods.
Source domain Target domain
Traditional machine learning
Source domain Target domain
Enhanced by transfer learning
Misclassify
Knowledge transferKnown samples Unknown samples
Figure 6.2 Illustrations of traditional ML methods and ML methods enhanced by
transfer learning
6.3.2. Wasserstein Generative Adversarial Networks
GANs are known to be difficult to train, and suffer from many problems such as
mode collapse, non-convergence, diminished gradient [187] even for DCGAN. The
Wasserstein generative adversarial with gradient penalty (WGAN-GP) was proposed to
address those problems. WGAN-GP replaces the Jensen-Shannon divergence by the
Wasserstein distance in the final objective function, and a gradient penalty term is
157
introduced to help the gradients flow back to the generator. The loss function of
WGAN-GP is written as:
min
𝐺
max
𝐷
𝐿𝐺𝐴𝑁(𝐷, 𝐺) = 𝔼𝑥[𝐷(𝑥)] + 𝔼𝓏[𝐷(𝐺(𝓏))]
+ 𝜂𝔼�̂�[(‖∇�̂�𝐷(�̂�)‖2 − 1)2]
(6.1)
where D stands for discriminator, and G stands for generator, Wasserstein distance is
estimated by the first two terms, the last term denotes the gradient penalty with a
weighting factor 𝜂, �̂�~𝑃�̂�(�̂�).𝑃�̂�(�̂�) is a sampling distribution uniformly sampled along
straight lines between pairs of points sampled from 𝑃𝑟𝑒𝑎𝑙(𝑥) and generator sample
distribution 𝑃𝐺(𝑥), and ‖∇�̂�𝐷(�̂�)‖2 is the gradient norm for random sample �̂�.
6.3.3. Procedures of the Proposed LTCNN-ADA
The experimental signals in the time domain and their frequency spectra are
illustrated in Figure 6.3. As can be seen from the figure, signals have sufficient
information for arc fault detection in frequencies below 10 kHz in both single phase and
three phase PV systems. Therefore, it also demonstrates that using 20 kHz sampling rate
is appropriate.
Note that every time-sequence sample in both source domain and target domain is
pre-processed using [-1, 1] minmax normalisation and arranged into an image-based
signal of 20×20 size before being fed into any neural network in LTCNN-ADA.
Given a target-domain dataset {𝑋𝑡, 𝑌𝑡} for training, including {𝑋𝑡,𝑎, 𝑌𝑡,𝑎} with 𝑁𝑡,𝑎
arcing samples and { 𝑋𝑡,𝑛, 𝑌𝑡,𝑛 } with 𝑁𝑡,𝑛 normal samples, the main steps of the
proposed LTCNN-ADA, as illustrated in Figure 6.4, comprise:
158
Without MinMax Normalisation With MinMax Normalisation
Inverter
switch-on
Inverter
initialisation
Instant of
partial shading
Inverter normal
operation
Arcing fault
Inverter
switch-on
Inverter
initialisation
Instant of
partial shading
Inverter normal
operation
Arcing fault
Without MinMax Normalisation With MinMax Normalisation
(a)
(b)
Initialisation
Initialisation
Figure 6.3 Examples of experimental signals (sampled at 200-kHz for analysis purpose)
and their frequency spectra under normal and arcing conditions in (a) single-phase PV
system; (b) three-phase PV system.
159
Step 1: Perform the ADA using {𝑋𝑡,𝑎, 𝑌𝑡,𝑎} and WGAN-GP according to (6.1)
through an adversarial training process. For G, a fully-connected layer with 3,200
neurons is used, followed by two up-sampling convolution layers with 128 and 64
filters. A BN-ReLu layer is used after each abovementioned layer [138], [139]. The last
layer of G is a convolution layer having one filter (5 × 5 in size) with a tanh activation
function in order to produce data in the range of [-1,1]. Therefore, a latent vector fed to
G can be firstly reshaped to a matrix with a size of 5 × 5 × 128 after the fully-
connected layer. Then, a synthesised sample with a desired size of 20 × 20 × 1 can be
generated after being processed by the rest of layers. For D, four convolution layers
with 16, 32, 64, and 128 filters are adopted and LeakyReLu (0.2) is used as activation
function. Unless otherwise mentioned, the filter size is 3 × 3 for all convolution layers.
The choices of the GAN architecture and hyperparameters are based on the
recommendations for image synthesis tasks in [136], [177], [187]. In this study, based
on [187], the learning rate is 𝛿 = 0.0001, batch size is 𝑁𝑏𝑎𝑡𝑐ℎ =64, gradient penalty
factor is 𝜂 = 10, and number of epochs is 14,000 for training WGAN-GP. It is common
practice to train a CNN with a balanced dataset, which is expected to lead to the best
performance [188], [189]. Therefore, after WGAN-GP is converged, to make the target-
domain dataset balanced, the trained generator is used to create 𝑁𝑡,𝑔𝑎 generated arcing
sample, where 𝑁𝑡,𝑛 = 𝑁𝑡,𝑎 + 𝑁𝑡,𝑔𝑎 . Then, { 𝑋𝑡,𝑔𝑎, 𝑌𝑡,𝑔𝑎 }, which consists of 𝑁𝑡,𝑔𝑎
generated arcing samples and corresponding labels, are included in {𝑋𝑡, 𝑌𝑡} for Step 3.
Step 2: Well-train a lightweight CNN classifier using {𝑋𝑠, 𝑌𝑠}, which consists of
sufficient source-domain arcing and normal samples, with cross-entropy loss, 𝐿𝐶𝐸, and
the popular SGD algorithm to initialise the LTCNN [136], as shown in (6.2)-(6.3):
160
𝐿𝐶𝐸 = −
1
𝑁
∑(𝑦𝑖 log
1
1 + ℯ−(𝑊𝑙𝑓
𝑇𝑓(𝓍𝑖)+𝑏𝑙𝑓)
𝑁
𝑖=1
+ (1 − 𝑦𝑖) log(1 −
1
1 + ℯ−(𝑊𝑙𝑓
𝑇𝑓(𝓍𝑖)+𝑏𝑙𝑓)
))
(6.2)
𝜃𝐿𝑇𝐶𝑁𝑁 ⟵ 𝑆𝐺𝐷(𝜃𝐿𝑇𝐶𝑁𝑁 , 𝛻𝜃𝐿𝑇𝐶𝑁𝑁
(𝐿𝐶𝐸(𝑋𝑠, 𝑌𝑠)), 𝛿) (6.3)
where the subscript lf denotes the last layer of the whole neural network, 𝜃 denotes the
network parameters, 𝛻(∙) stands for the calculated gradients of input ∙, and 𝛿 is the
learning rate. For LTCNN, a convolution layer with three 5 × 5 filters and a Max
Pooling layer with 2 × 2 kernel size is used at first. Then, two fully-connected layers
with eight and five neurons are connected before the classification layer. BN-ReLu
layers are applied as well. It is the same CNN structure as designed in Chapter 4.
Step 3: Fine-tune all the parameters of the pre-trained LTCNN obtained from Step
2 with the augmented dataset {𝑋𝑡, 𝑌𝑡} obtained from Step 1, using loss 𝐿𝐿𝑇𝐶𝑁𝑁, based on
(6.4)-(6.5)：
𝐿𝐿𝑇𝐶𝑁𝑁 = 𝐿𝐶𝐸 + 𝜆𝐿𝑟𝑒𝑔𝑢𝑙𝑎𝑟𝑖𝑧𝑒𝑟 (6.4)
𝜃𝐿𝑇𝐶𝑁𝑁 ⟵ 𝑆𝐺𝐷(𝜃𝐿𝑇𝐶𝑁𝑁, 𝛻𝜃𝐿𝑇𝐶𝑁𝑁
(𝐿𝐿𝑇𝐶𝑁𝑁(𝑋𝑡, 𝑌𝑡)), 𝛿) (6.5)
where 𝜆 is the regularisation factor. Note that {𝑋𝑠, 𝑌𝑠 } can be included during the
training process in Step 3 if the regulariser in (6.4) is enabled. After being fine-tuned,
the trained model can be deployed in real-time devices to achieve online DC series arc
fault detection for the target domain. In Step 2 and Step 3 in this study, 𝛿 = 0.001,
𝑁𝑏𝑎𝑡𝑐ℎ =64, 𝜆 = 0, the number of epochs for LTCNN pre-training in Step 2 is
2𝑁𝑠,𝑛
𝑁𝑏𝑎𝑡𝑐ℎ
×
10, and the number of epochs in Step 3 is
2𝑁𝑡,𝑛
𝑁𝑏𝑎𝑡𝑐ℎ
× 10.
161
Higher Level
Feature Reasoning
Decision
Layer
CNN
Blocks
Feature
Extractor
LCE,Source
LCE,Target
LRegularizer,Target
(Optional)
ƏLSource
ƏθHLFR
ƏLSource
ƏθFE
ƏLTarget
ƏθHLFR
ƏLTarget
ƏθFE
PV panels
Step 3:
Fine-Tuning using Target-
Domain Data and Generated
Arcing Data from WGAN-GP
Step 2:
Initialisation using Source-Domain Data
Normal
Arcing
Normal
Arcing
(limited)
Source Domain under Lab Environment
Data Feedforward
(Solid line)
Gradients back-propagation
to update model parameters
(Dot line)
PV emulator
Generator
CriticLatent
Generated arcing
ƏLG
ƏθG
ƏLC
ƏθC
Step 1:
Dataset
Preparation
Target Domain under Realistic Environment
Figure 6.4 Overview of the proposed LTCNN-ADA framework
6.4. Case Study 1: Offline Validation Results
6.4.1. Analysis and Evaluation of Generated Arcing Data
6.4.1.1. Training Loss Curves Analysis of ADA by WGAN-GP
Unlike other machine learning models that can be evaluated by loss functions, it is
difficult to evaluate the quality of a GAN generator based on its loss alone [190].
During the training of a GAN, the losses of generator and discriminator model should
maintain an equilibrium. After they converge to an equilibrium, the generator model can
be assessed by qualitative analysis (e.g. manual inspection or feature visualisation) and
quantitative analysis (e.g. performance on targeted classification tasks). Several
examples of training loss curves of WGAN-GP with 𝑁𝑡,𝑎 =10, 20, 30, 60, 150, and 300
from target domain dataset C are shown in Figure 6.5. It can be observed that WGAN-
162
GP with 𝑁𝑡,𝑎 =10 exhibits extreme instability and fails to converge. On the other hand,
with more samples included during the training process, WGAN-GP can easily reach
convergence. Therefore, to avoid the negative impacts from non-convergence problem
[136], 𝑁𝑡,𝑎 ≥20 is recommended.
Nt,a = 10 Nt,a = 20
Nt,a = 30 Nt,a = 60
Nt,a = 150 Nt,a = 300
Figure 6.5 Training loss curves of WGAN-GP with different 𝑁𝑡,𝑎
6.4.1.2. High Dimensional Feature Visualisation
To demonstrate the effectiveness of the ADA, the popular t-SNE is applied to
visualise the high-dimensional features of the generated fault samples by mapping them
from the feature space (output before the classification layer of the LTCNN classifier)
into a low-dimensional space (2D in this study) [181], [182]. Different transfer tasks
with 𝑁𝑡,𝑎=60 are studied for illustration purpose, and the results are shownin Figure
6.6. The visualisation results indicate that the generated fault samples by ADA share
similar features and distributions with the real arcing samples, including samples from
the training dataset and unseen samples from the testing dataset. Therefore, ADA can
163
generate meaningful samples to assist the training process of the LTCNN. This
observation is also confirmed by the quantitative analysis in Section 6.4.2.
B C B D
A DA C
Figure 6.6 2D visualisation using t-SNE under different transfer tasks (𝑁𝑡,𝑎 = 60)
6.4.1.3. Frequency Domain Analysis
Typically, manual inspection of generated synthetic samples is a common
technique for evaluating GANs. For example, in other applications such as hand-writing
digits recognition, the generated samples are understandable to human [176]. However,
in this research, we are dealing with signals which are less understandable merely
through manual inspection. The 2D matrix of normal, real arcing, and generated arcing
samples are visualised in Figure 6.7.
164
However, it is difficult to see the difference between those samples illustrated in
Figure 6.7(a)-(c). Therefore, FFT is applied to evaluate the generated fault samples in
the frequency domain. Those samples are firstly converted back to time series signals,
and then FFT is performed to obtain the frequency spectrum. As can be seen in Figure
6.7(d), the spectra of generated and real fault samples are similar. These results indicate
that the ADA method is able to generate samples with key characteristics of real fault
samples.
(b) (c)
(d)
(a)
Figure 6.7 (a) examples of three phase normal signal; (b) examples of three phase
arcing signal; (c) examples of generated three phase arcing signal; (d) frequency spectra
of the normalised time-series signals; those signals are randomly selected from the
training dataset D and generated arcing dataset by ADA in the transfer task of A→D
(𝑁𝑡,𝑎 = 60)
165
6.4.2. Results with Different Number of Fault Samples
In order to demonstrate the improvement and effectiveness of the proposed
framework, different approaches are investigated. For simplicity, no regularisers are
used during the training process in this chapter:
1. CNN (Baseline): Only the lightweight CNN is trained by simply using the
target-domain dataset, and no transfer learning or ADA are applied. It is the
original method developed in Chapter 4.
2. LTCNN without ADA: To demonstrate the effect of lightweight transfer
learning strategy, the lightweight CNN is firstly well-trained using the source-
domain dataset. Then its parameters are fine-tuned using the target-domain
dataset. No ADA is applied.
3. LTCNN-ADA: The lightweight CNN is firstly well-trained using the source-
domain dataset. Then its parameters are fine-tuned using a combination of the
target-domain dataset and generated fault dataset by ADA.
The effects of the amount of target-domain fault data are evaluated in this section
under different transfer tasks (A→C, A→D, B→C, and B→D). For Dataset C and
Dataset D, 11,000 out of 16,000 normal samples are used for training, and the
remaining 5,000 samples are reserved for testing. For the arcing fault dataset, 3,000
arcing samples are from those cases where arc faults are generated at the start of the PV
array, among which 20, 30, 60, 150, 300 of them are randomly selected for the target-
domain training dataset, respectively. On the other hand, another 2,500 new arcing
samples generated at the same fault location are prepared. Combined with 5,000 arcing
samples from those cases where arc faults are generated at the middle and the end of the
PV array, a total number of 7,500 fault samples are prepared for the testing dataset.
166
These samples are extracted from more than 40 cases, and the arc fault time duration of
each case is within a few seconds.
It should be emphasised that only fault samples generated from the start of the PV
array are used during the training. For testing, apart from using a new set of fault
samples generated from the same fault location, fault samples from two new cases (arc
fault generated at the middle and the end of the PV array) are also tested. Therefore, the
generalisation property of the proposed framework can be validated.
Figures 6.8-9 compare the overall accuracy (Acc), arcing accuracy (F), and normal
accuracy (N) of the four domain transfer tasks. These metrics can be calculated using
(4.26), (4.22), and (4.23), respectively. Different approaches (CNN, LTCNN, LTCNN-
ADA) are extensively evaluated using different number of target domain fault samples.
The experimental result of each case study is obtained by averaging results from 27
trials to mitigate the bias and randomness effects during the random selection of dataset.
It can be easily observed that more fault samples lead to higher accuracies. Also, it is
noticeable that the normal accuracies in all cases are all close to 100% since sufficient
normal samples are used for training, while the arcing accuracies heavily depend on the
number of fault samples.
167
Acc F
20
N Acc F
30
N Acc F
60
N Acc F
150
N Acc F
300
N
CNN 76.4 60.7 99.9 84 73.4 99.9 87.8 79.7 99.9 95.1 91.8 99.9 97.1 95.2 99.9
LTCNN 95 91.7 99.9 96 93.4 99.9 97 95 99.9 97.9 96.6 99.9 98.5 97.6 99.9
LTCNN-ADA 97.1 95.2 100 97.6 96 100 98 96.7 100 98.6 97.7 100 99 98.3 100
45
50
55
60
65
70
75
80
85
90
95
100
P
e
rc
e
n
ta
g
e
(
%
)
Acc F
20
N Acc F
30
N Acc F
60
N Acc F
150
N Acc F
300
N
CNN 76.4 60.7 99.9 84 73.4 99.9 87.8 79.7 99.9 95.1 91.8 99.9 97.1 95.2 99.9
LTCNN 96.1 93.5 99.9 96.5 94.2 99.9 97.5 95.8 100 98.3 97.1 99.9 98.7 97.8 99.9
LTCNN-ADA 97.8 96.4 100 98.1 96.8 100 98.4 97.4 100 98.8 98.1 99.9 99.2 98.7 100
45
50
55
60
65
70
75
80
85
90
95
100
P
e
rc
e
n
ta
g
e
(
%
)
(b)
B C
A C
(a)
Figure 6.8 Diagnosis results of different methods for Dataset C: (a) A→C; (b) B→C
168
Acc F
20
N Acc F
30
N Acc F
60
N Acc F
150
N Acc F
300
N
CNN 77.5 62.5 100 83.4 72.4 100 92.3 87.2 100 95.9 93.3 100 97.9 96.5 99.9
LTCNN 96.8 94.7 100 97.2 95.3 100 98.2 97 99.9 98.8 98.1 99.9 99.2 98.6 100
LTCNN-ADA 98.4 97.3 100 98.6 97.7 100 98.9 98.2 100 99.2 98.8 100 99.5 99.2 100
45
50
55
60
65
70
75
80
85
90
95
100
P
e
rc
e
n
ta
g
e
(
%
)
Acc F
20
N Acc F
30
N Acc F
60
N Acc F
150
N Acc F
300
N
CNN 77.5 62.5 100 83.4 72.4 100 92.3 87.2 100 95.9 93.3 100 97.9 96.5 99.9
LTCNN 96.5 94.2 99.9 96.9 94.8 100 98.1 96.8 100 98.7 97.9 99.9 99.1 98.6 99.9
LTCNN-ADA 98.2 97.1 100 98.5 97.5 100 98.8 98 100 99.2 98.7 100 99.4 99.1 100
45
50
55
60
65
70
75
80
85
90
95
100
P
e
rc
e
n
ta
g
e
(
%
)
(b)
B D
(a)
A D
Figure 6.9 Diagnosis results of different methods for Dataset D: (a) A→D; (b) B→D
169
In all cases, the arcing accuracies of the proposed LTCNN-ADA method
outperform the other methods, especially when the size of the target domain arcing data
is smaller. For instance, in A→C task with 𝑁𝑡,𝑎 =20 in Figure 6.8(a), 96.4% arcing
accuracy can be obtained by the proposed method, while the arcing accuracies of CNN
and LTCNN are only 60.7% and 93.5%, respectively. On the other hand, when the
number of fault samples increases to 300, the acing accuracies of CNN, LTCNN, and
the proposed method are 95.2%, 97.8%, and 98.7%, respectively. In more challenging
transfer tasks, where the inverters and power sources are completely different between
the source and target domains, e.g. B→C with 𝑁𝑡,𝑎 =20 as shown in Figure 6.8(b),
significant improvements can also be observed by LTCNN (31%) and LTCNN-ADA
(34.5%) as compared to CNN.
In general, the diagnosis performance of CNN is not reliable because of the low
accuracy and large standard deviations. LTCNN dramatically improves the performance
of CNN becauselightweight transfer learning is applied to leverage the learnt
knowledge from the source domain to assist the training process in the target domain.
LTCNN-ADA further increases the performance, and the increase is more noticeable
for smaller fault data size. When 𝑁𝑡,𝑎=20, compared to LTCNN, LTCNN-ADA can
improve the arc fault detection accuracies by 2.9%, 2.9%, 3.5%, and 2.6% for transfer
tasks A→C, A→D, B→C, and B→D, respectively. Besides, the standard deviations are
reduced by approximately half. When 𝑁𝑡,𝑎 increases to 300, 0.9%, 0.5%, 0.7%, and
0.6% improvements can also be achieved. These improvements are still good since the
accuracies of LTCNN have already reached around 98%. The mis-operation rates of
LTCNN, calculated using (6.6), are 2.2%, 1.4%, 2.4%, and 1.4% for four different
transfer tasks, respectively, while for the LTCNN-ADA are 1.3%, 0.9%, 1.7%, and
170
0.8% for four different transfer tasks. Therefore, remarkable reduction in mis-operation
rates ranging from 30%-43% can be achieved.
𝛼 = 1 − 𝐴𝑟𝑐𝑖𝑛𝑔 = (1 −
𝑇𝑃
𝑇𝑃 + 𝐹𝑃
) × 100% (6.6)
The impact of the ratio of real fault data over the ADA generated fault data,
𝑁𝑡,𝑎/(𝑁𝑡,𝑛 − 𝑁𝑡,𝑎), can also be revealed: ADA delivers higher accuracy improvement
under lower ratio. The reason is that, with the help of ADA, the proposed method can
capture more generalised features by looking at a broader set of augmented samples.
Furthermore, standard deviations of the proposed method are lower than other methods,
which reflects that ADA can stabilise the training process. In general, the proposed
method is more robust.
6.4.3. Comparative Study with Related Works
Although remarkable results have been achieved recently, limited work can be
found on lightweight transfer learning-based DC series arc fault detection. To
demonstrate the advantages of the proposed framework, a comparative study is carried
out against several state-of-the-art methods:
1. DA-DCGAN as presented in Chapter 5 [191];
2. Transfer RNN [192];
3. Deep transfer learning with VGG16 [185] (the input data is re-sized to fit
VGG16; the original output layer is replaced by a 1-neuron output layer for
binary classification);
4. Deep transfer learning with VGG-16 and ADA;
5. Transfer RNN with ADA.
171
The same dataset used for the proposed LTCNN-ADA is applied to method 1) - 5)
to benchmark the performance on different transfer tasks (𝑁𝑡,𝑎=60 as the case study).
Similarly, each result is obtained by averaging results of 27 trials. The performance
comparison between LTCNN-ADA and other methods are listed in Table 6.3 in detail.
DA-DCGAN was only validated by transfer task of A→C in [191], while this study
further explores the performance of DA-DCGAN in other transfer scenarios. The DA-
DCGAN based methodology achieves satisfactory arcing fault detection accuracy in
A→C with 96.85% and B→D with 97.41%. However, for more challenging tasks,
B→C and A→D, where the solar inverters are completely different in the source and
target domain, its fault diagnosis accuracies drop drastically to 80.4% and 78.12% on
average with large standard deviations. Similarly, the Transfer RNN in method 2) also
performs mediocrely in B→C and A→D. Method 3) proposed in [185] employs a pre-
trained deep CNN model called VGG16, and it is directly fine-tuned with the target-
domain dataset. Based on the evaluation, 93.96% and 97.09% arcing fault detection
accuracies can be achieved using dataset C and dataset D.
The feasibility of applying ADA to deep neural networks is also investigated.
Method 4) is a modification based on method 3), where the VGG16 is pre-trained with
the source domain dataset, and then fine-tuned with the target domain dataset and
artificial fault samples generated by ADA. Improvements of arc fault diagnosis rate in
range of 1.21% to 2.41% can be obtained in four different transfer tasks compared to
the original method in [185]. Meanwhile, the corresponding standard deviations also
reduce approximately by half. Likewise, method 5) is based on method 2), where ADA
is applied to increase the diagnosis performance and stabilise the training process. As
expected, noticeable improvements in different transfer tasks can be observed as well.
172
In Table 6.3, the proposed LTCNN-ADA demonstrates excellent performance in
arcing fault detection in all cases. It reaches 97.37% in A→C (highest), 96.66% in
B→C (highest), 97.95% in A→D (second highest), and 98.15% in B→D (second
highest), respectively. VGG16-ADA has slightly higher arcing fault detection rate for
A→D (98.32%) and B→D (98.44%) compared to LTCNN-ADA. Although VGG16-
ADA surpasses LTCNN-ADA in these two transfer tasks by 0.37% and 0.29%, the
trade-offs for accuracy improvements are the exponentially increased computation
resources: the inference time of VGG16-ADA is about 125 times longer than that of
LTCNN-ADA (36.34 s as compared to 0.29 s). Therefore, the proposed LTCNN-ADA
framework can achieve better DC series arc fault detection accuracy with significantly
reduced inference time, and so it is more suitable for cost-effective real-time
deployment in resource-constraint edge devices.
Table 6.3 Performance comparison of different algorithms on different transfer tasks
using the same dataset
Method
Target domain: C
(Mean±Std)
Target domain: D
(Mean±Std)
Inference
time
(per 10k
samples)
A → C (F/N) B → C (F/N) A → D (F/N) B → D (F/N)
DA-DCGAN
[191]
96.85±2.73
99.74±0.20
80.40±14.23
99.34±0.42
78.12±13.56
99.75±0.19
97.41±2.66
99.70±0.22
0.31 s
Trasnsfer RNN
[192]
93.85±1.14
99.95±0.03
87.36±2.29
99.94±0.03
89.31±2.16
99.94±0.03
92.38±1.71
99.94±0.04
13.40 s
Trasnsfer RNN -
ADA
96.19±0.86
99.95±0.04
93.71±1.25
99.93±0.03
93.00±1.38
99.95±0.04
95.68±0.98
99.96±0.02
13.31 s
VGG16 [185]
93.96±1.04
99.93±0.03
93.96±1.04
99.93±0.03
97.09±0.87
99.96±0.01
97.09±0.87
99.96±0.01
36.16 s
VGG16-ADA
96.37±0.61
99.94±0.04
95.93±0.76
99.95±0.02
98.30±0.32
99.97±0.02
98.44±0.29
99.96±0.04
36.34 s
LTCNN-ADA
(this work)
97.37±0.69
99.95±0.04
96.66±0.84
99.96±0.02
97.95±0.71
99.95±0.02
98.15±0.69
99.96±0.03
0.29 s
The inference time of the model is measured on a CentOS 7 Linux operating system
with a Tesla P100 GPU and an Intel (R) Xeon (R) Gold 6126 CPU.
173
6.5. Case Study 2: Online Validation Results
To validate the feasibility of the proposed LTCNN-ADA framework, a prototype
setup based on an NI CompactRIO is developed and tested on a single-phase and a
three-phase grid-connected PV systems, respectively. The trained LTCNN (𝑁𝑡,𝑎=60) is
deployed in the prototype using NI LabView programming. Similarly, multiple-window
strategy is applied to enhance the performance of the proposed algorithm in real-time as
described in Section 4.5.4. The flowchart of the overall real-time series arc fault
detection is illustrated in Figure 4.14. The final alarm signal will be issued when at least
m=3 samples are determined as arcing in a sliding window consisting of k=10 samples
with a step size of 1.
Extensive real-time experiments, including arc fault detection tests and unwanted
tripping tests, are conducted under the guidance of UL-1699B (2018) with applicable
cases. Normal transients, such as inrush current induced by inverter switch-on and load
current drop caused by partial shading from fast-moving clouds, are typical events to
cause nuisance tripping of traditional AFDs [37]. Some of the example results are
presented in the rest of this chapter. The real-time signals for each test are illustrated
using an oscilloscope including:
• The alarm signal from the prototype (yellow trace, CH1);
• Loop current (arc current) captured by the current probe with ratio of 10:1 (green
trace, CH2);
• Arcvoltage captured by the voltage probe with ratio of 200:1 (blue trace, CH3);
• Voltage of the PV inverter captured by the voltage probe with ratio of 200:1 (pink
trace, CH4).
Furthermore, those signals are saved using the DAQ system during the experiments
174
so that the arc fault energy can be determined. For AFD tests, the separation speed and
distance are set to 5 mm/s and 0.8 mm under high irradiance conditions, and 2.5 mm/s
and 0.8 mm under low irradiance conditions, respectively [15].
6.5.1. Three-Phase PV System
The prototype is firstly tested in a 5-kW three-phase PV system with 12 panels.
Two inrush currents can be observed in Figure 6.10 during the initialisation stage of the
three-phase inverter. After initialisation, the inverter starts to export power with MPPT
and feed the AC grid. No false tripping is experienced during the transients, while an
alarm signal is raised shortly after the inception of a series arcing. In Figure 6.11,
several series arc faults followed by a sustained series arcing can be observed, and the
prototype can produce correct logic output in a timely manner. In Figure 6.12, a partial
shading event with a relatively long duration (more than 30 seconds) followed by
several intermittent short partial shading events and a series arcing can be observed. The
logic output of the prototype is always false until the series arc fault is introduced. The
proposed algorithm is also tested under the low-irradiance level, where the arc fault
current is about 2.5A as shown in Figure 6.13.
Similar experiments are repeated in the same three phase PV system but with only
6 PV panels. The results are shown in Figure 6.14 to Figure 6.17. The proposed
algorithm does not respond to large load current drops caused by partial shading, inrush
currents caused by closing DC disconnect and initialisation of the inverter, and variation
introduced by MPPT operations as shown in Figure 6.14 and Figure 6.16. Different
types of series arc faults, including intermittent arc fault, low-irradiance arc fault, and
sustained arc fault can be accurately detected within a short time duration.
175
Three phase
inverter switch on
Inrush current
Exporting power
Arc inception
Alarm
Figure 6.10 Response to inverter switch on, start-up, MPPT, and series arcing in a three
phase PV system (12 panels)
Intermittent arc faults
Alarms
Sustained arcing
Figure 6.11 Response to several intermittent series arcing events generated at the
middle of the PV array followed by a sustained series arcing in a three phase PV system
(12 panels)
176
Partial
Shading
Intermittent Partial
Shading
Arc inception
Alarm
Figure 6.12 Response to partial shading and series arcing in a three phase PV system
(12 panels)
Arc inception
Alarm
Figure 6.13 Response to series arcing generated at very low irradiance level in a three
phase PV system (12 panels)
177
Three phase
inverter switch on
Inrush current
Exporting power
Arc inception
Alarm
MPPT
Figure 6.14 Response to inverter switch on, start-up, MPPT, and series arcing in a three
phase PV system (6 panels)
Intermittent arc faults
Alarms
Sustained arcing
Figure 6.15 Response to several intermittent series arcing events generated at the
middle of the PV array followed by a sustained series arcing in a three phase PV system
(6 panels)
178
Cover two PV
panels
Cover one PV
panel
MPPT
Arc inception
Alarm
Figure 6.16 Response to partial shading and series arcing in a three phase PV system (6
panels)
Arc inception
Alarm
Figure 6.17 Response to series arcing generated at the middle of the PV array at low
irradiance level in a three phase PV system (6 panels)
179
6.5.2. Single-Phase PV System
After that, the prototype is tested in a 1.5-kW single phase PV system. In Figure
6.18 to Figure 6.20, unwanted tripping tests against DC switch operation, MPPT
operation, and sharp irradiance step changes are performed, and the algorithm works as
expected without false alarm. Additionally, the proposed algorithm can accurately
detect different types of series arc fault such as intermittent arcing and sustained arcing.
An example of series arc fault detection results in the middle of the PV array at low
irradiance level is shown in Figure 6.21.
A short arc event
Alarm
Inrush current
Exporting power
1Ø inverter
switch on
Open-circuit
Figure 6.18 Response to inverter switch on, start-up, MPPT, and series arcing in a
single phase PV system (4 panels)
180
Intermittent arc faults
Alarms
Sustained
arcing
Figure 6.19 Response to several intermittent series arcing events followed by a
sustained series arcing in a single phase PV system (4 panels)
Alarm
Arc inception
Partial Shading
Figure 6.20 Response to partial shading and series arcing in a single phase PV system
(4 panels)
181
Arc inception
Alarm
Figure 6.21 Response to series arcing generated at the middle of the PV array at low
irradiance level in a single phase PV system (4 panels)
6.6. Conclusion
This chapter presents a new framework, LTCNN-ADA, which only requires
limited amount of target-domain fault data but can give better PV series arc fault
detection accuracy and reduced inference time. It provides an effective solution to
address the challenges when applied DL for practical series arc fault detection as
discussed in Section 4.6.1, Section 4.6.2, and Section 4.6.4.
In LTCNN-ADA, generalised fault features can be extracted from the raw CT
signals through a lightweight transfer network with source domain knowledge and small
amount of target domain fault data. ADA is performed to facilitate the process of
knowledge transfer. Offline experiments are carried out using 4 different datasets with
different power sources and inverters. Furthermore, real-time validation experiments are
designed and performed based on the recommendations of UL-1699B (2018) Standard
182
with applicable cases. Results from the comprehensive analysis demonstrate the
effectiveness of the proposed framework as well as its generalisation on a small fault
dataset.
Some of the work described in this chapter has been submitted to peer-review
journal:
Shibo Lu, Rui Ma, Tharmakulasingam Sirojan, B.T. Phung, and Daming Zhang,
“Lightweight Transfer Nets and Adversarial Data Augmentation for Photovoltaic Series
Arc Fault Detection with Limited Fault Data”, Submitted to Solar Energy, 2020.
183
7. Conclusion and Future Work
7.1. Conclusion
The main objective of this research is to develop practical and intelligent DC series
arc fault detection for PV systems protection. The findings are summarised as follows.
Firstly, this research presents an in-depth literature review covering various aspects
of DC arc fault detection in modern solar PV applications, including arc fault
mechanism and type, arc fault modelling, state-of-the-art techniques for arc fault
detection. The capabilities and limitations of different methods are presented and
discussed. Useful information about each applied method, such as key methodology,
sampling frequency, detection time, and accuracy, are summarised and compared. It is
found that DL has not yet been utilised for investigation in this field and thus it presents
a research gap to be filled.
Then, a characteristics study on DC series arc fault is performed, focusing on high-
frequency variation in the arc current. The arc current and its spectrum show
dependency on the source voltage, load current, and gap distance. The arc fault tends to
produce less arcing noise when the stable operating point is further away from the
interrupted point: the high frequency variation induced by the arc tends to decrease with
increasing source voltage, increasing load current, and smaller gap distance. Therefore,the worst-case scenario can be found accordingly, which can be also used to determine
the minimum threshold value for conventional detection methods. The UL-1699B
Standard also takes these factors into considerations and introduces some modifications
in arcing tests as compared to UL-1699B Outline. For example, the gap distance range
is reduced from 1.6-6.4 mm to 0.8-2.5mm. Furthermore, a method combines WPD and
entropy theory is developed to analyse arc current signals in time-frequency domain. It
184
demonstrates the capability of extracting consistent patterns of series arc fault under
different conditions.
In Chapter 4, several barriers preventing ML, especially DL, from implementation
in practical applications are identified: imbalanced dataset, small dataset, inconsistency
between training and testing dataset, unlabelled dataset, model complexity and real-time
capability, and interpretability. These aspects are still not extensively explored in the
current research and need further development. Potential solutions to these challenges
have been suggested to facilitate the applications of intelligent series arc fault detection
in the industry.
An intelligent DL based series arc fault detection using CNN is proposed. A
lightweight CNN structure is then designed through hyperparameter tuning. The overall
number of parameters is less than 1% of that of some state-of-the-art deep CNN models,
while the detection performance remains superior. It significantly reduces the required
computation resources during the inference process of the trained DL models, which is
more feasible for cost-effective real-time deployment in edge devices.
A comparative study on different popular ML classifiers is carried out using the
same datasets to examine their effectiveness in PV series arc fault detection. The first
dataset is prepared using raw CT time-series data without any hand-crafted feature. The
second dataset is formulated using wavelet packet entropy with 2D feature maps
extracted from the raw CT signal. 5 conventional ML methods, including shallow MLP,
Gaussian NB, SVM, and RF, and 5 DL methods, including deep MLP, CNN, SAE,
LSTM, and Bi-LSTM are evaluated. It is found that conventional ML methods can
benefit greatly from manual feature extraction. Furthermore, DL models with raw data
as input can achieve best-in-class overall accuracy. Of all ML methods tested, CNN
achieves the best overall classification accuracy regardless of the presence of feature
185
extraction or not.
In addition, two novel frameworks, DA-DCGAN and LTCNN-ADA, are newly
introduced in this research for optimising the performance of the proposed lightweight
CNN based series arc fault detection when lacking sufficient target-domain fault data.
They also solve the performance degradation of DL algorithms caused by data
inconsistency between the source-domain data used during the development and the
target-domain data encountered in operation in the field. The DA-DCGAN firstly learns
an intelligent normal-to-arcing transformation from the source-domain data using a
DCGAN. The target-domain dummy arcing data can be generated using target-domain
normal data by the generator of DCGAN (transformer) with the learnt transformation.
By including these fake arcing data into the training dataset, the target-domain arcing
detection accuracy is dramatically improved from 76.56% to 92.12% based on the case
study in Chapter 5. Then, domain adaptation is performed by including an MMD
punishment term during the training of the lightweight CNN, which minimises the
distribution discrepancy between the target-domain and source-domain features. The
target-domain arcing detection accuracy of the lightweight CNN is further enhanced by
approximately 5.5% through learning more domain-invariant/sharing features. Hence, a
robust and reliable fault diagnosis scheme is achieved for the target domain without
using target-domain fault data.
LTCNN-ADA is proposed in Chapter 6, which aims to enhance the detection
performance of the proposed lightweight CNN with limited amount of target-domain
fault data using transfer learning and ADA. Firstly, ADA using WGAN-GP is
performed to enlarge the target-domain fault dataset, making the target-domain dataset
balanced. Then, the lightweight CNN is initialised by the source-domain dataset. It
leverages the source domain knowledge by getting a better weight initialisation. Then,
186
all the parameters of the pre-trained lightweight CNN are fine-tuned using the
augmented target-domain dataset. Based on comprehensive offline analysis, the
lightweight CNN with transfer learning dramatically improves the performance of the
original CNN. With the help of ADA, it further increases the performance, and the
increase is more noticeable for smaller fault data size. For example, when only 20 fault
target-domain fault samples are used, more than 30% improvements in target-domain
arcing detection accuracy can be obtained in different transfer tasks. Comparative
studies with other recent work based on deep transfer network have been carried out.
Results show that the proposed method can achieve better series arc fault detection
accuracy in different transfer tasks with significantly reduced computation complexity
during model inference. The knowledge transfer capability of CNN with lightweight
structure is confirmed.
Under the guidance of UL-1699B Standard (2018) with applicable cases, multiple
real-time experiments on different types of PV systems are carried out to validate the
effectiveness of DL-based intelligent series arc fault detection. Experimental results
demonstrate the proposed methods enable fast and accurate detection of different types
of series arc faults at different fault locations. Robustness against unwanted tripping is
also confirmed through a series of tests covering different transients during normal
operating conditions, such as switching on DC disconnect, inverter start-up and
initialisation, MPPT operation, partial shadings, etc.
7.2. Suggestions for Future Work
Although there are a variety of techniques available to detect DC series arc faults in
PV systems, little research has been done on their localisation. Furthermore, in UL-
1699B Standard (2018), there are no clear requirements regarding localisation of DC
187
series arc faults. In fact, effective and accurate localisation of DC series arc faults
allows for rapid response to determine the root cause of failure and to replace of faulty
components. Therefore, in addition to the development of detection techniques, the
development of reliable methods for locating DC series arc faults is crucial, especially
for large-scale PV systems. DL can be used for further research on the localisation of
DC series arc faults.
With the development of advanced power electronics and control technologies, DC
systems are more capable of realising much more complex functions, and thus
becoming increasingly popular for different types of applications such as smart home,
energy storage systems, and electrical vehicle charging stations. Similar to PV systems,
other DC systems also suffer the problems of arcing faults. Up to the completion of this
thesis, DL has not yet been investigated in these systems in the context of DC arc fault
detection. Therefore, further investigations on this point are also of interest. In the
future, the proposed intelligent DC series arc fault detection may be adapted to a
different application.
This research has solved many practical challenges that prevent intelligent DC
series arc fault detection from being applied to industry including imbalanced dataset,
small dataset, data inconsistency, and model complexity for real-time deployment.
However, there are still some remaining problems, such as unlabelled dataset and model
interpretability,................................................. 138
5.5. CASE STUDY 2: REAL-TIME IMPLEMENTATION AND VALIDATION RESULTS ....... 143
5.6. CONCLUSION ....................................................................................................... 150
6. INTELLIGENT DC SERIES ARC FAULT DETECTION IN PV SYSTEMS
USING LTCNN-ADA WITH LIMITED TARGET-DOMAIN FAULT DATA ........ 151
6.1. INTRODUCTION.................................................................................................... 151
XIII
6.2. EXPERIMENTAL SETUP ........................................................................................ 152
6.2.1. Experimental Setup and Conditions in Source Domain ................................. 152
6.2.2. Experimental Setup and Conditions in Target Domain ................................. 153
6.3. PROPOSED LTCNN-ADA ................................................................................... 154
6.3.1. Transfer Learning ........................................................................................... 155
6.3.2. Wasserstein Generative Adversarial Networks .............................................. 156
6.3.3. Procedures of the Proposed LTCNN-ADA .................................................... 157
6.4. CASE STUDY 1: OFFLINE VALIDATION RESULTS ................................................. 161
6.4.1. Analysis and Evaluation of Generated Arcing Data ...................................... 161
6.4.1.1. Training Loss Curves Analysis of ADA by WGAN-GP ................... 161
6.4.1.2. High Dimensional Feature Visualisation ........................................... 162
6.4.1.3. Frequency Domain Analysis .............................................................. 163
6.4.2. Results with Different Number of Fault Samples .......................................... 165
6.4.3. Comparative Study with Related Works ........................................................ 170
6.5. CASE STUDY 2: ONLINE VALIDATION RESULTS .................................................. 173
6.5.1. Three-Phase PV System ................................................................................. 174
6.5.2. Single-Phase PV System ................................................................................ 179
6.6. CONCLUSION ....................................................................................................... 181
7. CONCLUSION AND FUTURE WORK ............................................................... 183
7.1. CONCLUSION ....................................................................................................... 183
7.2. SUGGESTIONS FOR FUTURE WORK ...................................................................... 186
XIV
REFERENCE ................................................................................................................... 189
XV
List of Acronyms
AFCI Arc Fault Circuit Interrupter
AFD Arc Fault Detector
ANN Artificial Neural Network
ASIC Application Specific Integrated Circuit
Bi-LSTM Bidirectional Long Short-Term Memory
BN Batch Normalisation
BPNN Backpropagation Neural Network
CNN Convolutional Neural Network
CT Current Transformer
DA-DCGAN Domain Adaptation and Deep Convolutional Generative Adversarial
Network
DAQ Data Acquisition
DCGAN Deep Convolutional Generative Adversarial Network
DL Deep Learning
DT Decision Tree
DWT Discrete Wavelet Transform
EMR Electromagnetic Radiation
FFT Fast Fourier Transform
FL Fuzzy Logic
FPGA Field-Programmable Gate Array
GAN Generative Adversarial Network
HMM Hidden Markov Model
XVI
kNN k-Nearest Neighbours
LR Logistic Regression
LSTM Long Short-Term Memory
LTCNN-ADA Lightweight Transfer Convolutional Neural Network with
Adversarial Data Augmentation
ML Machine Learning
MLP Multilayer Perceptron
MMD Maximum Mean Discrepancy
MPPT Maximum Power Point Tracking
NB Naïve Bayes
PCA Principal Component Analysis
PV Photovoltaic
ReLU Rectified Linear Unit
RF Random Forest
RNN Recurrent Neural Network
SAE Stack Autoencoder
SSTDR Spread Spectrum Time Domain Reflectometry
STC Standard Test Conditions
STFT Short-Time Fourier Transform
SVM Support Vector Machine
t-SNE t-distributed Stochastic Neighbour Embedding
WGAN-GP Wasserstein Generative Adversarial Network with Gradient Penalty
WPD Wavelet Packet Decomposition
XVII
List of Figures
Figure 1.1 Examples of arc current waveforms: (a) AC arc fault; (b) DC arc fault ........ 2
Figure 1.2 Fire due to DC arc faults in Bakersfield, USA in 2009 [10]: (a) Overview of
burned PV panels; (b) Burned PV conductors .................................................................. 3
Figure 1.3 Fires due to DC arc faults in Australia in recent years [11] ........................... 4
Figure 2.1 Typical structure of PV systems ................................................................... 12
Figure 2.2 Example of possible locations where arcing may occur in PV systems ....... 14
Figure 2.3 Typical disturbance sources .......................................................................... 17
Figure 2.4 V-I characteristic of arc ................................................................................ 18
Figure 2.5 Equivalent circuit representation of series arc fault...................................... 23
Figure 2.6 An example of hardware structure of the DC series AFD [72] .................... 27
Figure 2.7 Time and frequency resolutions of the original time-series signal, FFT
spectrum, STFT spectrogram, and wavelet transform spectrogram. .............................. 32
Figure 2.8 Comparison of DWT and WPD analysis (3-level as an example) ............... 34
Figure 2.9 Ensemble ML techniques: (a) Bagging; (b) Boosting; (c) Stacking ............. 44
Figure 2.10 Simplified transmission line model for SSTDR ......................................... 49
Figure 3.1 Experimental setup for characteristics study of DC series arc fault ............. 57
Figure 3.2 Diagram of arc fault generator in UL-1699B Standard ................................ 58
Figure 3.3 V-I characteristic under different gap distance ............................................. 60
Figure 3.4 The condition for a stable arcing point ......................................................... 62
XVIII
Figure 3.5 Waveforms of a typical DC series arc fault: (a) arc current; (b) arc voltage 63
Figure 3.6 Average arc resistance under (a) fixed source voltage; (b) fixed load current
......................................................................................................................................... 64
Figure 3.7 High frequency variation of DC series arc fault at 11A/200V ..................... 67
Figure 3.8 Average frequency spectrum of the 2-second data at non-arc state and arc
state at difference arc phase (FFT analysis window is 0.2 seconds) ............................... 68
Figure 3.9 Wavelet-packet entropy of DC series arc fault at 11A/200V ....................... 69
Figure 3.10 DC current dependency for different arc phases ........................................ 70
Figure 3.11 DC load current dependent arc spectrogram under fixed source voltage ... 71
Figure 3.12 V-I curve for fixed source voltage .............................................................. 72
Figure 3.13 Wavelet-packet entropy under fixed DC source voltage ............................ 73
Figure 3.14 DC source voltage dependent arc spectrogram under fixed load current ... 74
Figure 3.15 V-I curve for fixed load current .................................................................. 74
Figure 3.16 Wavelet-packet entropy under fixed load current ...................................... 75
Figure 3.17 Average frequency spectrum of the first 2 seconds data after the gap
distance reached the desired value (FFTwhich need further investigations. Also, implementing DL-based
algorithms in FPGA or ASIC can be done in the future for further cost reduction and
reliability improvement.
This research presents two novel frameworks for cross-domain DC series arc fault
detection in PV systems. They are achieved by leveraging the knowledge of a single-
source domain. For example, for LTCNN-ADA, the accurate series arc fault detection
188
in target-domain PV system is achieved by utilising transfer learning of single-source
domain and a small number of target-domain labelled fault data. Future research can
focus on investigation of transfer learning across multiple source domains. In addition,
transfer learning using unsupervised data from the target domain can also be explored.
189
Reference
[1] Australian Photovoltaic Institute, “Australian PV market since April 2001.”
https://pv-map.apvi.org.au/analyses (accessed May 31, 2020).
[2] D. Smith, “Arc Flash Hazards on Photovoltaic Arrays,” Fort Collins, Colorado,
USA, 2013.
[3] M. Earley and J. Sargent, “National Electrical Code 2011 Handbook,” Natl. Fire
Prot. Assoc. Quincy, MA, 2010.
[4] K. Klement, “DC Arc Flash Studied for Solar Photovoltaic Systems: Challenges
and Recommendations,” IEEE Trans. Ind. Appl., vol. 51, no. 5, pp. 4239–4244,
2015, doi: 10.1109/ESW.2015.7094947.
[5] M. Sedighizadeh, A. Rezazadeh, and N. I. Elkalashy, “Approaches in High
Impedance Fault Detection - A Chronological Review,” Adv. Electr. Comput.
Eng., vol. 10, no. 3, pp. 114–128, Aug. 2010, doi: 10.4316/aece.2010.03019.
[6] A. Ghaderi, H. L. Ginn, and H. A. Mohammadpour, “High impedance fault
detection: A review,” Electr. Power Syst. Res., vol. 143, pp. 376–388, Feb. 2017,
doi: 10.1016/j.epsr.2016.10.021.
[7] M. Mishra and R. R. Panigrahi, “Taxonomy of high impedance fault detection
algorithm,” Measurement, vol. 148, p. 106955, Dec. 2019, doi:
10.1016/j.measurement.2019.106955.
[8] B. Hao, “AI in arcing-HIF detection: A brief review,” IET Smart Grid, vol. in
press, Feb. 2020, doi: 10.1049/iet-stg.2019.0091.
[9] L. Zhu, S. Ji, and Y. Liu, “Generation and developing process of low voltage
series DC arc,” IEEE Trans. Plasma Sci., vol. 42, no. 10, pp. 2718–2719, Oct.
2014, doi: 10.1109/TPS.2014.2330419.
[10] T. Zgonena, L. Ji, and D. Dini, “Photovoltaic DC Arc-Fault Circuit Protection
and UL Subject 1699B,” in Photovoltaic Module Reliability Workshop, 2011,
Accessed: May 29, 2020. [Online]. Available:
https://www.researchgate.net/publication/266054761.
[11] AC Solar Wholesalers, “Solar Fires - DC Arc Faults.”
https://www.acsolarwarehouse.com/news/solar-fires-dc-arc-faults-on-solar-
190
systems/ (accessed Aug. 01, 2020).
[12] B. Brook, “Report of the results of investigation of failure of the 1.1135 MW
photovoltaic (PV) plant at the national gypsum facility in Mount Holly,” North
Carolina, USA, 2011.
[13] National Fire Protection Association, 70(R): National electrical code (R). Quincy,
MA, USA: (NFPA) National Fire Protection Association, 2014.
[14] UL 1699B - Outline of investigation for photovoltaic (PV) DC arc-fault circuit
protection. Underwriters Laboratories, 2013.
[15] UL 1699B - Standard for Photovoltaic (PV) DC Arc-Fault Circuit Protection.
Underwriters Laboratories, 2018.
[16] C. Strobl and P. Meckler, “Arc faults in photovoltaic systems,” in Annual Holm
Conference on Electrical Contacts, 2010, pp. 216–222, doi:
10.1109/HOLM.2010.5619538.
[17] J. Yuventi, “DC electric arc-flash hazard-risk evaluations for photovoltaic
systems,” IEEE Trans. Power Deliv., vol. 29, no. 1, pp. 161–167, 2014, doi:
10.1109/TPWRD.2013.2289921.
[18] S. Dhar, R. K. Patnaik, and P. K. Dash, “Fault Detection and Location of
Photovoltaic Based DC Microgrid Using Differential Protection Strategy,” IEEE
Trans. Smart Grid, vol. 9, no. 5, pp. 4303–4312, Sep. 2018, doi:
10.1109/TSG.2017.2654267.
[19] S. McCalmont, “Low Cost Arc Fault Detection and Protection for PV Systems,”
Golden, CO (United States), Oct. 2013. doi: 10.2172/1110454.
[20] X. Yao, J. Wang, and D. L. Schweickart, “Review and recent developments in
DC arc fault detection,” in IEEE International Power Modulator and High
Voltage Conference (IPMHVC), Aug. 2017, pp. 467–472, doi:
10.1109/IPMHVC.2016.8012887.
[21] M. K. Alam, F. Khan, J. Johnson, and J. Flicker, “A Comprehensive Review of
Catastrophic Faults in PV Arrays: Types, Detection, and Mitigation Techniques,”
IEEE J. Photovoltaics, vol. 5, no. 3, pp. 982–997, May 2015, doi:
10.1109/JPHOTOV.2015.2397599.
191
[22] M. G. Villalva, J. R. Gazoli, and E. R. Filho, “Comprehensive approach to
modeling and simulation of photovoltaic arrays,” IEEE Trans. Power Electron.,
vol. 24, no. 5, pp. 1198–1208, 2009, doi: 10.1109/TPEL.2009.2013862.
[23] G. Petrone, C. A. Ramos-Paja, G. Spagnuolo, and Weidong Xiao, Photovoltaic
sources modeling. Chichester, West Sussex, UK: John Wiley & Sons, 2017.
[24] F. Reil, A. Sepanski, W. Herrmann, J. Althaus, W. Vaaßen, and H. Schmidt,
“Qualification of arcing risks in PV modules,” in IEEE Photovoltaic Specialists
Conference (PVSC), 2012, pp. 727–730, doi: 10.1109/PVSC.2012.6317709.
[25] J. Flicker and J. Johnson, “Electrical simulations of series and parallel PV arc-
faults,” in IEEE Photovoltaic Specialists Conference (PVSC), 2013, pp. 3165–
3172, doi: 10.1109/PVSC.2013.6745127.
[26] Z. Wang and R. S. Balog, “Arc Fault and Flash Signal Analysis in DC
Distribution Systems Using Wavelet Transformation,” IEEE Trans. Smart Grid,
vol. 6, no. 4, pp. 1955–1963, Jul. 2015, doi: 10.1109/TSG.2015.2407868.
[27] J. Strauch, M. Quintana, J. Granata, W. Bower, and S. Kuszmaul, “Solar module
arc fault modeling at Sandia National Laboratories,” in 2011 NREL Module
Reliability Workshop, 2010.
[28] J. K. Hastings, M. A. Juds, C. J. Luebke, and B. Pahl, “A study of ignition time
for materials exposed to DC arcing in PV systems,” in IEEE Photovoltaic
Specialists Conference (PVSC), 2011, pp. 3724–3729, doi:
10.1109/PVSC.2011.6185959.
[29] Z. Wang, S. McConnell, R. S. Balog, and J. Johnson, “Arc fault signal detection -
Fourier transformation vs. wavelet decomposition techniques using synthesized
data,” in IEEE Photovoltaic Specialists Conference (PVSC), Oct. 2014, pp.
3239–3244, doi: 10.1109/PVSC.2014.6925625.
[30] K. M. Armijo, J. Johnson, M. Hibbs, and A. Fresquez, “Characterizing fire
danger from low-power photovoltaic arc-faults,” in IEEE Photovoltaic
Specialists Conference (PVSC), Oct. 2014, pp. 3384–3390, doi:
10.1109/PVSC.2014.6925658.
[31] F. Erhard, B. Schaller, and F. Berger, “Field test results of serial DC arc fault
investigationson real photovoltaic systems,” in Universities Power Engineering
192
Conference (UPEC), Oct. 2014, doi: 10.1109/UPEC.2014.6934689.
[32] S. C. Wang, C. J. Wu, and Y. J. Wang, “Detection of arc fault on low voltage
power circuits in time and frequency domain approach,” Int. J. Circuits, Syst.
Signal Process., vol. 6, no. 5, pp. 324–331, 2012.
[33] Z. Meng, L. Wang, and Q. Sun, “The characteristics of DC arc faults current,” in
European Conference on Power Electronics and Applications (EPE), 2013, doi:
10.1109/EPE.2013.6631914.
[34] J. Johnson and K. Armijo, “Parametric study of PV arc-fault generation methods
and analysis of conducted DC spectrum,” in IEEE Photovoltaic Specialists
Conference (PVSC), Oct. 2014, pp. 3543–3548, doi:
10.1109/PVSC.2014.6924874.
[35] J. Johnson, S. S. Kuszmaul, W. I. Bower, and D. A. Schoenwald, “Using PV
module and line frequency response data to create robust arc fault detectors,” in
European Photovoltaic Solar Energy Conference, 2011, pp. 3745–3750.
[36] B. England, “An Investigation into Arc Detection and Fire Safety Aspects of
Photovoltaic Installations,” Murdoch University, Perth, WA, Australia, 2012.
[37] J. Johnson, K. M. Armijo, M. Avrutsky, D. Eizips, and S.Kondrashov, “Arc-
fault unwanted tripping survey with UL 1699B-listed products,” in IEEE
Photovoltaic Specialists Conference (PVSC), Dec. 2015, doi:
10.1109/PVSC.2015.7356427.
[38] J. Johnson, C. Oberhauser, M. Montoya, A. Fresquez, S. Gonzalez, and A. Patel,
“Crosstalk nuisance trip testing of photovoltaic DC arc-fault detectors,” in IEEE
Photovoltaic Specialists Conference (PVSC), 2012, pp. 1383–1387, doi:
10.1109/PVSC.2012.6317857.
[39] S. Chen, X. Li, and J. Xiong, “Series Arc Fault Identification for Photovoltaic
System Based on Time-Domain and Time-Frequency-Domain Analysis,” IEEE J.
Photovoltaics, vol. 7, no. 4, pp. 1105–1114, Jul. 2017, doi:
10.1109/JPHOTOV.2017.2694421.
[40] J. Schmid, E. Kancsar, M. Drapalik, V. Schlosser, and G. Klinger, “A Study of
the Current Disturbance Caused by Wind Induced Vibrations of Photovoltaic
Modules,” in International Conference on Renewable Energies and Power
193
Quality (ICREPQ), 2010, pp. 135–140, doi: 10.24084/repqj08.256.
[41] A. D. Stokes, “Electric arcs in open air,” J. Phys. D. Appl. Phys., vol. 24, pp. 26–
35, 1991.
[42] J. J. Lowke, “Simple theory of free burning arc,” J. Phys. D Appl., vol. 12, pp.
1873–1886, 1979.
[43] R. F. Ammerman, T. Gammon, P. K. Sen, and J. P. Nelson, “DC-arc models and
incident-energy calculations,” IEEE Trans. Ind. Appl., vol. 46, no. 5, pp. 1810–
1819, Sep. 2010, doi: 10.1109/TIA.2010.2057497.
[44] H. B. Estes, “Horizontal series fault comparison in AC & DC micro-grid
architectures,” University of Texas at Austin, Austin, TX, USA, 2011.
[45] N. Gustavsson, “Evaluation and Simulation of Black-box Arc Models for High
Voltage Circuit-breakers,” Linköping University, Linköping, Sweden, 2004.
[46] M. Ju and L. Wang, “Arc fault modeling and simulation in DC system based on
Habedank model,” in Prognostics and System Health Management Conference
(PHM), Jan. 2017, doi: 10.1109/PHM.2016.7819827.
[47] L. van der Sluis, W. R. Rutgers, and C. G. A. Koreman, “A physical ARC model
for the simulation of current zero behavior of high-voltage circuit breakers,”
IEEE Trans. Power Deliv., vol. 7, no. 2, pp. 1016–1022, 1992, doi:
10.1109/61.127112.
[48] W. B. Nottingham, “A new equation for the static characteristic of the normal
electric arc,” J. Am. Inst. Electr. Eng., vol. 42, no. 1, pp. 12–19, Oct. 2013, doi:
10.1109/joaiee.1923.6591851.
[49] D. B. Miller and J. L. Hildenbrand, “DC Arc Model including Circuit
Constraints,” IEEE Trans. Power Appar. Syst., vol. PAS-92, no. 6, pp. 1926–
1934, 1973, doi: 10.1109/TPAS.1973.293572.
[50] NFPA, NFPA 70E: Standard for Electrical Safety in the Workplace. Quincy, MA,
USA: National Fire Protection Association, 2015.
[51] J. Paukert, “The arc voltage and arc resistance of LV fault arcs,” in international
symposium on switching arc phenomena, 1993, pp. 49–51.
[52] X. Yao, L. Herrera, S. Ji, K. Zou, and J. Wang, “Characteristic study and time-
194
domain discrete-wavelet-transform based hybrid detection of series DC arc
faults,” IEEE Trans. Power Electron., vol. 29, no. 6, pp. 3103–3115, Jun. 2014,
doi: 10.1109/TPEL.2013.2273292.
[53] F. M. Uriarte et al., “A DC arc model for series faults in low voltage microgrids,”
IEEE Trans. Smart Grid, vol. 3, no. 4, pp. 2063–2070, 2012, doi:
10.1109/TSG.2012.2201757.
[54] M. Weerasekara, M. Vilathgamuwa, and Y. Mishra, “Modelling of DC arcs for
photovoltaic system faults,” in IEEE Annual Southern Power Electronics
Conference (SPEC), 2016, doi: 10.1109/SPEC.2016.7846061.
[55] V. V. Terzija, M. Popov, V. Stanojevic, and Z. Radojevic, “EMTP simulation
and spectral domain features of a long arc in free air,” in International
Conference and Exhibition on Electricity Distribution (CIRED), 2005, pp. 389–
392, doi: 10.1049/cp:20050953.
[56] X. Yao, L. Herrera, and J. Wang, “Impact evaluation of series dc arc faults in dc
microgrids,” in IEEE Applied Power Electronics Conference and Exposition
(APEC), May 2015, pp. 2953–2958, doi: 10.1109/APEC.2015.7104771.
[57] J. Johnson et al., “Photovoltaic DC arc fault detector testing at Sandia National
Laboratories,” in IEEE Photovoltaic Specialists Conference (PVSC), 2011, pp.
3614–3619, doi: 10.1109/PVSC.2011.6185930.
[58] N. L. Georgijevic, M. V. Jankovic, S. Srdic, and Z. Radakovic, “The detection of
series arc fault in photovoltaic systems based on the arc current entropy,” IEEE
Trans. Power Electron., vol. 31, no. 8, pp. 5917–5930, Aug. 2016, doi:
10.1109/TPEL.2015.2489759.
[59] J. C. Kim, B. Lehman, and R. Ball, “DC Arc Fault Model Superimposing
Multiple Random Arc Noise States on an Average Model,” in IEEE Workshop on
Control and Modeling for Power Electronics (COMPEL), Jun. 2019, doi:
10.1109/COMPEL.2019.8769649.
[60] X. Shen, “on-line monitoring of arcing faults in medium voltage network,”
University of New South Wales, Sydney, NSW, Australia, 2016.
[61] B. Novak, “Implementing arc detection in solar applications: achieving
compliance with the new UL 1699B Standard,” 2012.
195
[62] J. Johnson and J. Kang, “Arc-fault detector algorithm evaluation method utilizing
prerecorded arcing signatures,” in IEEE Photovoltaic Specialists Conference
(PVSC), 2012, pp. 1378–1382, doi: 10.1109/PVSC.2012.6317856.
[63] Y. Ohta and H. Isoda, “Arc detecting device and aircraft equipped therewith,”
U.S. Patent 8,093,904, 2012.
[64] Y. Cao, J. Li, M. Sumner, E. Christopher, and D. Thomas, “Arc fault generation
and detection in DC systems,” in Asia-Pacific Power and Energy Engineering
Conference (APPEEC), 2013, doi: 10.1109/APPEEC.2013.6837123.
[65] F. Eger, “Arcing signals in PV power plants - Evaluation of AFD test standards
and alternative setups based on in-field arc measurements,” in IEEE Photovoltaic
Specialists Conference (PVSC), Nov. 2016, pp. 1250–1253, doi:
10.1109/PVSC.2016.7749814.
[66] J. Johnson, D. Schoenwald, S. Kuszmaul, J. Strauch, and W. Bower, “Creating
dynamic equivalent PV circuit models with impedance spectroscopy for arc fault
modeling,” in IEEE Photovoltaic Specialists Conference (PVSC), 2011, pp.
2328–2333, doi: 10.1109/PVSC.2011.6186419.
[67] M. I. Fitrianto, E. Wahjono, D. O. Anggriawan, E. Prasetyono, R. H. Mubarok,
and A. Tjahjono, “Identification and Protection of Series DC Arc Fault for
Photovoltaic Systems Based on Fast Fourier Transform,” in International
Electronics Symposium (IES), Sep. 2019, pp. 159–163, doi:
10.1109/ELECSYM.2019.8901605.
[68] K. Xia, Z. He, Y. Yuan, Y. Wang, and P. Xu, “An arc fault detection system for
the household photovoltaic inverter according to the DC bus currents,” in
International Conference on Electrical Machines and Systems (ICEMS), Jan.
2016, pp. 1687–1690, doi: 10.1109/ICEMS.2015.7385312.
[69] J. C. Gu, D. S. Lai, J. M. Wang, J. J. Huang, and M. T. Yang, “Design of a DC
series arc fault detector for photovoltaic system protection,” IEEE Trans. Ind.
Appl., vol. 55, no. 3, pp. 2464–2471, May 2019, doi: 10.1109/TIA.2019.2894992.
[70] G. S. Seo, H. Bae, B. H. Cho, and K. C. Lee, “Arc protection scheme for DC
distribution systems with photovoltaic generation,” in International Conference
on Renewable Energy Research and Applications (ICRERA), 2012, doi:
196
10.1109/ICRERA.2012.6477374.
[71] C. Aarstad, T. Taufik, A. Kean, and M. Muscarella, “Development of arc fault
interrupter laboratory testing for low voltage DC electricity,” in International
Seminar on Intelligent Technology and Its Application (ISITIA), Jan. 2016, pp.
583–588, doi: 10.1109/ISITIA.2016.7828725.
[72] G. S. Seo, K. A. Kim, K. C. Lee, K. J. Lee, and B. H. Cho, “A new DC arc fault
detection method using DC system component modeling and analysis in low
frequency range,” in IEEE Applied Power Electronics Conference and
Exposition (APEC), May 2015, pp. 2438–2444, doi:
10.1109/APEC.2015.7104690.
[73] M. Kanemaru, K. Kokura, M. Mori, T. Shindoi, and M. Yamamoto,“Identification technique of DC series arc-fault strings in photovoltaic systems,”
Electr. Eng. Japan (English Transl. Denki Gakkai Ronbunshi), vol. 207, no. 2, pp.
12–19, 2019, doi: 10.1002/eej.23204.
[74] S. Chae, J. Park, and S. Oh, “Series DC Arc Fault Detection Algorithm for DC
Microgrids Using Relative Magnitude Comparison,” IEEE J. Emerg. Sel. Top.
Power Electron., vol. 4, no. 4, pp. 1270–1278, Dec. 2016, doi:
10.1109/JESTPE.2016.2592186.
[75] W. Miao, X. Liu, K. H. Lam, and P. W. T. Pong, “Arc-Faults Detection in PV
Systems by Measuring Pink Noise with Magnetic Sensors,” IEEE Trans. Magn.,
vol. 55, no. 7, Jul. 2019, doi: 10.1109/TMAG.2019.2903899.
[76] W. Miao, X. Liu, K. H. Lam, and P. W. T. Pong, “DC-arcing detection by noise
measurement with magnetic sensing by TMR sensors,” IEEE Trans. Magn., vol.
54, no. 11, 2018, doi: 10.1109/TMAG.2018.2842187.
[77] M. Wendl, M. Weiss, and F. Berger, “HF Characterization of Low Current DC
Arcs at Alterable Conditions,” in International Conference on Electrical
Contacts (ICEC), 2014, pp. 439–444.
[78] J. Johnson et al., “Differentiating series and parallel photovoltaic arc-faults,” in
IEEE Photovoltaic Specialists Conference (PVSC), 2012, pp. 720–726, doi:
10.1109/PVSC.2012.6317708.
[79] L. Yuan, J. Shengchang, W. Jin, Y. Xiu, and Z. Yeye, “Study on characteristics
197
and detection of DC arc fault in power electronics system,” in IEEE International
Conference on Condition Monitoring and Diagnosis (CMD), 2012, pp. 1043–
1046, doi: 10.1109/CMD.2012.6416335.
[80] T. A. Ramamurthy and K. S. Swarup, “High Impedance Fault detection using
DWT for transmission and distribution networks,” in IEEE International
Conference on Power Systems (ICPS), Oct. 2016, doi:
10.1109/ICPES.2016.7584004.
[81] S. Chen and X. Li, “PV series arc fault recognition under different working
conditions with joint detection method,” in Annual Holm Conference on
Electrical Contacts, Dec. 2016, pp. 25–32, doi: 10.1109/HOLM.2016.7780002.
[82] S. V. Narasimhan, N. Basumallick, and S. Veena., Introduction to wavelet
transform: a signal processing approach. Alpha Science International, Ltd, 2011.
[83] G. Yunmei, W. Li, W. Zhuoqi, and J. Binfeng, “Wavelet packet analysis applied
in detection of low-voltage DC arc fault,” in IEEE Conference on Industrial
Electronics and Applications (ICIEA), 2009, pp. 4013–4016, doi:
10.1109/ICIEA.2009.5138962.
[84] Z. Wang and R. S. Balog, “Arc fault and flash detection in DC photovoltaic
arrays using wavelets,” in IEEE Photovoltaic Specialists Conference (PVSC),
2013, pp. 1619–1624, doi: 10.1109/PVSC.2013.6744455.
[85] X. Yao, L. Herrera, and J. Wang, “A series DC arc fault detection method and
hardware implementation,” in IEEE Applied Power Electronics Conference and
Exposition (APEC), 2013, pp. 2444–2449, doi: 10.1109/APEC.2013.6520638.
[86] W. Li, A. Monti, and F. Ponci, “Fault detection and classification in medium
voltage dc shipboard power systems with wavelets and artificial neural networks,”
IEEE Trans. Instrum. Meas., vol. 63, no. 11, pp. 2651–2665, Nov. 2014, doi:
10.1109/TIM.2014.2313035.
[87] R. S. Balog, “Method and system for detecting arc faults and flashes using
wavelets,” U.S. Patent 9,329,220, 2016.
[88] Y. Zhang, L. Wang, and S. Yang, “Research on Characteristics of DC Arc Fault
Based on Wavelet Transform,” in IEEE International Conference on Electrical
Systems for Aircraft, Railway, Ship Propulsion and Road Vehicles and
198
International Transportation Electrification Conference (ESARS-ITEC), Jan.
2019, doi: 10.1109/ESARS-ITEC.2018.8607619.
[89] H. Zhu, Z. Wang, and R. S. Balog, “Real time arc fault detection in PV systems
using wavelet decomposition,” in IEEE Photovoltaic Specialists Conference
(PVSC), Nov. 2016, pp. 1761–1766, doi: 10.1109/PVSC.2016.7749926.
[90] C. He, L. Mu, and Y. Wang, “The Detection of Parallel Arc Fault in Photovoltaic
Systems Based on a Mixed Criterion,” IEEE J. Photovoltaics, vol. 7, no. 6, pp.
1717–1724, Nov. 2017, doi: 10.1109/JPHOTOV.2017.2742143.
[91] S. Chen, X. Li, Y. Meng, and Z. Xie, “Wavelet-based protection strategy for
series arc faults interfered by multicomponent noise signals in grid-connected
photovoltaic systems,” Sol. Energy, vol. 183, pp. 327–336, 2019, doi:
10.1016/j.solener.2019.03.008.
[92] M. Dargatz and M. Fornage, “Method and apparatus for detection and control of
dc arc faults,” U.S. Patent 8,179,147, 2012.
[93] D. Kilrey and W. Oldenburg, “DC arc fault detection and protection,” U.S. Patent
0133135A1, 2007.
[94] H. Braun et al., “Signal processing for fault detection in photovoltaic arrays,” in
IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), 2012, pp. 1681–1684, doi: 10.1109/ICASSP.2012.6288220.
[95] S. Buddha et al., “Signal processing for photovoltaic applications,” in IEEE
International Conference on Emerging Signal Processing Applications (ESPA),
2012, pp. 115–118, doi: 10.1109/ESPA.2012.6152459.
[96] F. Schimpf and L. E. Narum, “Recognition of electric arcing in the DC-wiring of
photovoltaic systems,” in International Telecommunications Energy Conference
(INTELEC), 2009, doi: 10.1109/INTLEC.2009.5352037.
[97] Y. Gao, J. Zhang, Y. Lin, and Y. Sun, “An innovative photovoltaic DC arc fault
detection method through multiple criteria algorithm based on a new arc
initiation method,” in IEEE Photovoltaic Specialists Conference (PVSC), Oct.
2014, pp. 3188–3192, doi: 10.1109/PVSC.2014.6925613.
[98] Q. Lu, Z. Ye, M. Su, Y. Li, Y. Sun, and H. Huang, “A DC series arc fault
detection method using line current and supply voltage,” IEEE Access, vol. 8, pp.
199
10134–10146, 2020, doi: 10.1109/ACCESS.2019.2963500.
[99] C. Strobl, “Arc fault detection in DC microgrids,” in IEEE 1st International
Conference on Direct Current Microgrids (ICDCM), Jul. 2015, pp. 181–186, doi:
10.1109/ICDCM.2015.7152035.
[100] C. Strobl, “Arc fault detection – a model-based approach,” in international
conference on electrical contacts, 2014, pp. 367–373.
[101] A. R. Jordehi, “Parameter estimation of solar photovoltage (PV) cells: A review,”
Renew. Sustain. Energy Rev., vol. 61, pp. 354–368, 2016.
[102] D. T. Cotfas, P. A. Cotfas, and S. Kaplanis, “Methods and techniques to
determine the dynamic parameters of solar cells: Review,” Renew. Sustain.
Energy Rev., vol. 61, pp. 213–221, 2016, doi: 10.1016/j.rser.2016.03.051.
[103] R. D. Telford, S. Galloway, B. Stephen, and I. Elders, “Diagnosis of series DC
Arc faults - A machine learning approach,” IEEE Trans. Ind. Informatics, vol. 13,
no. 4, pp. 1598–1609, Aug. 2017, doi: 10.1109/TII.2016.2633335.
[104] J. A. Momoh and R. Button, “Design and Analysis of Aerospace DC Arcing
Faults using Fast Fourier Transformation and Artificial Neural Network,” in
IEEE Power Engineering Society General Meeting, 2003, pp. 788–793, doi:
10.1109/pes.2003.1270407.
[105] J. Yang and Y. Wang, “Identification and Detection of DC Arc Fault in
Photovoltaic Power Generation System,” in International Conference on
Intelligent Transportation, Big Data & Smart City (ICITBS), Jan. 2020, pp. 440–
444, doi: 10.1109/ICITBS49701.2020.00095.
[106] Z. Wang and R. S. Balog, “Arc fault and flash detection in photovoltaic systems
using wavelet transform and support vector machines,” in IEEE Photovoltaic
Specialists Conference (PVSC), Nov. 2016, pp. 3275–3280, doi:
10.1109/PVSC.2016.7750271.
[107] K. Yang, R. Zhang, J. Yang, C. Liu, S. Chen, and F. Zhang, “A Novel Arc Fault
Detector for Early Detection of Electrical Fires,” Sensors, vol. 16, no. 4, p. 500,
Apr. 2016, doi: 10.3390/s16040500.
[108] K. Xia, S. He, Y. Tan, Q. Jiang, J. Xu, and W. Yu, “Wavelet packet and support
vector machine analysis of series DC ARC fault detection in photovoltaic system,”
200
IEEJ Trans. Electr. Electron. Eng., vol. 14, no. 2, pp. 192–200, Feb.2019, doi:
10.1002/tee.22797.
[109] K. Xia, H. Guo, S. He, W. Yu, J. Xu, and H. Dong, “Binary classification model
based on machine learning algorithm for the DC serial arc detection in electric
vehicle battery system,” IET Power Electron., vol. 12, no. 1, pp. 112–119, Jan.
2019, doi: 10.1049/iet-pel.2018.5789.
[110] K. Xia et al., “Wavelet entropy analysis and machine learning classification
model of DC serial arc fault in electric vehicle power system,” IET Power
Electron., vol. 12, no. 15, pp. 3998–4004, Dec. 2019, doi: 10.1049/iet-
pel.2019.0375.
[111] Z. Yin, L. Wang, Y. Zhang, Y. Gao, and S. Yang, “The DC Arc Fault Detection
Method Taken Advantage of WT and MFE,” in Prognostics and System Health
Management Conference (PHM), Oct. 2019, doi: 10.1109/PHM-
Qingdao46334.2019.8942846.
[112] B. Grichting, J. Goette, and M. Jacomet, “Cascaded fuzzy logic based arc fault
detection in photovoltaic applications,” in International Conference on Clean
Electrical Power: Renewable Energy Resources Impact (ICCEP), Aug. 2015, pp.
178–183, doi: 10.1109/ICCEP.2015.7177620.
[113] Y. Zhao, R. Ball, J. Mosesian, J.-F. de Palma, and B. Lehman, “Graph-based
detection rules for fault detection in solar photovoltaic arrays,” IEEE Trans.
Power Electron., vol. 30, no. 5, pp. 2848–2858, 2015.
[114] Y. Zhao, L. Yang, B. Lehman, J. F. De Palma, J. Mosesian, and R. Lyons,
“Decision tree-based fault detection and classification in solar photovoltaic
arrays,” in IEEE Applied Power Electronics Conference and Exposition (APEC),
2012, pp. 93–99, doi: 10.1109/APEC.2012.6165803.
[115] V. Le, X. Yao, C. Miller, and B. H. Tsao, “Series DC Arc Fault Detection Based
on Ensemble Machine Learning,” IEEE Trans. Power Electron., vol. 35, no. 8,
pp. 7826–7839, Aug. 2020, doi: 10.1109/TPEL.2020.2969561.
[116] V. Le, X. Yao, C. Miller, and T. B. Hung, “Arc fault detection in DC distribution
using semi-supervised ensemble machine learning,” in IEEE Energy Conversion
Congress and Exposition (ECCE), Sep. 2019, pp. 2939–2945, doi:
201
10.1109/ECCE.2019.8913286.
[117] M. Rabla, E. Tisserand, P. Schweitzer, and J. Lezama, “Arc fault analysis and
localization by cross-correlation in 270 V DC,” in IEEE 59th Holm Conference
on Electrical Contacts, 2013, pp. 1–6.
[118] M. Ahmadi, H. Samet, and T. Ghanbari, “Series Arc Fault Detection in
Photovoltaic Systems Based on Signal-to-Noise Ratio Characteristics Using
Cross-Correlation Function,” IEEE Trans. Ind. Informatics, vol. 16, no. 5, pp.
3198–3209, May 2020, doi: 10.1109/TII.2019.2909753.
[119] M. Ahmadi, H. Samet, and T. Ghanbari, “Kalman filter–based approach for
detection of series arc fault in photovoltaic systems,” Int. Trans. Electr. Energy
Syst., vol. 29, no. 5, p. e2823, May 2019, doi: 10.1002/2050-7038.2823.
[120] K. Gajula and L. Herrera, “Detection and Localization of Series Arc Faults in DC
Microgrids using Kalman Filter,” IEEE J. Emerg. Sel. Top. Power Electron., pp.
1–1, Apr. 2020, doi: 10.1109/jestpe.2020.2987491.
[121] N. L. Georgijevic, D. Stojic, and Z. Radakovic, “Series arc fault detection in
photovoltaic system by small-signal impedance and noise monitoring,” Int. Trans.
Electr. Energy Syst., vol. 30, no. 2, pp. 1–15, 2020, doi: 10.1002/2050-
7038.12234.
[122] M. K. Alam, F. H. Khan, J. Johnson, and J. Flicker, “PV arc-fault detection using
spread spectrum time domain reflectometry (SSTDR),” in IEEE Energy
Conversion Congress and Exposition (ECCE), Nov. 2014, pp. 3294–3300, doi:
10.1109/ECCE.2014.6953848.
[123] H. L. Lan and R. C. Zhang, “Current research and development trends on faults
arc detection method in switch cabinet,” Gaodianya Jishu/High Volt. Eng., vol.
34, no. 3, pp. 496–499, Mar. 2008.
[124] Q. Xiong, S. Ji, L. Zhu, L. Zhong, and Y. Liu, “A novel DC arc fault detection
method based on electromagnetic radiation signal,” IEEE Trans. Plasma Sci., vol.
45, no. 3, pp. 472–478, Mar. 2017, doi: 10.1109/TPS.2017.2653817.
[125] C. J. Kim, “Electromagnetic radiation behavior of low-voltage arcing fault,”
IEEE Trans. Power Deliv., vol. 24, no. 1, pp. 416–423, 2009, doi:
10.1109/TPWRD.2008.2002873.
202
[126] J. J. Shea, C. J. Luebke, and K. L. Parker, “RF current produced from DC
electrical arcing,” in international conference on electrical contacts, 2012, pp. 1–
6, doi: 10.1049/cp.2012.0612.
[127] Q. Xiong et al., “Electromagnetic Radiation Characteristics of Series DC Arc
Fault and Its Determining Factors,” IEEE Trans. Plasma Sci., vol. 46, no. 11, pp.
4028–4036, Nov. 2018, doi: 10.1109/TPS.2018.2864605.
[128] L. Zhu, J. Li, Y. Liu, and S. Ji, “Initial features of the unintended atmospheric
pressure dc arcs and their application on the fault detection,” IEEE Trans.
Plasma Sci., vol. 45, no. 4, pp. 742–748, Apr. 2017, doi:
10.1109/TPS.2017.2676821.
[129] G. S. Seo, J. I. Ha, B. H. Cho, and K. C. Lee, “Series arc fault detection method
based on statistical analysis for dc Microgrids,” in IEEE Applied Power
Electronics Conference and Exposition (APEC), May 2016, pp. 487–492, doi:
10.1109/APEC.2016.7467916.
[130] L. Ryves, D. R. McKenzie, and M. M. M. Bilek, “Cathode-spot dynamics in a
high-current pulsed arc: A noise study,” IEEE Trans. Plasma Sci., vol. 37, no. 2,
pp. 365–368, 2009, doi: 10.1109/TPS.2008.2007734.
[131] S. Lu, B. T. Phung, and D. Zhang, “A comprehensive review on DC arc faults
and their diagnosis methods in photovoltaic systems,” Renew. Sustain. Energy
Rev., vol. 89, pp. 88–98, Jun. 2018, doi: 10.1016/j.rser.2018.03.010.
[132] P. G. Slade, “The Arc and Interruption,” in Electrical Contact: Principles and
Applications, 2nd Editio., CRC Press, 2013, pp. 554–616.
[133] A. I. Aio, “Modelization and analysis of the electric arc in low voltage circuit
breakers,” University of The Basque Country, Spain, 2013.
[134] D. Hanbay, I. Turkoglu, and Y. Demir, “Prediction of wastewater treatment plant
performance based on wavelet packet decomposition and neural networks,”
Expert Syst. Appl., vol. 34, no. 2, pp. 1038–1043, Feb. 2008, doi:
10.1016/j.eswa.2006.10.030.
[135] V. Kecman, Support Vector Machines: Theory and Applications. Springer
Science & Business Media, 2005.
[136] I. Goodfellow, B. Yoshua, and C. Aaron, Deep Learning. MIT Press, 2016.
203
[137] Y. Lecun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553,
pp. 436–444, May 2015, doi: 10.1038/nature14539.
[138] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network
training by reducing internal covariate shift,” in International Conference on
Machine Learning (ICML), Feb. 2015, pp. 448–456.
[139] V. Nair and G. E. Hinton, “Rectified Linear Units Improve Restricted Boltzmann
Machines,” in international conference on machine learning (ICML), 2010, pp.
807–814.
[140] A. Graves and J. Schmidhuber, “Framewise phoneme classification with
bidirectional LSTM and other neural network architectures,” Neural Networks,
vol. 18, no. 5–6, pp. 602–610, Jul. 2005, doi: 10.1016/j.neunet.2005.06.042.
[141] J. P. Ram, H. Manghani, D. S. Pillai, T. S. Babu, M. Miyatake, and N. Rajasekar,
“Analysis on solar PV emulators: A review,” Renew. Sustain. Energy Rev., vol.
81, pp. 149–160, Jan. 2018, doi: 10.1016/j.rser.2017.07.039.
[142] X. Jiang, J. Sun, C. Li, and H. Ding, “Video Image Defogging Recognition
Based on Recurrent Neural Network,” IEEE Trans. Ind. Informatics, vol. 14, no.
7, pp. 3281–3288, Jul. 2018, doi: 10.1109/TII.2018.2810188.
[143] Z. Liu, Z. Wu, T. Li, J. Li, and C. Shen, “GMM and CNN Hybrid Method for
Short Utterance Speaker Recognition,” IEEE Trans. Ind. Informatics, vol. 14, no.
7, pp. 3244–3252, Jul. 2018, doi: 10.1109/TII.2018.2799928.
[144] B. Luo, H. Wang, H. Liu, B. Li, and F. Peng, “Early Fault Detection of Machine
Tools Based on Deep Learning and Dynamic Identification,” IEEE Trans. Ind.
Electron., vol. 66, no. 1, pp. 509–518, Jan.2018, doi: 10.1109/TIE.2018.2807414.
[145] Z. Zeng, Z. Li, D. Cheng, H. Zhang, K. Zhan, and Y. Yang, “Two-Stream
Multirate Recurrent Neural Network for Video-Based Pedestrian
Reidentification,” IEEE Trans. Ind. Informatics, vol. 14, no. 7, pp. 3179–3186,
Jul. 2018, doi: 10.1109/TII.2017.2767557.
[146] C. Sun, M. Ma, Z. Zhao, and X. Chen, “Sparse Deep Stacking Network for Fault
Diagnosis of Motor,” IEEE Trans. Ind. Informatics, vol. 14, no. 7, pp. 3261–
3270, Jul. 2018, doi: 10.1109/TII.2018.2819674.
[147] M. F. Guo, X. D. Zeng, D. Y. Chen, and N. C. Yang, “Deep-Learning-Based
204
Earth Fault Detection Using Continuous Wavelet Transform and Convolutional
Neural Network in Resonant Grounding Distribution Systems,” IEEE Sens. J.,
vol. 18, no. 3, pp. 1291–1300, Feb. 2018, doi: 10.1109/JSEN.2017.2776238.
[148] G. Jiang, H. He, J. Yan, and P. Xie, “Multiscale Convolutional Neural Networks
for Fault Diagnosis of Wind Turbine Gearbox,” IEEE Trans. Ind. Electron., vol.
66, no. 4, pp. 3196–3207, Apr. 2019, doi: 10.1109/TIE.2018.2844805.
[149] M. Zhao, M. Kang, B. Tang, and M. Pecht, “Deep Residual Networks with
Dynamically Weighted Wavelet Coefficients for Fault Diagnosis of Planetary
Gearboxes,” IEEE Trans. Ind. Electron., vol. 65, no. 5, pp. 4290–4300, May
2018, doi: 10.1109/TIE.2017.2762639.
[150] J. Wang, R. Zhao, D. Wang, R. Yan, K. Mao, and F. Shen, “Machine health
monitoring using local feature-based gated recurrent unit networks,” IEEE Trans.
Ind. Electron., vol. 65, no. 2, pp. 1539–1548, Jul. 2017, doi:
10.1109/TIE.2017.2733438.
[151] S. Chen, X. Li, Z. Xie, and Y. Meng, “Time-Frequency Distribution
Characteristic and Model Simulation of Photovoltaic Series Arc Fault With
Power Electronic Equipment,” IEEE J. Photovoltaics, vol. 9, no. 4, pp. 1128–
1137, Jul. 2019, doi: 10.1109/JPHOTOV.2019.2915337.
[152] M. Ahmadi, H. Samet, and T. Ghanbari, “Series Arc Fault Detection in
Photovoltaic Systems Based on Signal-to-Noise Ratio Characteristics Using
Cross-Correlation Function,” IEEE Trans. Ind. Informatics, vol. 16, no. 5, pp.
3198–3209, May 2020, doi: 10.1109/TII.2019.2909753.
[153] L. Wen, X. Li, L. Gao, and Y. Zhang, “A New Convolutional Neural Network-
Based Data-Driven Fault Diagnosis Method,” IEEE Trans. Ind. Electron., vol. 65,
no. 7, pp. 5990–5998, Jul. 2018, doi: 10.1109/TIE.2017.2774777.
[154] R. Liu, G. Meng, B. Yang, C. Sun, and X. Chen, “Dislocated Time Series
Convolutional Neural Architecture: An Intelligent Fault Diagnosis Approach for
Electric Machine,” IEEE Trans. Ind. Informatics, vol. 13, no. 3, pp. 1310–1320,
Jun. 2017, doi: 10.1109/TII.2016.2645238.
[155] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
applied to document recognition,” Proc. IEEE, vol. 86, no. 11, pp. 2278–2323,
205
1998, doi: 10.1109/5.726791.
[156] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classification with
Deep Convolutional Neural Networks,” 2012. Accessed: Aug. 04, 2020. [Online].
Available: http://code.google.com/p/cuda-convnet/.
[157] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-
Scale Image Recognition,” in International Conference on Learning
Representations (ICLR), 2015, Accessed: Aug. 04, 2020. [Online]. Available:
http://arxiv.org/abs/1409.1556.
[158] A. Khan, A. Sohail, U. Zahoora, and A. S. Qureshi, “A survey of the recent
architectures of deep convolutional neural networks,” Artif. Intell. Rev., pp. 1–62,
Apr. 2020, doi: 10.1007/s10462-020-09825-6.
[159] Y. Wang, F. Zhang, and S. Zhang, “A New Methodology for Identifying Arc
Fault by Sparse Representation and Neural Network,” IEEE Trans. Instrum.
Meas., vol. 67, no. 11, pp. 2526–2537, 2018, doi: 10.1109/TIM.2018.2826878.
[160] S. Lu, H. Chai, A. Sahoo, and B. T. Phung, “Condition Monitoring based on
Partial Discharge Diagnostics using Machine Learning Methods : A
Comprehensive State-of-the-Art Review,” IEEE Trans. Dielectr. Electr. Insul.,
vol. In Press, 2020.
[161] H. Zheng et al., “Cross-Domain Fault Diagnosis Using Knowledge Transfer
Strategy: A Review,” IEEE Access, vol. 7, pp. 129260–129290, 2019, doi:
10.1109/ACCESS.2019.2939876.
[162] S. Lu, B. T. Phung, and D. Zhang, “Study on DC series arc fault in photovoltaic
systems for condition monitoring purpose,” in Australasian Universities Power
Engineering Conference (AUPEC), Feb. 2018, pp. 1–6, doi:
10.1109/AUPEC.2017.8282464.
[163] J. Andrea, D. Jung, P. Schweitzer, B. Vidales, E. Calderon, and S. Weber,
“Simulation of Arcing Fault in PV Panel Network,” in Electrical Contacts,
Proceedings of the Annual Holm Conference on Electrical Contacts, Jan. 2019,
pp. 329–335, doi: 10.1109/HOLM.2018.8611743.
[164] E. Strubell, A. Ganesh, and A. McCallum, “Energy and Policy Considerations for
Deep Learning in NLP,” in Annual Meeting of the Association for Computational
206
Linguistics (ACL), Jun. 2019, pp. 3645–3650, doi: 10.18653/v1/p19-1355.
[165] Q. shi Zhang and S. chun Zhu, “Visual interpretability for deep learning: a
survey,” Front. Inf. Technol. Electron. Eng., vol. 19, no. 1, pp. 27–39, Jan. 2018,
doi: 10.1631/FITEE.1700808.
[166] G. Montavon, W. Samek, and K. R. Müller, “Methods for interpreting and
understanding deep neural networks,” Digit. Signal Process., vol. 73, pp. 1–15,
Feb. 2018, doi: 10.1016/j.dsp.2017.10.011.
[167] T. Sirojan, S. Lu, B. T. Phung, D. Zhang, and E. Ambikairajah, “Sustainable
Deep Learning at Grid Edge for Real-time High Impedance Fault Detection,”
IEEE Trans. Sustain. Comput., Nov. 2018, doi: 10.1109/tsusc.2018.2879960.
[168] T. De Bruin, K. Verbert, and R. Babuska, “Railway Track Circuit Fault
Diagnosis Using Recurrent Neural Networks,” IEEE Trans. Neural Networks
Learn. Syst., vol. 28, no. 3, pp. 523–533, Mar. 2017, doi:
10.1109/TNNLS.2016.2551940.
[169] F. Reil, A. Sepanski, S. Raubach, M. Vosen, and E. Dietrich, “Comparison of
different DC Arc spectra - Derivation of proposals for the development of an
international arc fault detector standard,” in IEEE Photovoltaic Specialists
Conference (PVSC), 2013, pp. 1589–1593, doi: 10.1109/PVSC.2013.6744449.
[170] E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell, “Adversarial discriminative
domain adaptation,” in IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), Nov. 2017, pp. 2962–2971, doi: 10.1109/CVPR.2017.316.
[171] C. Tan, F. Sun, T. Kong, W. Zhang, C. Yang, and C. Liu, “A survey on deep
transfer learning,” in Lecture Notes in Computer Science (including subseries
Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Oct.
2018, pp. 270–279, doi: 10.1007/978-3-030-01424-7_27.
[172] V. M. Patel, R. Gopalan, R. Li, and R. Chellappa, “Visual Domain Adaptation: A
survey of recent advances,” IEEE Signal Process. Mag., vol. 32, no. 3, pp. 53–69,
May 2015, doi: 10.1109/MSP.2014.2347059.
[173] P. Cao, S. Zhang, and J. Tang, “Preprocessing-Free Gear Fault Diagnosis Using
Small Datasets with Deep Convolutional Neural Network-Based Transfer
Learning,” IEEE Access, vol. 6, pp. 26241–26253, May 2018, doi:
207
10.1109/ACCESS.2018.2837621.
[174] L. Guo, Y. Lei, S. Xing, T. Yan, and N. Li, “Deep Convolutional Transfer
Learning Network: A New Method for Intelligent Fault Diagnosis of Machines
with Unlabeled Data,” IEEE Trans. Ind. Electron., vol. 66, no. 9, pp. 7316–7325,
Sep. 2019, doi: 10.1109/TIE.2018.2877090.
[175] L. Wen, L. Gao, and X. Li, “A new deep transfer learning based on sparse auto-
encoder for fault diagnosis,” IEEE Trans. Syst. Man, Cybern. Syst., vol. 49, no. 1,
pp. 136–144, Jan. 2019, doi: 10.1109/TSMC.2017.2754287.
[176] I. J. Goodfellow et al., “Generative Adversarial Nets,” in Conference on Neural
Information Processing Systems (NIPS), 2014, pp. 2672–2680, Accessed: May
31, 2020. [Online]. Available: http://www.github.com/goodfeli/adversarial.
[177] A. Radford, L.Metz, and S. Chintala, “Unsupervised representation learning with
deep convolutional generative adversarial networks,” in International Conference
on Learning Representations (ICLR), 2016.
[178] K. M. Borgwardt, A. Gretton, M. J. Rasch, H.-P. Kriegel, B. Schö Lkopf, and A.
J. Smola, “Integrating structured biological data by Kernel Maximum Mean
Discrepancy,” Bioinformatics, vol. 22, no. 14, pp. 49–57, 2006, doi:
10.1093/bioinformatics/btl242.
[179] S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang, “Domain adaptation via transfer
component analysis,” IEEE Trans. Neural Networks, vol. 22, no. 2, pp. 199–210,
Feb. 2011, doi: 10.1109/TNN.2010.2091281.
[180] W. Lu, B. Liang, Y. Cheng, D. Meng, J. Yang, and T. Zhang, “Deep Model
Based Domain Adaptation for Fault Diagnosis,” IEEE Trans. Ind. Electron., vol.
64, no. 3, pp. 2296–2305, Mar. 2017, doi: 10.1109/TIE.2016.2627020.
[181] L. Van Der Maaten, A. Courville, R. Fergus, and C. Manning, “Accelerating t-
SNE using Tree-Based Algorithms,” J. Mach. Learn. Res., vol. 15, pp. 3221–
3245, 2014, Accessed: May 31, 2020. [Online]. Available:
http://homepage.tudelft.nl/19j49/tsne;
[182] L. Van Der Maaten and G. Hinton, “Visualizing Data using t-SNE,” J. Mach.
Learn. Res., vol. 9, pp. 2579–2605, 2008.
[183] L. Bottou, O. Chapelle, D. DeCoste, and J. Weston, “Support Vector Machine
208
Solvers,” in Large Scale Kernel Machines, 2007, pp. 1–27.
[184] K. He and J. Sun, “Convolutional neural networks at constrained time cost,” in
IEEE Computer Society Conference on Computer Vision and Pattern
Recognition (CVPR), Oct. 2015, pp. 5353–5360, doi:
10.1109/CVPR.2015.7299173.
[185] S. Shao, S. McAleer, R. Yan, and P. Baldi, “Highly Accurate Machine Fault
Diagnosis Using Deep Transfer Learning,” IEEE Trans. Ind. Informatics, vol. 15,
no. 4, pp. 2446–2455, Apr. 2019, doi: 10.1109/TII.2018.2864759.
[186] H. Oh, J. H. Jung, B. C. Jeon, and B. D. Youn, “Scalable and Unsupervised
Feature Engineering Using Vibration-Imaging and Deep Learning for Rotor
System Diagnosis,” IEEE Trans. Ind. Electron., vol. 65, no. 4, pp. 3539–3549,
Apr. 2018, doi: 10.1109/TIE.2017.2752151.
[187] I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville, “Improved
Training of Wasserstein GANs,” in Conference on Neural Information
Processing Systems (NIPS), 2017, [Online]. Available:
https://arxiv.org/abs/1704.00028.
[188] A. Brittomattos, D. Augustoborges Oliveira, and E. Dasilva Morais, “Improving
CNN-Based Viseme Recognition Using Synthetic Data,” in IEEE International
Conference on Multimedia and Expo, Oct. 2018, doi:
10.1109/ICME.2018.8486470.
[189] N. Anantrasirichai, J. Biggs, F. Albino, and D. Bull, “A deep learning approach
to detecting volcano deformation from satellite imagery using synthetic datasets,”
Remote Sens. Environ., vol. 230, p. 111179, Sep. 2019, doi:
10.1016/j.rse.2019.04.032.
[190] H. Arnout, J. Kehrer, J. Bronner, and T. Runkler, “Visual Evaluation of
Generative Adversarial Networks for Time Series Data,” in AAAI Fall
Symposium, 2019, [Online]. Available: https://arxiv.org/abs/2001.00062.
[191] S. Lu, T. Sirojan, B. T. Phung, D. Zhang, and E. Ambikairajah, “DA-DCGAN:
An Effective Methodology for DC Series Arc Fault Diagnosis in Photovoltaic
Systems,” IEEE Access, vol. 7, pp. 45831–45840, 2019, doi:
10.1109/ACCESS.2019.2909267.
209
[192] A. Zhang et al., “Transfer Learning with Deep Recurrent Neural Networks for
Remaining Useful Life Estimation,” Appl. Sci., vol. 8, no. 12, p. 2416, Nov. 2018,
doi: 10.3390/app8122416.
Title Page : INTELLIGENT DC SERIES ARC FAULT DETECTION USING DEEP LEARNING IN PHOTOVOLTAIC SYSTEMS
Acknowledgement
Abstract
Table of contents
List of Acronyms
List of Figures
List of Tables
1. Introduction
1.1. Background and Research Motivation
1.2. Summary of Research Contributions
1.3. Thesis Organisation
1.4. List of Publications
2. Literature Review
2.1. Introduction
2.2. DC Arc in Photovoltaic Systems
2.2.1. Photovoltaic Systems Structure and Arc Hazards
2.2.2. Challenges to Detect DC Arc Faults
2.3. DC Arc Models
2.3.1. Physics-based Arc Model
2.3.2. V-I Characteristic-based Arc Model
2.3.2.1. Nottingham Arc Model
2.3.2.2. Hall, Myer, and Viicheck Arc Model
2.3.2.3. Stokes and Oppenlander Arc Model
2.3.2.4. Paukert Arc Model
2.3.2.5. Modified Paukert Arc Model
2.3.3. Heuristic Arc Model
2.3.4. High-Frequency Variation Caused by Arc Faults
2.4. DC Arc Faults Detection Methods in PV Systems
2.4.1. Sensors for Measurement
2.4.2. Fast Fourier Transform
2.4.3. Short Time Fourier Transform
2.4.4. Wavelet Transform
2.4.5. Statistical Analysis
2.4.6. Model-based Methods
2.4.7. Machine Learning based Methods
2.4.8. Other Types of Methods
2.5. Discussion and Conclusion
3. Characteristics Study on DC Series Arc Fault
3.1. Introduction
3.2. Experimental Setup
3.3. Static Characteristics
3.3.1. V-I Characteristics
3.3.2. Stable Operating Point
3.3.3. Load Current Effect and Source Voltage Effect
3.4. High-frequency Variation in Arc Current
3.4.1. Wavelet Packet Entropy
3.4.2. Effect of Arc Phase
3.4.3. Load Current Effect
3.4.4. Source Voltage Effect
3.4.5. Gap Distance Effect
3.5. Discussion and Conclusion
4. DC Series Arc Fault Detection in PV systems using Deep Learning
4.1. Introduction
4.2. Classical Machine Learning
4.2.1. Artificial Neural Network
4.2.2. Support Vector Machine
4.2.3. Decision Tree and Random Forest
4.2.4. k-Nearest Neighbours
4.2.5. Others
4.3. Deep Learning
4.3.1. Deep Fully-Connected Neural Network
4.3.2. Autoencoder
4.3.3. Convolutional Neural Network
4.3.4. Recurrent Neural Network
4.4. Experimental Setup
4.5. Proposed Deep Learning based Series Arc Fault Detection Method using Convolutional Neural Network
4.5.1. Dataset Preparation
4.5.2. Hyperparameters Setting and Offline Validation Results
4.5.2.1. Size of Filter
4.5.2.2. Number of Filters in the Convolution Layer
4.5.2.3. Number of Convolution Layers
4.5.2.4. Number of Fully Connected Layers and Number of Neurons in Each Layer
4.5.2.5. Comparison with Very Deep CNNs
4.5.3. Evaluation of Different ML Classifiers
4.5.3.1. Datasets Preparation
4.5.3.2. Settings for Different ML Classifiers
4.5.3.3. Results of Comparative Study
4.5.4. Real-time Implementation and Validation Results
4.6. Discussion and Recommendations
4.6.1. Imbalanced Dataset or Small Dataset
4.6.2. Inconsistency between Training and Testing Dataset
4.6.3. Unlabelled Dataset
4.6.4. ML Model Complexity and Real-time Capability
4.6.5. Model Interpretability
4.7. Conclusion
5. Intelligent DC Series Arc Fault Detection in PV Systems using DA-DCGAN without Target-Domain Fault Data
5.1. Introduction
5.2. Experimental Setup
5.2.1. Experimental Setup and Conditions in Source Domain
5.2.2. Experimental Setup and Conditions in Target Domain
5.3. Proposed DA-DCGAN
5.3.1. Generative Adversarial Networks
5.3.2. Optimisation Procedures and Deep Learning Model Structures
5.4. Case Study 1: Offline Validation Results
5.5. Case Study 2: Real-time Implementation and Validation Results
5.6. Conclusion
6. Intelligent DC Series Arc Fault Detection in PV Systems using LTCNN-ADA with Limited Target-Domain Fault Data
6.1. Introduction
6.2. Experimental Setup
6.2.1. Experimental Setup and Conditions in Source Domain
6.2.2. Experimental Setup and Conditions in Target Domain
6.3. Proposed LTCNN-ADA
6.3.1. Transfer Learning
6.3.2. Wasserstein Generative Adversarial Networks
6.3.3. Procedures of the Proposed LTCNN-ADA
6.4. Case Study 1: Offline Validation Results
6.4.1. Analysis and Evaluation of Generated Arcing Data
6.4.1.1. Training Loss Curves Analysis of ADA by WGAN-GP
6.4.1.2. High Dimensional Feature Visualisation
6.4.1.3. Frequency Domain Analysis
6.4.2. Results with Different Number of Fault Samples
6.4.3. Comparative Study with Related Works
6.5. Case Study 2: Online Validation Results6.5.1. Three-Phase PV System
6.5.2. Single-Phase PV System
6.6. Conclusion
7. Conclusion and Future Work
7.1. Conclusion
7.2. Suggestions for Future Work
Referenceanalysis window is 0.2 seconds) .................... 76
Figure 4.1 General procedures for ML-based DC arc fault detection methods ............. 80
Figure 4.2 Structure of a back propagation neural network (MLP) ............................... 81
Figure 4.3 Illustration of the backpropagation algorithm .............................................. 82
Figure 4.4 Gradient descent with different learning rate ............................................... 83
Figure 4.5 A simple linear SVM for classification ........................................................ 85
XIX
Figure 4.6 A simple illustration of decision tree for series arc fault detection .............. 86
Figure 4.7 Diagram of a SAE for series arc fault detection ........................................... 91
Figure 4.8 General structure for a convolutional neural network .................................. 93
Figure 4.9 Diagram of a LSTM block in a LSTM model .............................................. 95
Figure 4.10 Schematic diagram of the experimental setup ............................................ 96
Figure 4.11 Original CNN architecture: (a) LeNet 5; (b) AlexNet; (c) VGG 16 ......... 101
Figure 4.12 The optimal lightweight CNN structure and feature visualisation ........... 103
Figure 4.13 Feature extraction process and visualisation of normal/arcing feature map
....................................................................................................................................... 110
Figure 4.14 Flowchart of real-time series arc fault detection ...................................... 116
Figure 4.15 Response to small step changes induced by irradiance level changes ...... 118
Figure 4.16 Response to a relatively large step change ............................................... 118
Figure 4.17 A mis-operation is experienced during normal conditions ....................... 119
Figure 4.18 Response to start-up transients and MPPT operation of the inverter ....... 119
Figure 4.19 Response to series arc fault during inverter start-up and MPPT operation
....................................................................................................................................... 120
Figure 4.20 Response to several intermittent series arc faults followed by a sustained
arcing ............................................................................................................................. 121
Figure 4.21 Response to several intermittent series arc faults followed by an arcing
with increasing gap distance ......................................................................................... 121
Figure 4.22 A malfunction experienced during arc conditions .................................... 122
XX
Figure 5.1 Experimental setup and schematic representation for target-domain data
collection and real-time validation tests ........................................................................ 131
Figure 5.2 General framework of a generative adversarial network ............................ 133
Figure 5.3 Overview of DA-DCGAN for DC series arc fault diagnosis in PV systems
....................................................................................................................................... 133
Figure 5.4 Healthy signal capture by CT from source domain and target domain ...... 140
Figure 5.5 Visualisation of high-level features in the lightweight CNN before
classification layer using t-SNE method: (a) DA-DCGAN without MMD; (b) proposed
DA-DCGAN. ................................................................................................................ 141
Figure 5.6 Impact of kernel numbers in the convolution layer of the lightweight CNN
on the performance of target-domain DC series arc fault diagnosis ............................. 142
Figure 5.7 Response to DC disconnect switch closing, inrush current during
initialisation of inverter, start-up, and MPPT operation ............................................... 145
Figure 5.8 Response to fast moving cloud and a series arc fault at high irradiance level
(10A, full loading): (a) 5s per division; (b) 200ms per division (zoomed) ................... 146
Figure 5.9 Response to a fast shading disturbance....................................................... 147
Figure 5.10 Response to a series arc fault at low irradiance level in a cloudy day ...... 147
Figure 5.11 Response to several intermittent series arc faults followed by a sustained
arc fault: (a) 2s per division; (b) 200ms per division (zoomed) .................................... 148
Figure 5.12 Response to a series arc fault generated at middle of the PV string on a
sunny day: (a) 1s per division; (b) 100ms per division (zoomed)................................. 149
Figure 6.1 Experimental setup in: (a) source domain; (b) target domain .................... 153
XXI
Figure 6.2 Illustrations of traditional ML methods and ML methods enhanced by
transfer learning ............................................................................................................ 156
Figure 6.3 Examples of experimental signals (sampled at 200-kHz for analysis purpose)
and their frequency spectra under normal and arcing conditions in (a) single-phase PV
system; (b) three-phase PV system. .............................................................................. 158
Figure 6.4 Overview of the proposed LTCNN-ADA framework ................................ 161
Figure 6.5 Training loss curves of WGAN-GP with different 𝑁𝑡, 𝑎 ........................... 162
Figure 6.6 2D visualisation using t-SNE under different transfer tasks (𝑁𝑡, 𝑎 = 60) . 163
Figure 6.7 (a) examples of three phase normal signal; (b) examples of three phase
arcing signal; (c) examples of generated three phase arcing signal; (d) frequency spectra
of the normalised time-series signals; those signals are randomly selected from the
training dataset D and generated arcing dataset by ADA in the transfer task of A→D
(𝑁𝑡, 𝑎 = 60) .................................................................................................................. 164
Figure 6.8 Diagnosis results of different methods for Dataset C: (a) A→C; (b) B→C
....................................................................................................................................... 167
Figure 6.9 Diagnosis results of different methods for Dataset D: (a) A→D; (b) B→D
....................................................................................................................................... 168
Figure 6.10 Response to inverter switch on, start-up, MPPT, and series arcing in a three
phase PV system (12 panels)......................................................................................... 175
Figure 6.11 Response to several intermittent series arcing events generated at the
middle of the PV array followed by a sustained series arcing in a three phase PV system
(12 panels) ..................................................................................................................... 175
XXII
Figure 6.12 Response to partial shading and series arcing in a three phase PV system
(12 panels) ..................................................................................................................... 176
Figure 6.13 Response to series arcing generated at very low irradiance level in a three
phase PV system (12 panels)......................................................................................... 176
Figure 6.14 Response to inverter switch on, start-up, MPPT, and series arcing in a three
phase PV system (6 panels)........................................................................................... 177
Figure 6.15 Response to several intermittent series arcing events generated at the
middle of the PV array followed bya sustained series arcing in a three phase PV system
(6 panels) ....................................................................................................................... 177
Figure 6.16 Response to partial shading and series arcing in a three phase PV system (6
panels) ........................................................................................................................... 178
Figure 6.17 Response to series arcing generated at the middle of the PV array at low
irradiance level in a three phase PV system (6 panels) ................................................. 178
Figure 6.18 Response to inverter switch on, start-up, MPPT, and series arcing in a
single phase PV system (4 panels) ................................................................................ 179
Figure 6.19 Response to several intermittent series arcing events followed by a
sustained series arcing in a single phase PV system (4 panels) .................................... 180
Figure 6.20 Response to partial shading and series arcing in a single phase PV system
(4 panels) ....................................................................................................................... 180
Figure 6.21 Response to series arcing generated at the middle of the PV array at low
irradiance level in a single phase PV system (4 panels) ............................................... 181
XXIII
List of Tables
Table 2.1 Summary of DC arc fault models for simulation ........................................... 25
Table 2.2 Summary of FFT based detection methods .................................................... 30
Table 2.3 Summary of STFT based detection methods ................................................. 31
Table 2.4 Summary of wavelet transform based detection methods .............................. 35
Table 2.5 Summary of statistical analysis based detection methods .............................. 38
Table 2.6 Summary of model based detection methods ................................................. 39
Table 2.7 Summary of ML detection methods ............................................................... 46
Table 2.8 Summary of other detection methods ............................................................. 51
Table 2.9 Comparison of detection methods for DC arc fault detection ........................ 51
Table 3.1 Experimental conditions for characteristics study of DC series arc fault ...... 59
Table 3.2 Wavelet-packet entropy level for different arc phases (11A/200V) .............. 69
Table 3.3 Wavelet-packet entropy level (Non-arc state) for different load current and
source voltage.................................................................................................................. 72
Table 3.4 Wavelet-packet entropy level (Arc state) for different load current and source
voltage ............................................................................................................................. 72
Table 3.5 Wavelet-packet entropy level for different gap distance (6.5A/158V) .......... 76
Table 4.1 Activation functions used in this thesis .......................................................... 89
Table 4.2 Structure and parameters of the optimal lightweight CNN .......................... 104
Table 4.3 Influence of filter size on CNN performance ............................................... 105
XXIV
Table 4.4 Influence of number of filters on CNN performance ................................... 106
Table 4.5 Influence of number of convolution layers on CNN performance ............... 106
Table 4.6 Influence of fully connected layer settings on CNN performance ............... 107
Table 4.7 Performance comparison of different CNNs ................................................ 108
Table 4.8 Evaluation of different popular ML methods ............................................... 114
Table 5.1 Specification of PV model JINKO (JKM350M-72) at STC ........................ 132
Table 5.2 The architecture of different neural networks in DA-DCGAN .................... 136
Table 5.3 Testing accuracy comparison for target domain series arc fault detection .. 139
Table 6.1 Description of datasets for LTCNN-ADA case study .................................. 154
Table 6.2 Symbols and descriptions ............................................................................. 155
Table 6.3 Performance comparison of different algorithms on different transfer tasks
using the same dataset ................................................................................................... 172
1
1. Introduction
1.1. Background and Research Motivation
With the development of new technology and increasing concerns about
environmental pollution, renewable energies come into the world stage and gradually
substitute traditional fossil fuels. Solar energy is one eminent source. For example, as of
30 September 2019, there are more than 2.2 million photovoltaic (PV) installations with
a combined capacity of more than 13.9 GW in Australia [1]. Solar power development
is increasing worldwide, and residential rooftop solar panels and grid-connected PV
generation will play an important role to support the main loads and micro-grids. The
increasing amount of PV systems coupled with increased operating DC voltage level
has a high potential of creating DC arc faults (utility-scaled PV solar farms typically
produce voltage between 600 and 1500 volts, and typical building PV systems produce
voltage between 120 and 600 volts in the USA) [2], [3]. Because the deterioration of
cables, connectors, conductors, and other system components caused by long-time
weathering and aging effect, without adequate scheduled maintenance, the possibility of
DC arc occurrence in PV systems is sharply going up [4].
AC arc fault has been extensively investigated for decades resulting in many
detection methods proposed. Extensive literature reviews on high impedance and AC
arc faults can be found in [5]–[8]. There are various physical and electrical
characteristics that can be exploited for AC arc fault detection, such as intermittency of
the arc, asymmetry in the current waveform, current buildup, randomness of the current
magnitude, high frequency variation in the current waveform, etc. [6]. AC arcs are more
likely to produce some distortions around the zero-crossing points in the current
waveform, because it periodically crosses zero points and the voltage near zero-crossing
2
points is too small to break down the air gap. Therefore, the AC arcs periodically result
in ignition and re-ignition in each power frequency cycle, which introduces many
harmonics into the signal. This kind of feature is commonly used as one of the key
characteristics for AC arc fault detection. On the other hand, DC arcs tend to be more
sustainable and difficult to be extinguished compared to AC arcs due to the absence of
natural alternating current at zero crossing. Furthermore, the absence of many
distinguishable features (e.g. distortions around zero-crossing points) makes DC arcs
more difficult to be detected. Examples of AC and DC arc fault current waveforms,
including their key characteristics in the time domain, are illustrated in Figure 1.1.
Distortions around
zero-crossing points
Asymmetry
Intermittence
A sudden current drop
Before arcing After arcing
More variations
(a)
(b)
Figure 1.1 Examples of arc current waveforms: (a) AC arc fault; (b) DC arc fault
3
Ironically, there is less scheduled maintenance for PV systems because the PV
components are considered reliable. Therefore, arc faults are often ignored. But they
could last for a long time. The high-temperature plasma generated by a sustained arc
could produce significant amount of thermal energy [9], which can lead to severe
damage to the PV systems such asshown in Figure 1.2 and Figure 1.3 [10], [11].
(a)
(b)
Burned PV panels
Burned PV conductors
Figure 1.2 Fire due to DC arc faults in Bakersfield, USA in 2009 [10]: (a) Overview of
burned PV panels; (b) Burned PV conductors
DC arc faults are becoming more common incidents in PV systems nowadays.
They have caused many catastrophic fires around the world in the past decade. The
events in Bakersfield (USA) and Mount Holly (USA) in 2009 and 2011, respectively,
raised attention and triggered the formation and improvement of the relevant standards
and codes [10], [12]. Many solar fire incidents were reported recently across Australia,
such as Haddon in Victoria, Cairns in Queensland, Felixstow in South Australia, New
4
Lambton in New South Wales, etc [11].
Although arc faults could be mitigated by improving the construction and design of
the PV system and its components, implementing a detecting device that continuously
monitors the arc faults would dramatically increase the system reliability and reduce the
fire risk. The 2011 National Electrical Code requires all rooftop PV systems of DC
operating voltage above 80 volts equip with series arc fault circuit interrupters (AFCIs),
and then the requirement extends to all types of PV systems greater than 80 volts in
2014 to reduce the fire hazard due to arc faults [13]. UL-1699B Outline was initially
introduced to fulfil these requirements in 2013 [14], and the first edition of the formal
UL-1699B Standard became available as a guideline to develop and test arc fault
detectors (AFDs) and AFCIs from August 2018 [15]. All these factors call for the
development of reliable, robust, cost-effective, and intelligent DC arc fault detection
algorithms for PV system protection.
Figure 1.3 Fires due to DC arc faults in Australia in recent years [11]
5
Series arc fault is considered to be more dangerous because it is more difficult to be
detected as compared to parallel arc fault [16]. Unlike the parallel arc fault, the series
arc fault increases the circuit impedance so the load current level will go down, and thus
cannot reach 156% of the maximum normal current defined by National Electrical Code
to melt the overcurrent protection fuse [17], [18]. Furthermore, it has been shown in [19]
that the overwhelming majority of faults that result in fires in PV systems are series arc
fault and grounding fault; other types of parallel arc faults in PV systems are less likely
to happen. The undetected grounding faults can contribute to parallel arc faults.
However, it is better to prevent this type of fault by improving the detection and
protection of grounding faults. Because of this, the relevant standards and codes are
mainly focused on detection of series arc faults in PV systems. Therefore, series arc
faults will be the focus of this thesis.
1.2. Summary of Research Contributions
The contributions by the author from this research are:
• A comprehensive review of DC arc faults in PV systems and their state-of-the-
art diagnosis methods is carried out. Useful information and technical details
of applied methods are summarised. The capabilities and limitations of
different detection methods are presented and discussed.
• The characteristics of DC series arc fault are investigated. A time-frequency
domain method based on wavelet packet decomposition (WPD) and entropy
theory is developed to extract consistent patterns characterising the arc current.
• A novel deep learning (DL) DC series arc fault detection method for PV
systems using convolutional neural network (CNN) is developed. To the best
of the author’s knowledge, DL is applied and its feasibility is investigated for
6
the first time in this specific application. Different machine learning (ML)
methods, including conventional ML and DL, are evaluated and compared
using the same experimental datasets collected in the laboratory to examine
their effectiveness. Technical roadblocks preventing intelligent DC series arc
fault detection from being applied to real-world industry are identified,
including imbalanced dataset, small dataset, data inconsistency, unlabelled
dataset, model complexity and real-time capability, and model interpretability.
Detailed recommendations and potential solutions to these challenges are
provided.
• An optimal lightweight CNN structure is developed and comprehensively
tested. It can achieve the same level of detection accuracy while the model
complexity is substantially reduced as compared to well-known deep CNN
models. Thus, the proposed method is more suitable for cost-effective real-
time deployment in resource-constraint edge devices.
• A novel and effective methodology, namely domain adaptation and deep
convolutional generative adversarial network (DA-DCGAN), is proposed. It
overcomes the performance degradation of DL-based detection algorithms
caused by data inconsistency between the source-domain data (e.g. laboratory)
used during the development and the target-domain data (e.g. real PV systems)
encountered in operation in the field. It also tackles the problem of lack of
faults data in the target domain, even the extreme case of no target-domain
fault data available.
• A novel framework, namely lightweight transfer convolutional neural network
with adversarial data augmentation (LTCNN-ADA), is developed to address
the same problem, i.e. data inconsistency and insufficient data. It optimises the
7
detection performance of DL-based algorithms where only a limited amount of
target-domain fault data is available.
1.3. Thesis Organisation
The organisation of this thesis is as follows:
Chapter 1 - Introduction. This chapter provides the background context, research
motivation, novel contributions of this research, and achievements by the author.
Chapter 2 - Literature Review. This chapter presents an in-depth review of DC arc
faults in PV systems and their diagnosis methods.
Chapter 3 – Characteristics Study on DC Series Arc Fault. This chapter examines
the characteristics of DC series arc fault in detail. The impact of the load current, source
voltage, and arc gap distance on the static characteristics and high frequency
characteristics are investigated. Wavelet packet entropy is applied to analyse the arc
current signals.
Chapter 4 – DC Series Arc Fault Detection in PV Systems using Deep Learning.
This chapter presents a DL-based detection algorithm using CNN. A lightweight CNN
structure is designed to achieve a good balance between model complexity and
detection accuracy. A comparative study among different conventional ML methods
and DL methods is performed using the same experimental dataset. Finally, challenges
in applying ML methods for practical PV series arc fault detection are identified and
potential solutions are provided.
Chapter 5 - Intelligent DC Series Arc Fault Detection in PV Systems using DA-
DCGAN without Target-Domain Fault Data. This chapter presents a novel
methodology based on DL for DC series arc fault detection in PV systems when the
8
fault data from the target PV system is not available. The proposed method is
implemented in an embedded system and validated in real-time under different test
conditions.
Chapter 6 - Intelligent DC Series Arc Fault Detection in PV Systems using
LTCNN-ADA with Limited Target-Domain Fault Data. This chapter presents a
novel framework based on DL for DC series arc fault detection in PV systems when the
fault data from the target PV system is available but limited. Its effectiveness is
validated through comprehensive off-line analysis, comparative studies with other
recent work, and real-time experiments under different test conditions.
Chapter 7 - Conclusion and Future Work. This chapter gives a summary of the major
research outcomes of this research and provides suggestions for furtherwork.
1.4. List of Publications
Journal Papers
1. Shibo Lu, Rui Ma, Tharmakulasingam Sirojan, B.T. Phung, and Daming
Zhang, “Lightweight Transfer Nets and Adversarial Data Augmentation for
Photovoltaic Series Arc Fault Detection with Limited Fault Data”, Submitted
to Solar Energy, 2020.
2. Shibo Lu, Hua Chai, Animesh Sahoo, and B. T. Phung, “Condition
Monitoring based on Partial Discharge Diagnostics using Machine Learning
Methods: A Comprehensive State-of-the-Art Review,” accepted for
publication in IEEE Transactions on Dielectrics and Electrical Insulation, 3
July, 2020.
3. Shibo Lu, Tharmakulasingam Sirojan, B. T. Phung, Daming Zhang, and
Eliathamby Ambikairajah, “DA-DCGAN: An Effective Methodology for DC
9
Series Arc Fault Diagnosis in Photovoltaic systems,” IEEE Access, vol. 7, pp.
45831-45840, April 2019.
4. Tharmakulasingam Sirojan, Shibo Lu, B. T. Phung, Daming Zhang, and
Eliathamby Ambikairajah, “Sustainable Deep Learning at Grid Edge for Real-
time High Impedance Fault Detection,” IEEE Transactions on Sustainable
Computing, doi: 10.1109/TSUSC.2018.2879960
5. Shibo Lu, B. T. Phung and Daming Zhang, “A Comprehensive Review on DC
Arc Faults and Their Diagnosis Methods in Photovoltaic Systems,” Renewable
& Sustainable Energy Reviews, vol. 89, pp.88-98, June 2018.
Conference Papers
1. Shibo Lu, Animesh Sahoo, Rui Ma, and B. T. Phung, “DC Series Arc Fault
Detection using Machine Learning in Photovoltaic Systems: Recent
Developments and Challenges,” International Conference on Condition
Monitoring (CMD), Phuket, Thailand, 25-28 Oct. 2020.
2. Tharmakulasingam Sirojan, Shibo Lu, B. T. Phung, Daming Zhang, and
Eliathamby Ambikairajah, “Embedded Edge Computing for Smart Meter Data
Analytics,” International Conference on Smart Energy Systems and
Technologies (SEST), Porto, Portugal, 9-11 Sep. 2019.
3. Shibo Lu, B. T. Phung, Daming Zhang, and Hua Chai, “An Experimental Study
of Low-Current DC Series Arc Faults for Condition Monitoring Purpose,”
International Conference and Exhibition on Electricity and Distribution
(CIRED), Madrid, Spain, 3-6 June 2019.
4. Hua Chai, Shibo Lu, B. T. Phung, Daming Zhang “Comparative Study of
Partial Discharge Localization based on UHF Detection Methods,” International
Conference and Exhibition on Electricity and Distribution (CIRED), Madrid,
10
Spain, 3-6 June 2019.
5. Miao Li, Shibo Lu, Daming Zhang, and B. T. Phung “Series Arc Fault
Detection in DC Microgrid Using Hybrid Detection Method,” Annual
Conference of the IEEE Industrial Electronics Society (IECON), Washington
D.C., USA, 21-23 Oct. 2018.
6. Tharmakulasingam Sirojan, Shibo Lu, B. T. Phung, Daming Zhang and
Eliathamby Ambikairajah, “High Impedance Fault Detection by Convolutional
Deep Neural Network,” IEEE International Conference on High Voltage
Engineering and Application (ICHVE), Athens, Greece, 10-13 Sept. 2018.
7. Shibo Lu, B. T. Phung, and Daming Zhang, “Study on DC Series Arc Fault in
Photovoltaic systems for Condition Monitoring Purpose,” Australasian
Universities Power Engineering Conference (AUPEC), Melbourne, Australia,
Nov. 2017.
8. Shibo Lu, Daming Zhang and B. T. Phung, “Arcing Fault Detection in the
Scenario with Renewable Energy Generation,” Annual Conference of the IEEE
Industrial Electronics Society (IECON), Beijing, China, Oct/Nov. 2017.
9. Ruihao Song, Shibo Lu, Tharmakulasingam Sirojan, and B.T. Phung, “Power
Quality Monitoring of Single-Wire-Earth-Returned Distribution Feeders,”
International Conference on High Voltage Engineering and Power Systems
(ICHVEPS), Bali, Indonesia, Oct. 2017.
Patent
1. Tharmakulasingam Sirojan, Shibo Lu, B. T. Phung, and Eliathamby
Ambikairajah, “Apparatus and process for real-time detection of high-
impedance faults in power lines,” No. PCT/AU2019/051219, Nov. 2019.
11
2. Literature Review
2.1. Introduction
There are two surveys related to DC arc faults in PV systems [20], [21]. Yao et al.
briefly reviewed a limited number of arc fault detection techniques for DC systems,
including PV systems [20]. Alam et al. conducted a comprehensive survey on detection
and mitigation techniques of catastrophic faults, such as line-line faults, ground faults,
and arc faults in PV systems [21]. However, both studies did not present arc fault
diagnosis techniques for PV systems in detail. Moreover, both surveys did not discuss
the capabilities and limitations of different detection algorithms, such as the required
sampling frequency and computation load, which are very useful when they are to be
implemented in microprocessor controllers.
In this chapter, the primary objective is to present the state-of-the-art detection
methods for diagnosis of DC arc faults in PV systems. The capabilities and limitations
of different methods are discussed, compared, and summarised. Besides, in order to
develop effective detection algorithms, it is of significance to know the arc fault
characteristics and mechanisms. Since carrying out field testing is difficult, costly, time-
consuming, and still not exhaustive (covering all fault types), precisely modelling arc
faults becomes more critical. Therefore, the different types of DC arc fault model and
their capability for PV systems application are presented and compared. Furthermore,
the development trend of detection methods in PV systems is discussed at the end.
2.2. DC Arc in Photovoltaic Systems
2.2.1. Photovoltaic Systems Structure and Arc Hazards
12
The basic structure and the procedure of forming the solar array are shown in
Figure 2.1. The solar cells are connected together to form a solar module, and solar
modules are usually connected in series to form a solar string to increase the DC
operating level of the whole system. These solar strings could be connected in parallel
to increase the DC current level, which accordingly increases the power generation
capability [22], [23].
Figure 2.1 Typical structure of PV systems
There are mainly two types of possible arc faults in PV systems: series and parallel
arc faults, the latter includes grounding arc fault. Parallel and grounding arc faults often
draw a large fault current because of the sizeable different potential, which is easier to
be detected by traditional protection devices [17], [18]. In contrast, due to its inherent
nature, the series arc fault current (lower than normal operating current level) will not
be sufficient to melt the fuse or activate the overcurrent protection devices. Parallel arc
faults except grounding arc fault are less likely to happen in PV systems compared to
13
series arc fault [19]. Therefore, in PV systems, the relevant standards and codes are
mainly focused on series arc fault detection and protection.
Series arcs can be created across small gaps between two connecting terminals such
as busbar-ribbon connection in PV modules and connection in a combiner box as shown
in Figure 2.2. The lack of scheduled maintenance, aging effect, weather effect (e.g.
corrosion caused by rain), mechanical damage induced by wind, animal bites, improper
wiring can cause bad joints to occur. Bad joints decrease the cross-section area,
effectively increase the connection resistance, and significantly increase the heat loss. It
introduces more thermal stress due to the higher operating temperature, and accelerates
the deterioration in connections, which leads to loose connections [24]. After that, a
small gap may develop between two connecting terminals without interrupting the
current flow. When the electric field across the gap exceeds approximately 3 V/μm (the
breakdown strength depends on surrounding environment), the air in the gap starts to
ionise and arc plasma is developed, which will finally form a series arc. The oxygen
flow into the plasma stream further sustains the arc discharge. The gap distance for
seriesarcs is typically less than few millimetres. The fault current magnitude is low,
typically few amps, at the solar cell, the module, or the string level. However, it can
reach several hundred amps at the combiner box and more than a thousand amps at the
DC side of the inverter in large PV systems.
Parallel arcs have a similar mechanism as series arc faults. They can be developed
between two conductors in the same string, two conductors of two different strings, and
conductor and grounding point as shown in Figure 2.2 [25]. The parallel arc faults are
mainly caused by degradation and breakdown of insulation due to various reasons such
as animal bites, mechanical damage, and aging effect. This is because most cables and
wires in PV systems are exposed to the open environment (no protective enclosure) [26].
14
Both the fault current level and the gap width are larger than those of series arc fault.
Figure 2.2 Example of possible locations where arcing may occur in PV systems
Due to the nature of PV modules and design of the PV systems, there could be
thousands of connectors and a significant amount of cables in the system as shown in
Figure 2.1 and Figure 2.2. Every connecting point can create an arc fault, which
substantially increases the possibility of arc fault occurrence, especially the series arc
fault. Because of the inverter operation for maximum power point tracking (MPPT), the
fault current level can return to normal while the arcing fault still exists, which makes it
more challenging in DC arc fault detection. It has been shown that only 0.4 mm2 of arc
area could cause ignition of surrounding materials and burn off the metal coating within
2 seconds [27]. In [28], it is found that, at a radius of 10 mm, the surface ignition time
for plastics are 4 s, 1 s, and 0.3 s with arc power of 200 W, 400 W, and 800 W,
respectively. As a result, the heat energy generated by undetected arc over a long time
can lead to serious damage to system components, and it presents severe threats to
15
system stability and human safety [29]. UL-1699B Outline requires every arc fault
circuit interrupter (AFCI) or arc fault detector (AFD) pass the arc test with the power of
300-900W [14]. However, low power arcs are often the case at string level, which can
also lead to a fire. Therefore, 100W arc fault test has been recommended to be added
into UL-1699B [30].
The formal UL-1699B Standard became available from August 2018 [15]. The
total response time 𝑇𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒 of AFCI and AFD, with a limit of 2.5 seconds, is defined
in (2.1):
𝑇𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒 = max (
750 (𝑗𝑜𝑢𝑙𝑒)
𝐼𝑎𝑟𝑐𝑉𝑎𝑟𝑐
, 2.5) (2.1)
Besides the fire hazard, arc faults will have severe impact on the operating points of
the PV system [31]. The presence of series arc faults will inject extra impedance in the
PV system, which can cause mismatch loss, heating loss, and decreasing fill factor.
Thus, it effectively decreases the efficiency of PV system. Although the operation of the
solar inverter can keep tracking the maximum operating point, the overall output of the
system is essentially going down. Therefore, it is important to protect PV systems from
such a situation.
2.2.2. Challenges to Detect DC Arc Faults
Majority of AC arc detection algorithms mainly rely on the most iconic feature: the
“shoulder” in the arc current waveform [5]. However, the DC arcs do not have zero
crossing points, and the absence of “shoulder” makes DC arc more sustainable and
difficult to be detected [32]. In order to develop an effective detection algorithm and
differentiate DC arc fault from normal condition, besides studying the DC arc
characteristics to extract the arc features, other disturbances from normal operations
16
must be identified. The different disturbances can cause AFCI and AFD fail to response
or make a wrong decision to interrupt the electricity delivery. As shown in Figure 2.3,
there are both internal and external disturbances that can mask or modify the arc
features:
• The intensity of arc noise and variation would be affected by several factors, such
as different electrode materials, geometries, current and voltage levels, load type,
and source type. For example, when the loop current of non-arc state is the same,
the arc discharge tends to be more stable at higher voltage level, which results in a
smaller fault current standard deviation (less variation of current) [33]. Another
example is that the frequency components of fault current with rounded tip
electrodes are clearly lower than that with flat tip electrodes, because the arc is
symmetrically created through the radial centre of electrodes [34].
• The long PV cables would act as an antenna to pick up radio-frequency noise
especially in frequency band of 100 kHz to 500 MHz. Also, different system
topologies and sizes would affect the radio-frequency response, and thus it may
modify the arc noise profile [35]. PV cables also have inductive component, which
acts as a low pass filter, decreasing the noise level increased by arc fault [36].
• The electronics loads, such as DC/DC converter and DC/AC inverter would
produce high frequency electromagnetic interference noise in the circuit, which
could cause nuisance tripping [37].
• Crosstalk effect may introduce radio-frequency noise, causing tripping [38].
• The transformer-less inverters are used in some PV system applications, and AC
noise will be injected into the circuit from the AC side (50 Hz grid fundamental
frequency component and its harmonics).
• Step change caused by load shifting (turn off the converter or inverter), fast moving
17
cloud, partial shading operation, mechanical vibration induced by wind, system
shutdown, and power adjustment by inverter exhibit similar behaviours to arc faults
[31], [39], [40].
Figure 2.3 Typical disturbance sources
2.3. DC Arc Models
The arc is a very complicated and chaotic physical phenomenon. The behaviours of
arc vary from different voltage and current levels, environmental conditions (i.e.
ambient temperature, pressure, and moisture level), and arc length. These factors make
its physical constants difficult to be defined.
Nowadays, the majority of arc studies are based on observation of experiments and
analysis of acquired data sets, and scholars mainly use the V-I curve to characterise this
complex phenomenon. The quasi-static arcing V-I characteristic with different arc
length is shown in Figure 2.4. It can be seen that in the lower current region, the smaller
the current, the larger the voltage, where the power of the arc (𝑃𝑎𝑟𝑐 = 𝑉𝑎𝑟𝑐𝐼𝑎𝑟𝑐) tends to
remain the same; while in the higher current region, due to the magnetic effect, the
18
voltage approximately remains unchanged with the increasing current [41], [42]. The
transition point line demarcates those two regions. Note that the arc length is not
equivalent to the gap width. In real situations, the arc length could be significantly
larger than the gap width unless the arc current is low, and the gap width is short. More
comprehensive study on DC arc fault can be found in [43] and [44].
Figure 2.4 V-I characteristic of arc
Acquiring arc fault data from real PV systems is too costly and difficult as the
features of arc fault vary for different fault locations and current levels. Therefore, an
alternative approach is using computer simulation of a realistic arc fault model to
develop and verify the arc fault detection algorithms, and it can indeed reduce the cost.
There are three main types of arc model that can be used for simulation: models
based on physical principles, traditional V-I empirical models obtained from
measurement data, and heuristic models. A summary of arc models that can be used for
arc fault simulation in PV systems is presented in Table 2.1.
19
2.3.1. Physics-basedMais conteúdos dessa disciplina
Screenshot_20251113_170941_UNIASSELVI Leo App- Screenshot_20251113_171042_UNIASSELVI Leo App
- Screenshot_20251113_170949_UNIASSELVI Leo App
- Screenshot_20251113_171314_UNIASSELVI Leo App
- Screenshot_20251113_171229_UNIASSELVI Leo App
- Screenshot_20251113_171000_UNIASSELVI Leo App
- Screenshot_20251113_171055_UNIASSELVI Leo App
- Screenshot_20251113_171336_UNIASSELVI Leo App
- Screenshot_20251113_171345_UNIASSELVI Leo App
- Screenshot_20251113_170926_UNIASSELVI Leo App
- Screenshot_20251113_170913_UNIASSELVI Leo App
- Screenshot_20251113_171114_UNIASSELVI Leo App
- Screenshot_20251113_171106_UNIASSELVI Leo App
- A função das(os) _______ no gerador DC é coletar corrente do comutador e entregá-la ao circuito externo. Assinale a alternativa que preenche cor...
- Um transformador ideal tem 200 espiras no primário e 50 no secundário. Qual é a corrente no secundário se a corrente no primário é 2 A? A 0,5 A. B 0,2
- Um transformador ideal tem 200 espiras no primário e 50 no secundário. Qual é a corrente no secundário se a corrente no primário é 2 A? A 0,5 A. B ...
- Em dispositivos eletromecânicos, como já visto, existe a indução de tensões devido ao campo magnético estabelecido pela excitação. Assinale a alternat
- No gerador DC, qual dos seguintes dispositivos desempenha um papel importante no fornecimento de corrente contínua: A Rotor. B Escovas. C Bobinas. D C
- Um transformador ideal tem 200 espiras no primário e 50 no secundário. Qual é a corrente no secundário se a corrente no primário é 2 A? A 0,5 A. B 0,2
- A respeito dos transformadores trifásicos, assinale a alternativa incorreta. A Um transformador trifásico pode ser construído a partir de um banco for
- O sentido de rotação do motor DC em série pode ser revertido trocando os terminais: A Armadura e campo. B Campo. C Fonte e armadura. D Fonte. ...
- No gerador DC, qual dos seguintes dispositivos desempenha um papel importante no fornecimento de corrente contínua: A Bobinas. B Escovas. C Anéi...
- Motores bifásicos são construtivamente semelhantes aos motores monofásicos de fase dividida. Os motores monofásicos são compostos por dois enrolamento
- O uso do motor com enrolamento bifásico de forma simétrica permite algumas vantagens, ou seja: campo magnético girante uniforme e constante, força mag
- As máquinas de relutância variável (MRV) podem ser configuradas de várias maneiras. Comparadas às máquinas de relutância variável com rotor saliente,
- Os motores de indução são amplamente utilizados na indústria e em diversas aplicações oque é um motor de indução
- Bioenergia AULA 3 - Combustão da Madeira
- callen cap 15 15-3
