8546 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO.

11, NOVEMBER 2017

Line-to-Line Fault Detection for Photovoltaic

Arrays Based on Multiresolution Signal
Decomposition and Two-Stage Support
Vector Machine
Zhehan Yi, Student Member, IEEE, and Amir H. Etemadi, Member, IEEE

Abstract—Fault detection in photovoltaic (PV) arrays output of a PV array under different operating conditions, may
becomes difficult as the number of PV panels increases. further increase the difficulty of detecting such faults [2].
Particularly, under low irradiance conditions with an active A number of methods for PV fault detection have been pre-
maximum power point tracking algorithm, line-to-line
(L-L) faults may remain undetected because of low fault sented in the literature. Detecting the change of PV internal
currents, resulting in loss of energy and potential fire resistance using the signals in the extremum-seeking control
hazards. This paper proposes a fault detection algorithm (ESC)-based MPPT method is proposed in [3]. Internal resis-
based on multiresolution signal decomposition for feature tance can be used for monitoring the performance of a PV array,
extraction, and two-stage support vector machine (SVM) but needs further development to be used for fault detection. Ad-
classifiers for decision making. This detection method only
requires data of the total voltage and current from a PV ditionally, it may not be applicable in PV systems with different
array and a limited amount of labeled data for training the MPPT algorithms. Based on the quantum probability model the-
SVM. Both simulation and experimental case studies verify ory, [4] presents a series arc-fault detection algorithm for PV
the accuracy of the proposed method. systems. Regarding the islanding protection for grid-connected
Index Terms—Fault detection, multiresolution signal PV system, [5] introduces a method using the active frequency
decomposition (MSD), photovoltaic (PV) power system, pro- drift method with the help of a positive feedback, while [6]
tection, renewable energy, support vector machine (SVM), uses the extension neural networks and chaos synchronization
wavelet transformation. (CS) to detect faults. An active method to detect hot spots (due
to partial shading conditions) within a series of PV cells using
I. INTRODUCTION ac parameter characterization has been proposed in [7]. These
methods seek to detect PV problems other than L-L faults.
ENEWABLE ENERGY SOURCES (RES) have been sig-
R nificantly contributing to sustainable power generation,
considering the increasing demand and global environmental
There are a few algorithms for detecting PV faults by compar-
ing the estimated and measured PV parameters. A maximum-
power-based PV performance validation method is proposed in
degradation. Among various RES, solar energy is one of the [8], which monitors and analyzes three types of silicon-based
most attractive power sources because of its technical, econom- PV arrays in the Italian alpine region, by comparing the mea-
ical, and environmental benefits. Increasing deployment of PV sured and rated maximum power of the PV systems. Another
systems is revealing various technical issues that used to be un- maximum-power-comparison-based algorithm to detect ground
known to the PV industry. PV array L-L short-circuit fault is one faults, mismatch faults, and bridged faults in PV systems has
such issues that can cause PV failure and lower system efficiency been presented in [9], where a fractional-order color relation
[1]. L-L faults are hardly detectable under low irradiance condi- classifier is employed to quantify the output power degradation
tions or when the fault occurs between points with close electric and determine faults. Platon et al. [10] proposed an online fault
potential, i.e., low mismatch faults. Moreover, maximum power detection method by comparing the model predicted and mea-
point tracking (MPPT) schemes, while optimizing the power sured values of the ac power output for a Canadian PV system.
Based on descriptive and inferential statistics, [11] compares
Manuscript received September 5, 2017; revised January 11, 2017 the offline estimation and online measurement of a PV system
and March 20, 2017; accepted April 25, 2017. Date of publication May to monitor its operation and detect faults. Guerriero et al. [12]
11, 2017; date of current version October 9, 2017. (Corresponding introduced a method, which installs a wireless self-powered sen-
author: Zhehan Yi.)
The authors are with the Department of Electrical and Computer En- sor on each PV panel, and compares the measured and estimated
gineering, The George Washington University, Washington, DC 20052 maximum string power based on the iMPPT algorithm proposed
USA (e-mail: zhehanyi@gwu.edu; etemadi@gwu.edu). in [13]. A panel-based multipurpose sensor, which is equipped
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. with voltage, current, irradiance, temperature, and inertial mon-
Digital Object Identifier 10.1109/TIE.2017.2703681 itoring, is used to detect faults in PV arrays by analyzing its

0278-0046 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on June 17,2021 at 04:07:33 UTC from IEEE Xplore. Restrictions apply.
 YI AND ETEMADI: L-L FAULT DETECTION FOR PV ARRAYS BASED ON MSD AND TWO-STAGE SVM 8547

estimated and real efficiency in [14]. Although detecting faults

by comparing estimated and measured parameters may be effec-
tive, these fault detection methods may suffer from the following
limitations:
1) inaccurate estimation due to discrepancies between the
actual system and its simulated model;
2) maloperations due to the similarity between low
mismatch fault conditions and operation under low
irradiance;
3) the need for numerous sensors leading to costly schemes.
Hu et al. [15] introduced a method to reduce the number of
sensors for PV fault detection by optimizing the locations of
the voltage sensors and omitting the current sensors, based on
which a two-section fault detection scheme is proposed. This
algorithm still requires a number of voltage sensors within PV
strings, especially for large-scale arrays. A method that uses
the voltage, current, and irradiance of a PV array to detect PV Fig. 1. Typical configuration of a grid-connected PV system.
mismatch and partial shading conditions is presented in [16]
based on power loss and the ratio of instantaneous dc power and
irradiance. However, irradiance level can vary significantly over
a large-area PV array, which will affect the performance of the
method. Additionally, the power loss is calculated based on the
difference between the measured and estimated power that may
not be accurate. Chen et al. [17] introduced a quick detection
algorithm for PV faults based on the sequential change detec-
tion framework using a minimal amount of meters. This method
mainly aims at eliminating the delay of fault detection and it
needs further techniques to differentiate L-L faults from irradi-
ance fluctuations. A probabilistic neural network is proposed in
[18] to classify PV open- and short-circuit faults and locate the Fig. 2. Characteristic curves of a PV array at different irradiance levels.
faulted string. The validation of this method is based on severe
faults that result in high fault currents.
Overall, little attention has been paid to small-mismatch additional sensors. It provides rapid and accurate detection of L-
faults, high-impedance faults, or those under low irradiance, L faults, which can be combined with fault locating techniques
which are challenges for PV fault detection due to the minimal to clear faults at an early stage. Both simulation and experimen-
disturbances they may introduce. Also, the influence of MPPT tal results validate the accuracy and reliability of the method.
has not been taken into consideration in the aforementioned de- The rest of this paper is organized as follows: Section II intro-
tection methods. The authors in [19]–[21] presented multiple duces the background of the L-L fault in PV arrays and analyzes
algorithms as an attempt to detect PV faults under these sit- its detection blind spots resulting from low irradiance, low mis-
uations; however, the accuracy for faults with low-percentage match percentage, and MPPT. The proposed method and the
mismatch or high impedance can be improved. There are also detection procedures are explained in Section III, followed by
methods aiming at locating PV faults [12], [14], [22]. Neverthe- case studies in Section IV where simulations and experiments
less, these methods may require numerous costly sensors at the are carried out to verify the method. Section V concludes this
panel level and do not provide rapid and accurate detection for paper.
L-L faults, especially for high-impedance faults, low-mismatch
faults, and those occur under low irradiance conditions, which II. PV SYSTEM CONFIGURATION AND FAULT DETECTION
are challenging for detection. CHALLENGES
To address the issues discussed previously, this paper pro-
poses an algorithm to detect L-L faults in PV arrays based on A. Typical Configuration of a PV System
multiresolution signal decomposition (MSD) and a two-stage A typical PV system consists of a PV array, a dc/dc converter
support vector machine (SVM) classifier. The proposed method and an inverter (see Fig. 1) [23]. PV panels are connected in a
can detect L-L faults under a wide range of situations, includ- series-and-parallel configuration to achieve the required levels
ing faults occur in low irradiance conditions, high-impedance of output voltage and power [2], [24]. Characteristic curves and
faults, and low-mismatch faults. This is an economical and fast the operating point of a PV array change with the irradiance or
detection system since it is based on the overall voltage and cur- temperature. As is illustrated in Fig. 2, the power generated by
rent of the PV array instead of using numerous costly sensors. It a PV array increases as its terminal voltage increases from 0 to
can be implemented into a low-cost controller and do not require VM PP i , where the PV array reaches its maximum power point

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on June 17,2021 at 04:07:33 UTC from IEEE Xplore. Restrictions apply.
 8548 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 11, NOVEMBER 2017

(MPPi ), and decreases thereafter, where i refers to different op-

erating conditions. Therefore, MPPT algorithms are employed
in PV systems to extract the maximum power from the PV ar-
ray, i.e., forcing the PV array to work around VM PP i , in any
operating conditions [25].
Grounding of PV systems depends on the scale, installation
location, local standards, and operating voltage level. While
some PV plants are system grounded through ground fault detec-
tor interrupters, residual current monitoring devices, or dc insu-
lation resistance (Riso ) measurements, others are only equipment
grounded, which connect all noncurrent carrying conductors to-
gether to the earth through equipment grounding conductors Fig. 3. Back-feeding current of a faulted PV string (L-L fault at t = 2 s).
[1], [26].

diodes are not OCPD according to the UL Standard 1703 [30];

B. L-L Fault and Its Detection Blind Spots and they may fail and cause unnecessary losses [31]. Therefore,
An L-L fault may occur within the same string or across as an attempt to address these issues, this research focuses on
different strings of a PV array, as is depicted in Fig. 1 (red detecting L-L faults under a wide range of weather conditions,
dotted arrows). Commonly, L-L faults result from short especially under low light conditions, where the fuses may fail
circuits between the current carrying conductors, mainly due to interrupt the back-feeding currents.
to water ingress, animal chewing, mechanical damage, or dc Furthermore, MPPT schemes that enable PV systems to ex-
junction box corrosion [27]. Although L-L faults may not occur tract the maximum PV power under different irradiance levels,
frequently in a PV systems, their occurrence is unpredictable may mask these faults. An L-L fault will lead the PV array to
and detection schemes should be installed to detect them at the operate on another set of characteristic curves for the fault con-
early stage to prevent further losses or fire hazards. The severity ditions, for instance, red dash-dotted curves in Fig. 2. The MPPT
of an L-L fault is represented by the mismatch percentage, notices this change without recognizing its root cause, and will
which indicates the number of panels involved. For instance, F1 optimize the system by moving the operating voltage to VM PP f ,
and F2 in Fig. 1 are denoted as L-L faults with 10% mismatch, which consequentially increases the output current of the array
assuming each PV string contains ten panels in series. Likewise, and reduces the back-feeding current [1], [2], [32]. Fig. 3 il-
the mismatch percentage of F3 and F4 are 20% and 90%, lustrates this process: without MPPT, the back-feeding current
respectively. Essentially, the lower the mismatch percentage, caused by an L-L fault (black dotted curve, 40% mismatch) ex-
the more difficult are the L-L faults to be detected. Fuses are ceeds the fuse threshold and will be interrupted by the fuse; with
required in PV strings by the NEC Artical 690 and UL Standard MPPT, the back-feeding current of the same fault (blue solid
2579 to prevent overcurrent that might result from L-L faults curve) exceeds the threshold but is immediately pulled back by
[28], [29]. the MPPT to a permissible level, rendering it undetectable by
When an L-L fault occurs, for instance, F4 in Fig. 1, the the fuse because of the insufficient melting time. Moreover, if
voltage of the faulted string falls suddenly, which draws a back- the L-L fault involves less panels, e.g., 20% mismatch (green
feeding current from other strings. Normally, overcurrent pro- dash-dotted curve in Fig. 3), the fault current will be too small
tection devices (OCPD), i.e., fuses, will blow to interrupt this to be detectable by conventional protection methods.
back-feeding current. The threshold value of a fuse should be
greater than 2.1 ISC , according to UL Standard 2579-7, where III. PROPOSED DETECTION ALGORITHM FOR PV L-L FAULTS
ISC is the short-circuit current of the PV string under standard
The proposed fault detection method is based on the pattern-
testing condition (STC, irradiance = 1000 W/m2 , temperature
recognition theory, which takes advantage of the MSD process
= 25 ◦ C) [2], [28], [29]. However, as irradiance decreases, e.g.,
to decompose the collected signals by wavelet transformation,
in cloudy days (irradiance may be less than 500 W/m2 ), the cur-
and to create the feature space that can characterize L-L faults in
rent generated by the PV panels become generally smaller. In
PV arrays. A two-stage SVM classifier is trained using a small
this situation, the back-feeding current induced by an L-L fault
amount of the collected feature samples and labels, after which
may be insufficient to melt the fuse, making the faults being
it is employed as the decision-making stage to detect L-L faults.
undetected. In this case, the faulted strings will become a load
The detection procedure is elaborated in this section.
to consume power from other healthy strings, which hinders
the optimal efficiency of the PV array and potentially results
in damage to the panels and fire hazards. Blocking diodes may A. MSD and the Feature Extraction Process
be installed in each PV string to prevent damages from back- MSD is a digital signal processing (DSP) method that al-
feeding currents. However, since the back-feeding current is lows simultaneous time and frequency analysis of a signal [33],
blocked by the diode, fuse will fail to be melted even a severe and has been successfully applied to fault diagnosis in power
fault has occurred [2]. If there is no further detection schemes, systems [34], [35]. It is applicable in analyzing nonperiodic sig-
faults will be masked in this situation. Additionally, blocking nals, such as power quality disturbances, which consist of both

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on June 17,2021 at 04:07:33 UTC from IEEE Xplore. Restrictions apply.
 YI AND ETEMADI: L-L FAULT DETECTION FOR PV ARRAYS BASED ON MSD AND TWO-STAGE SVM 8549

Fig. 4. MSD tree.

A2
Fig. 5. Fill factor (F F = A1 ) of a PV array.

stationary and transient components [36]. The process is based

on the multilevel digital wavelet transform (DWT) of an input
signal X(n), which then uses a series of high- and low-pass of l samples, is mathematically expressed as
filters (HPFi and LPFi in Fig. 4) to extract multiple levels of 
k
1
the signal details DX (n )i and approximations AX (n )i , respec- f1 = |DX (i)6 |2 (5)
tively. Each level of DWT for a digitized signal X(n) can be l
i=k −l+1
mathematically expressed as
  where k denotes the last sample.
1  k−b 2) Feature #2 (f2 ): The second feature f2 for fault detec-
DWT(m, k) = √ X(n) · ψ √ (1)
a n a tion uses the rate of change of the fill factor (F F ) of the PV
array, where F F is defined as
where ψ(t) is the mother wavelet, k refers to a sample of the
input signal, and a and b are the scale and translation factors (a = VM PP IM PP
FF = (6)
0 , b = nb0 a0 ). To most accurately diagnose L-L faults, the
am m VOC ISC
following three features are extracted to form the fault detection where V OC and I SC are the open-circuit voltage and short-
feature space. circuit current of the PV array under STC, respectively.
1) Feature #1 (f1 ): In order to make the proposed method Graphically, F F equals to the ratio of the area (A2 ) of the
scalable, the voltage and current of the PV array (vPV and iPV largest rectangle that fits in the I − V curve to the area (A1 )
in Fig. 1) are normalized by the open-circuit voltage VOC and of the rectangle delimited by VOC and ISC of a PV array
short-circuit current ISC of the array under STC (provided by (see Fig. 5) [39]. For other irradiances or temperatures, F F
the datasheets from manufacturers), respectively, as follows: represents the normalized maximum power that an array can
vPV provide. Obviously, a higher F F indicates a higher efficiency.
N
vPV (t) = (2)
VOC Both faults and irradiance disturbances will lead to fluctuations
iPV in F F . However, unlike irradiance, faults normally causes
PV (t) =
iN . (3) an abrupt change of F F , since faults result in a sudden
ISC
change of current and voltage. Therefore, the rate of change
These signals are then discretized, denoted as VPV N
(n) and of F F , is sampled using the same sampling rate as f1 , and
IPV (n), and the following signal can be constructed:
N
used as another feature (f2 ) to distinguish L-L faults in PV
N
IPV (n + 1) − IPV
N
(n) arrays.
X(n) = (4) 3) Feature #3 (f3 ): To improve the classification accu-
VPV (n + 1) − VPV (n)
N N
racy, f2 is decomposed by MSD using DB as the mother wavelet.
where X(n) indicates the rate of change of the conduc- DB is one of the most widely employed mother wavelets in
N N
tance of a PV array, based on two consecutive VPV and IPV power system transient analysis, whose orthogonal property
samples. makes it helpful for classification and localization problems
By numerous studies of multiple orders among the wavelets of power system disturbances [40]. The wavelet energy of third-
Harr, Symlets (Sym), Coiflets (Coif), and Daubechies (DB), order MSD details of f2 , |Df 2 (n )3 |2 , is used to construct the
Coif, which was derived from DB, is selected as the MSD third feature using a sliding averaging window (length = l),
mother wavelet for X(n). Coif has more symmetry, regular- which is mathematically described as follows:
ity, and vanishing moments for the scaling function than other
wavelets and it presents a good performance in approximating 1 
k

a polynomial function at different resolution levels [37], [38]. f3 = |Df 2 (n )3 |2 . (7)

l
i=k −l+1
X(n) is decomposed into six levels by the process in Fig. 4 and
the details DX (n )1 , DX (n )2 , ..., DX (n )6 are obtained. DX (n )6 The selection of the decomposition level of a signal is based
is selected for feature extraction, since it shows higher energy on numerous tests and case studies. Df 2 (n )3 responds with a
levels during an L-L fault than normal operations, making it dis- sudden increase in energy at the inception of an L-L fault, which
tinctive for fault detection. Hence, the first feature f1 for fault is a distinct feature for differentiating faults with temporary
detection, based on a sliding averaging window with the length disturbances, e.g., change in irradiance or temperature.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on June 17,2021 at 04:07:33 UTC from IEEE Xplore. Restrictions apply.
 8550 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 11, NOVEMBER 2017

Fig. 6. Illustration of SVM operation.

B. SVM Fig. 7. Training and testing process of the proposed fault detection
method.
The decision making stage employs an SVM, which has been
successfully used for fault detection in power systems [41], [42].
It has emerged as a machine learning technique for efficiently data clouds of some L-L faults, e.g., small-mismatch L-L faults,
learning from minimum amount of raw data, and particularly, and environmental disturbances is fuzzy in the feature space.
for fast solving binary classification problems [43], making it a Therefore, to increase the accuracy, a two-stage SVM classifier
powerful decision making stage for detecting L-L faults in PV is proposed in this method to detect L-L faults in PV arrays.
arrays from normal operation cases. It differentiates data clouds Fig. 7 illustrates the training and testing process of the proposed
in the feature space by a hyperplane that separates data with two-stage SVM. The training set contains the 3-D feature sam-
the maximized margins for both sides. The nearest data sam- ples F  train1 (f1i , f2i , f3i ) from collected L-L faults and normal
i
ples to the hyperplane in the feature space are named “support operating situation, which are then divided into two groups:
vectors.” Fig. 6 illustrates the basic mechanism of a linear SVM 50% to train SVM1, the other 50% to self-evaluate SVM1. The
that separates two data clouds in a 2-D feature space, where the outcome of SVM1 will be a binary quantity string, which con-
hyperplane in 2-D space is a line. For data clouds that are not sists of the self-evaluation result of each sample, where L-L
linearly separable, a kernel function is used to map the original fault is labeled as “1” and normal situation is “0.” Two sliding
training instances xi (i = 1, ..., γ) to a higher dimensional space windows, mean μ and variance var (length l, same as the length
where data clouds are more likely to be linearly separable [44], of the moving average windows of f1 and f3 ), are applied to the
which is also known as the nonlinear SVM, whose learning pro- outcome string, the result of which, F  train2 (μi , vari ), are used
i
cess can be mathematically formulated as an optimization task to train the second stage, SVM2. The two-stage SVM classi-
 γ fier is fully trained at this point. The real-time testing stream,
1 T F test (f1i , f2i , f3i ), that is collected from PV arrays can be fed
min ω ω+C ζi (8) i
2 into the classifier to detect if there are L-L faults. Following the
i=1
testing process in Fig. 7, the testing set evaluates the trained
subject to yi (ω T χω + b) ≥ 1 − ζi (9) SVM1, where the results are applied by the sliding mean and
where ω is inversely proportional to the SVM separation mar- variance windows, and are fed to the trained SVM2 to see the
gin, χ is the transformation from the original feature space to final detection results. The final outcome will be L-L fault or
higher dimensional space, i denotes the current sample, γ is the normal operation. This proposed algorithm focuses on distin-
number of training samples, C is the penalty of misclassifying guishing L-L faults from normal operations, such as normal
the training samples, ζi is the slack variable, and y ∈ {1, −1}γ change of irradiance or temperature.
[45]. By numerous examinations and simulations, Gaussian It is noteworthy that, the proposed two-stage SVM is a binary
radial basis function K(xi , xj ) = χ(xi )T χ(xj ) is selected as classifier that requires training using only a minimal amount of
the kernel function [46], where history data from the PV system being tested, and the proposed
algorithm mainly focuses on detecting L-L faults in PV ar-
−||xi − xj ||22 rays. However, since vPV and iPV have been normalized using
K(xi , xj ) = exp . (10)
σ2 VOC and ISC under STC, respectively, before feature extrac-
tions, the proposed method is applicable in different scales of
C. Proposed Two-Stage SVM Method and Detection PV systems without significant tuning. Additionally, to detect
Procedures
other types of PV faults, more stages of the SVM can be em-
As is discussed in Section II, there are challenges that may ployed in parallel or in further steps to provide more detailed
prevent L-L faults from being detectable. In fact, the boundary of classifications.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on June 17,2021 at 04:07:33 UTC from IEEE Xplore. Restrictions apply.
 YI AND ETEMADI: L-L FAULT DETECTION FOR PV ARRAYS BASED ON MSD AND TWO-STAGE SVM 8551

TABLE I
L-L FAULT SIMULATION SCENARIOS

Variable Simulated Values

Temperature (◦ C) 10, 25, 40

Irradiance (W/m2 ) 300, 500, 800, 1000, 1200
Mismatch Percentage (%) 10, 20, 30, 40, 50, 60
Fault Impedance (Ω) 0, 5, 15, 25
Fault Instant Positive and negative peaks and
both zero crossings of the grid-
phase-A voltage in 1 cycle.
Fault Location Across the same string, and
across different strings.

Fig. 9. Features f1 , f2 , and f3 of an L-L fault with 30% mismatch under

STC, fault at t = 5 s.
Fig. 8. |D X (n )6 |2 and |D f 2 (n )3 |2 of an L-L fault with 30% mismatch
under STC, fault at t = 5 s.
TABLE II
L-L FAULT DETECTION RESULTS FOR THE SVM TRAINED
UNDER STC (SET A)
IV. CASE STUDIES
To evaluate the performance of the proposed method, both Mismatch Fault Impedance (Ω) Average
simulation and experimental case studies are carried out and
Percentage(%) 0 5 15 25 Accuracy
results are presented in this section.
60 100% 100% 98.33% 88.33% 96.67%
50 100% 100% 95.00% 73.33% 92.08%
A. Simulation Verification
40 100% 98.33% 81.67% 66.67% 86.67%
The simulation platform is set up based on the 30 100% 91.67% 71.67% 56.67% 80.00%
20 100% 73.33% 58.33% 46.67% 69.58%
PSCAD/EMTDC and MATLAB is used for data analysis. A 10 66.77% 41.67% 33.33% 20.00% 40.42%
grid-connected PV system (not system grounded) is built using
the configuration presented in Fig. 1, which consists of 100 PV
panels (ten PV panels in a string, ten strings in parallel). The
panel features the following parameters under STC (1000 W/m2 , To evaluate the proposed method comprehensively, and to
25◦ C): VOC = 19.20 V, ISC = 0.59 A. compare the performance of the algorithm under different train-
1) L-L Fault Cases Studies: Section II elaborates on ing sets, the two-stage SVM classifier is trained multiple times
the challenging cases in detecting L-L faults of PV arrays. The using the following training sets, respectively.
presented case studies focus on verifying the performance of the 1) Set a: 24 L-L faults under STC (1000 W/m2 , 25 ◦ C);
proposed method under these challenging cases. To this end, the 2) Set b: 24 L-L faults under 800 W/m2 , 25 ◦ C;
PV voltage and current (vPV and iPV in Fig. 1) with numerous L- 3) Set c: 24 L-L faults under 500 W/m2 , 25 ◦ C;
L faults under different conditions are collected, based on which 4) Set d: 24 L-L faults under 300 W/m2 , 25 ◦ C.
the features (f1 , f2 , and f3 in Section III) for all the studied cases Note that all these faults occur at the positive zero crossings
are extracted to evaluate the detection system. A rather small of the grid phase-A voltage. Another set of 100 nonfault samples
portion of the collected data will be used for training the SVM. under STC are collected to train the SVM with each fault set.
The simulated fault scenarios (2880 in total) are combinations of The trained SVM from each set is then evaluated using the rest
the variables and values listed in Table I. Fig. 8 plots the transient of the collected faults, e.g., the SVM trained by the 24 faults
waveforms of |DX (n )6 |2 and |Df 2 (n )3 |2 of an example L-L fault from Set a will be tested by the other 2 856 faults.
(30% mismatch) at t = 5 s, which is the sixth and third level 2) L-L Detection Results and Discussions: The case
wavelet energy of X(n) and f2 , respectively. Based on vPV , studies demonstrate the high accuracy of the proposed algorithm
iPV , and following the data processing procedures elaborated in in detecting PV L-L faults. Table II lists the detailed detection
Section III, features f1 , f2 , and f3 can be extracted, which are accuracy of the proposed method trained by data of STC (Set
shown in Fig. 9. a). The proposed method is able to detect L-L faults in a wide

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on June 17,2021 at 04:07:33 UTC from IEEE Xplore. Restrictions apply.
 8552 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 11, NOVEMBER 2017

Fig. 10. L-L fault detection accuracy as a function of irradiance and

mismatch percentage, SVM trained under STC (Set a).

Fig. 11. L-L fault detection results for different training sets.
range of situations, including those that have not been used
for training the two-stage SVM. For short-circuit faults with
zero fault impedance, the proposed method achieves 100% ac- TABLE III
curacy in most situations. As fault impedance increases, or the L-L FAULT DETECTION RESULTS FOR A PV ARRAY WITH BLOCKING DIODES
mismatch percentage decreases, L-L faults becomes more and
more difficult to be detected, because the impact of such faults Mismatch Fault Impedance (Ω)
on the entire system is minimum and is similar to that caused Percentage(%) 0 5 15 25
by environmental factors, such as irradiance and temperature
changes. For instance, an L-L fault with 10% mismatch and 60 5/5 5/5 5/5 5/5
50 5/5 5/5 5/5 5/5
25-Ω impedance would be an extremely difficult detection case. 40 5/5 5/5 5/5 4/5
However, the average accuracy of the proposed method over 30 5/5 5/5 5/5 4/5
these numerous cases demonstrates its promising performance. 20 5/5 4/5 4/5 4/5
10 5/5 4/5 4/5 4/5
Compared to recent works on PV fault detection, the approach
introduced in Ref.[20] is able to detect L-L faults with high
mismatch percentage; however, it fails to detect any L-L faults
with 20% mismatch, while this proposed method can detect all 3) Case Studies for PV Array with String Blocking
these faults when there is no fault impedance in the short-circuit Diodes: The performance of the proposed scheme is further
path. Another method proposed in [19] is able to detect L-L validated for detecting L-L faults in a PV array with a blocking
faults with 10% mismatch, which is a difficult detection case, diode in each string. The PV system uses the same parameters as
with an accuracy of 1.43% when there is no fault impedance. previous tests. 120 additional cases of L-L faults are applied to
The proposed method improves the accuracy for this situation the PV array, under different combinations of irradiance (300,
to 66.77%. 500, 800, 1000, and 1200 W/m2 ), mismatch percentage (10,
Fig. 10 illustrates the accuracy of the proposed method 20, 30, 40, 50, and 60 %), and fault impedance (0, 5, 15, and
(trained by Set a) as a function of mismatch percentage for 25 Ω). The two-stage SVM is trained using Set d. Table III lists
different irradiance levels. Faults under low irradiance tend to the detection results under different situations, which proves
remain hidden because of low fault current and response of that, blocking diodes does not affect the performance of the
active MPPT algorithms, as is discussed in Section II. It is note- proposed method. It can detect L-L faults with a high accuracy
worthy that, to reduce the complexity and costs of training for in PV systems with string blocking diodes, even when proposed
each training set, the SVM only uses 24 cases as labeled faults, method is trained using data without these blocking diodes. This
which is less than 1% of the test cases (2856). The detection is primarily because the features (f1 , f2 , and f3 ) are extracted
accuracy is expected to increase as the training set expands. based on the rate of change of the array voltage and current.
Moreover, simulation results indicate that the proposed method Although blocking diodes may prevent back-feeding currents,
is able to detect faults under different situations (irradiance, an L-L fault will still lead to a sudden change in the terminal
temperature, fault impedance, point on wave) in spite of the current, which can be detected by the proposed method. The
conditions of training sets. Fig. 11 compares the average ac- performance of the proposed method is expected to be improved
curacy of the proposed method trained by the aforementioned if it can be fine tuned using data from systems with blocking
sets (Set a, b, c, and d). As the SVM is trained by samples un- diodes.
der lower irradiance condition, the proposed classifier is more 4) Case Studies for Disturbances From Change of
sensitive to low-percentage mismatch faults, and the detection Irradiance or Temperature: Weather disturbances, e.g., ir-
accuracy is enhanced: 68.75% for L-L fault with 10% mismatch radiance and temperature fluctuations, are temporary distur-
from Set d (300 W/m2 ), compared to 40.42% of that from Set a bances that occur frequently in PV systems. Since these may
(1000 W/m2 ). also cause disturbances to the operating point, fault detection

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on June 17,2021 at 04:07:33 UTC from IEEE Xplore. Restrictions apply.
 YI AND ETEMADI: L-L FAULT DETECTION FOR PV ARRAYS BASED ON MSD AND TWO-STAGE SVM 8553

methods must be able to differentiate between them and actual

L-L faults. To verify this, the proposed method (trained using
Set a under STC) is further tested by the following cases:
1) 24 irradiance fluctuation cases, with combinations of the
temperature (10, 25, and 40 ◦ C) and the change in irradi-
ance (±200, ±400, ±600, and ±800 W/m2 );
2) 50 temperature fluctuation cases, with combinations of
the irradiance (200, 400, 600, 800, and 1000 W/m2 )
and the change in temperature (±2, ±4, ±6, ±8, and
±10 ◦ C).
Each transition of these changes is finished within 0.1 s. De-
tection results show that, the proposed fault detection scheme
does not show any false positive response to any of the afore-
mentioned disturbance cases. Although the characteristic curves
and operating point of a PV array are changed by a sudden fluc-
tuation of irradiance or temperatures, the transitions caused by
these changes are different from those of L-L faults. One of the
Fig. 12. Experimental setup: 1) Polycrystalline-silicon PV array, 2) in-
differences between the effects from an L-L fault and a tem- terface inverter, 3) data processing station, 4) circuit breakers, 5) fault
porary disturbance is the rate of change of the array voltage impedance, 6) reference panel for irradiance measurement (for reference
and current, which consequently results in distinctive wavelet only).
responses and enables the SVM to separate them in the feature
space by a hyperplane. TABLE IV
5) Case Studies for Disturbances From Partial EXPERIMENTAL SETUP
Shading Conditions: Partial shading condition is another
challenge that may confuse the proposed method, since the Device Parameter

characteristics and the operating point of a PV array are af- PV Panel Polycrystalline Silicon,
fected similarly compared with L-L faults. To prevent hot spots, I SS T
C
C STC
= 0.15 A, V O C = 6V
bypass diodes are installed in the dc junction box of a PV panel. Inverter Model: GTI300W, built-in MPPT,
input: 10.5 to 28 V, output: 90 to 140 V
In normal situations, bypass diodes are reverse biased. When Fuse Threshold = 2.1 I SS T C
= 0.315 A
C
a panel is shadowed, bypass diodes will conduct the extract
current from the healthy panels to protect the shaded panel.
To verify the performance of the proposed algorithm in partial
shading conditions, 40 examples of partial shading disturbances B. Experimental Verification
under 25 ◦ C are performed, with the combinations of the shaded 1) Experimental Setup and Procedures: To further
area on the array(5–50%, with an increment of 5%) and the re- evaluate the performance of the proposed fault detection method
duced irradiance (200, 400, 600, and 800 W/m2 ). The two-stage in practice, a small-scale grid-connected PV system is set up for
SVM is trained by Set a. Five partial shading samples from the experiments. As is shown in Fig. 12, the PV array consists
aforementioned tests (10 %, 20 %, 30 %, 40 %, and 50 % shaded of 7 (parallel) × 4 (series) polycrystalline-silicon PV panels,
area with an irradiance reduced by 400 W/m2 ) are added to the where each string contains a fuse to protect the PV panels. A
nonfault training set. The other 35 partial shading cases are used microinverter is used to interface the PV array and deliver power
to evaluate the proposed method. to the grid, and a data station is used to collect the voltage and
Results show that, the proposed method is able to distinguish current of the PV array. Features are extracted and analyzed
L-L faults from partial shading situation. None of the 35 par- via the data station. Table IV shows the detailed parameter of
tial shading cases is incorrectly classified as an L-L fault. This the experimental setup. One hundred seventy-seven L-L faults
is mainly because L-L faults and partial shading disturbances with different combinations of mismatch percentages (50% and
result in different wavelet patterns that are used as the fault de- 25%) and fault impedances (0, 5, 15, 25 Ω) are tested. The data
tecting features. The signal transients caused by an L-L fault sampling rate is approximately 5000 Hz and the length of sliding
and a shaded panel with a conducting bypass diode are also windows l for features f1 and f3 is 100 samples.
different. Moreover, adding a few partial shading cases in the 2) Experiment Results and Discussions: Among the
nonfault training set allows the SVM to adjust the hyperplane 177 cases, data from four faults with 50% mismatch and another
that separates L-L faults and the nonfault cases, which contain four faults with 25% mismatch (only eight training samples in
partial shading samples. Furthermore, the proposed two-stage total) are used to train the SVM, while others are used to test
SVM increases the capability and accuracy of classification. To the detection system. Table V presents the results of the exper-
further distinguish partial shading situations from other distur- iments. Although the proposed method is trained using only 8
bances, more stages of SVM can be trained and employed in out of the 177 collected faults, it performs satisfactorily with
parallel or further classification steps. a high detection accuracy. More than 90% of the faults are

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on June 17,2021 at 04:07:33 UTC from IEEE Xplore. Restrictions apply.
 8554 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 11, NOVEMBER 2017

TABLE V immediately at their inceptions. Although the experiments are

FAULT DETECTION RESULTS OF EXPERIMENTS
carried out based on a small-scale grid-connected PV system,
the characteristics of this PV system are the same as those in
Mismatch Fault Impedance (Ω) Average
realistic PV power plants, and the impacts of L-L faults to these
Percentage(%) 0 5 15 25 Accuracy systems are similar. Moreover, the proposed method only uses
the array voltage and current, which are usually available in PV
25 22/26 29/29 18/19 16/19 91.40%
50 17/19 19/19 19/19 17/19 94.74%
systems. They are normalized by the open-circuit voltage and
short-circuit current, respectively, which makes the proposed
method scalable. Therefore, it has a great potential to be ap-
plied in large-scale PV systems at a reasonable cost. In practice,
the proposed method can be implemented in a microcontroller,
which monitors the array voltage and current and detects faults.
For different PV systems, the proposed method is applicable
without significant modifications if it can be trained by samples
from these systems. If the actual data of a PV array are not avail-
able, the training set can be generated from simulations, where
a PV model can be built using the same parameters as the ac-
tual system. It is noteworthy that the proposed method requires
only a minimum amount of training data, which considerably
Fig. 13. Normalized voltage and current of a collected L-L fault: 50% reduces the costs and difficulties of training and implementation
mismatch, fault impedance = 0.
in reality. Only 0.83% and 4.7% of the collected faults are used
as the training data in these simulation and experimental case
studies, respectively.

V. CONCLUSION
A fault detection algorithm for PV systems based on pattern
recognition and machine learning techniques is proposed to im-
prove the detection accuracy for challenging L-L fault scenarios
that occur under low irradiance, through a high impedance, or in
interaction with the MPPT scheme. The method takes advantage
of the MSD technique to extract the feature space of L-L faults.
A two-stage SVM classifier is proposed for decision making.
Both simulations and experiments are carried out, which verify
the promising performance of the proposed scheme. The pro-
posed method is economical as it only requires measurements
of the overall voltage and current of a PV array instead of nu-
merous costly sensors. Trained by a minimum portion of data,
this algorithm presents satisfactory accuracy in detecting L-L
faults under different operating conditions.

REFERENCES
[1] M. K. Alam, F. Khan, J. Johnson, and J. Flicker, “A comprehensive re-
view of catastrophic faults in PV arrays: Types, detection, and mitigation
techniques,” IEEE J. Photovolt., vol. 5, no. 3, pp. 982–997, May 2015,
doi: 10.1109/JPHOTOV.2015.2397599.
[2] Y. Zhao, J. F. de Palma, J. Mosesian, R. Lyons, and B. Lehman, “Line-
Fig. 14. Features extracted from the collected L-L fault: 50% mismatch, line fault analysis and protection challenges in solar photovoltaic arrays,”
fault impedance = 0. IEEE Trans. Ind. Electron., vol. 60, no. 9, pp. 3784–3795, Sep. 2013, doi:
10.1109/TIE.2012.2205355.
detected. For faults with 50% mismatch, the accuracy reaches [3] X. Li, Y. Li, J. E. Seem, and P. Lei, “Detection of internal resistance change
for photovoltaic arrays using extremum-seeking control MPPT signals,”
almost 95%. Note that none of the fuses succeeds in interrupting IEEE Trans. Control Syst. Technol., vol. 24, no. 1, pp. 325–333, Jan. 2016,
these faults. The normalized voltage and current of an example doi: 10.1109/TCST.2015.2424857.
L-L fault (50% mismatch, zero impedance) is plotted in Fig. 13, [4] N. L. Georgijevic, M. V. Jankovic, S. Srdic, and Z. Radakovic, “The
detection of series arc fault in photovoltaic systems based on the
where the fault occurs at around 0.6 s. The current (red curve arc current entropy,” IEEE Trans. Power Electron., vol. 31, no. 8,
in Fig. 13) dips at the inception of fault but returns to normal pp. 5917–5930, Aug. 2016, doi: 10.1109/TPEL.2015.2489759.
[5] M. E. Ropp, M. Begovic, and A. Rohatgi, “Analysis and performance
level immediately thereafter because of the MPPT, which veri- assessment of the active frequency drift method of islanding prevention,”
fies the analysis in Section II. The features extracted from this IEEE Trans. Energy Convers., vol. 14, no. 3, pp. 810–816, Sep. 1999, doi:
fault are presented in Fig. 14. All the faults are detected almost 10.1109/60.790956.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on June 17,2021 at 04:07:33 UTC from IEEE Xplore. Restrictions apply.
 YI AND ETEMADI: L-L FAULT DETECTION FOR PV ARRAYS BASED ON MSD AND TWO-STAGE SVM 8555

[6] M. H. Wang, M. L. Huang, and K. J. Liou, “Islanding detection [25] L. N. Khanh, J. J. Seo, Y. S. Kim, and D. J. Won, “Power-management
method for grid connected photovoltaic systems,” IET Renewable strategies for a grid-connected PV-FC hybrid system,” IEEE Trans.
Power Gener., vol. 9, no. 6, pp. 700–709, 2015, doi: 10.1049/iet-rpg. Power Del., vol. 25, no. 3, pp. 1874–1882, Jul. 2010, doi: 10.1109/TP-
2014.0264. WRD.2010.2047735.
[7] K. A. Kim, G. S. Seo, B. H. Cho, and P. T. Krein, “Photovoltaic hot-spot [26] T. Klassen, “Ground fault protection—Ground fault protection for
detection for solar panel substrings using ac parameter characterization,” solar applications,” Internet, 2012. [Online]. Available: http://m.
IEEE Trans. Power Electron., vol. 31, no. 2, pp. 1121–1130, Feb. 2016, littelfuse.com/∼/media/electrical/white-papers/littelfuse_ground_fault_
doi: 10.1109/TPEL.2015.2417548. solar_applications_white_paper.pdf
[8] A. Colli and W. J. Zaaiman, “Maximum-power-based PV perfor- [27] “Line-line fault analysis and protection in PV arrays,” Mersen, Newbury-
mance validation method: Application to single-axis tracking and fixed- port, MA, USA, Tech. Rep., Photovoltaic Protection, Note 2, Issue 1,
tilt c-Si systems in the italian alpine region,” IEEE J. Photovolt., 2011.
vol. 2, no. 4, pp. 555–563, Oct. 2012, doi: 10.1109/JPHOTOV.2012. [28] Article 690—Solar Photovoltaic Systems, National Electrical Code
2203794. NFPA70, 2014.
[9] C. L. Kuo, J. L. Chen, S. J. Chen, C. C. Kao, H. T. Yau, and C. [29] Outline of Investigation for Low-Voltage Fuses—Fuses for Photovoltaic
H. Lin, “Photovoltaic energy conversion system fault detection using Systems, Underwriters Laboratories, Inc, UL Subject 2579, 2010.
fractional-order color relation classifier in microdistribution systems,” [30] Standard for Safety for Flat-Plate Photovoltaic Modules and Panels, Un-
IEEE Trans. Smart Grid, vol. 8, no. 3, pp. 1163–1172, May 2017, doi: derwriters Laboratories, UL Standard 1703, 2004.
10.1109/TSG.2015.2478855. [31] J. C. Schaefer, “Review of photovoltaic power plant performance and
[10] R. Platon, J. Martel, N. Woodruff, and T. Y. Chau, “Online fault detection economics,” IEEE Trans. Energy Convers., vol. 5, no. 2, pp. 232–238,
in PV systems,” IEEE Trans. Sustain. Energy, vol. 6, no. 4, pp. 1200–1207, Jun. 1990, doi: 10.1109/60.107215.
Oct. 2015, doi: 10.1109/TSTE.2015.2421447. [32] Z. Yi and A. H. Etemadi, “Fault detection for photovoltaic systems based
[11] S. Vergura, G. Acciani, V. Amoruso, G. E. Patrono, and F. Vacca, “Descrip- on multi-resolution signal decomposition and fuzzy inference systems,”
tive and inferential statistics for supervising and monitoring the operation IEEE Trans. Smart Grid, vol. 8, no. 3, pp. 1274–1283, May 2017, doi:
of PV plants,” IEEE Trans. Ind. Electron., vol. 56, no. 11, pp. 4456–4464, 10.1109/TSG.2016.2587244.
Nov. 2009, doi: 10.1109/TIE.2008.927404. [33] A. H. Etemadi and M. Sanaye-Pasand, “High-impedance fault detec-
[12] P. Guerriero, F. D. Napoli, G. Vallone, V. d’Alessandro, and tion using multi-resolution signal decomposition and adaptive neural
S. Daliento, “Monitoring and diagnostics of PV plants by a wire- fuzzy inference system,” IET Gener., Transmiss. Distrib., vol. 2, no. 1,
less self-powered sensor for individual panels,” IEEE J. Photovolt., pp. 110–118, Jan. 2008, doi: 10.1049/iet-gtd:20070120.
vol. 6, no. 1, pp. 286–294, Jan. 2016, doi: 10.1109/JPHOTOV.2015. [34] C.-H. Kim, H. Kim, Y.-H. Ko, S.-H. Byun, R. K. Aggarwal, and A.
2484961. T. Johns, “A novel fault-detection technique of high-impedance arcing
[13] P. Guerriero, G. Vallone, V. d’Alessandro, and S. Daliento, “Innovative faults in transmission lines using the wavelet transform,” IEEE Trans.
algorithm for true maximum detection based on individual PV panel sen- Power Del., vol. 17, no. 4, pp. 921–929, Oct. 2002, doi: 10.1109/TP-
sor network,” in Proc. 2013 Int. Conf. Clean Elect. Power, Jun. 2013, WRD.2002.803780.
pp. 42–47, doi: 10.1109/ICCEP.2013.6586963. [35] S.-J. Huang, T.-M. Yang, and J.-T. Huang, “FPGA realization of wavelet
[14] B. And, S. Baglio, A. Pistorio, G. M. Tina, and C. Ventura, “Sen- transform for detection of electric power system disturbances,” IEEE
tinella: Smart monitoring of photovoltaic systems at panel level,” IEEE Trans. Power Del., vol. 17, no. 2, pp. 388–394, Apr. 2002, doi:
Trans. Instrum. Meas., vol. 64, no. 8, pp. 2188–2199, Aug. 2015, doi: 10.1109/61.997905.
10.1109/TIM.2014.2386931. [36] J. Barros, R. I. Diego, and M. de Apraiz, “Applications of wavelet trans-
[15] Y. Hu et al., “Online two-section PV array fault diagnosis with optimized form for analysis of harmonic distortion in power systems: A review,”
voltage sensor locations,” IEEE Trans. Ind. Electron., vol. 62, no. 11, IEEE Trans. Instrum. Meas., vol. 61, no. 10, pp. 2604–2611, Oct. 2012,
pp. 7237–7246, Nov. 2015, doi: 10.1109/TIE.2015.2448066. doi: 10.1109/TIM.2012.2199194.
[16] R. Hariharan, M. Chakkarapani, G. S. Ilango, and C. Nagamani, “A [37] I. Daubechies, “Orthonormal bases of compactly supported wavelets II.
method to detect photovoltaic array faults and partial shading in PV sys- Variations on a theme,” SIAM J. Math. Anal., vol. 24, no. 2, pp. 499–519,
tems,” IEEE J. Photovolt., vol. 6, no. 5, pp. 1278–1285, Sep. 2016, doi: 1993.
10.1109/JPHOTOV.2016.2581478. [38] S.-J. Huang and C.-T. Hsieh, “Coiflet wavelet transform applied to inspect
[17] L. Chen, S. Li, and X. Wang, “Quickest fault detection in photo- power system disturbance-generated signals,” IEEE Trans. Aerosp. Elec-
voltaic systems,” IEEE Trans. Smart Grid, vol. PP, no. 99, p. 1 doi: tron. Syst., vol. 38, no. 1, pp. 204–210, Jan. 2002, doi: 10.1109/7.993240.
10.1109/TSG.2016.2601082. [39] M. A. Green, “Solar cell fill factors—General graph and empirical expres-
[18] M. N. Akram and S. Lotfifard, “Modeling and health monitoring of dc sions,” Solid State Electron., vol. 24, pp. 788–789, 1981.
side of photovoltaic array,” IEEE Trans. Sustain. Energy, vol. 6, no. 4, [40] K. L. V. Iyer, X. Lu, Y. Usama, V. Ramakrishnan, and N. C. Kar, “A
pp. 1245–1253, Oct. 2015, doi: 10.1109/TSTE.2015.2425791. twofold Daubechies-wavelet-based module for fault detection and volt-
[19] Y. Zhao, R. Ball, J. Mosesian, J. F. de Palma, and B. Lehman, “Graph- age regulation in SEIGs for distributed wind power generation,” IEEE
based semi-supervised learning for fault detection and classification in Trans. Ind. Electron., vol. 60, no. 4, pp. 1638–1651, Apr. 2013, doi:
solar photovoltaic arrays,” IEEE Trans. Power Electron., vol. 30, no. 5, 10.1109/TIE.2012.2188258.
pp. 2848–2858, May 2015, doi: 10.1109/TPEL.2014.2364203. [41] J. A. Jiang et al., “A hybrid framework for fault detection, classifica-
[20] Y. Zhao, L. Yang, B. Lehman, J.-F. de Palma, J. Mosesian, and R. Lyons, tion, and location—Part II: Implementation and test results,” IEEE Trans.
“Decision tree-based fault detection and classification in solar photovoltaic Power Del., vol. 26, no. 3, pp. 1999–2008, Jul. 2011, doi: 10.1109/TP-
arrays,” in Proc. IEEE 27th Annu. Appl. Power Electron. Conf. Expo., WRD.2011.2141158.
Feb. 2012, pp. 93–99, doi: 10.1109/APEC.2012.6165803. [42] Z. Moravej, M. Pazoki, and M. Khederzadeh, “New pattern-recognition
[21] Y. Zhao, B. Lehman, R. Ball, J. Mosesian, and J.-F. de Palma, “Out- method for fault analysis in transmission line with UPFC,” IEEE Trans.
lier detection rules for fault detection in solar photovoltaic arrays,” in Power Del., vol. 30, no. 3, pp. 1231–1242, Jun. 2015, doi: 10.1109/TP-
Proc. IEEE 28th Annu. Appl. Power Electron. Conf. Expo., Mar. 2013, WRD.2014.2365674.
pp. 2913–2920, doi: 10.1109/APEC.2013.6520712. [43] R. K. Begg, M. Palaniswami, and B. Owen, “Support vector machines
[22] Y. Hu, W. Cao, J. Ma, S. J. Finney, and D. Li, “Identifying PV mod- for automated gait classification,” IEEE Rev. Biomed. Eng., vol. 52, no. 5,
ule mismatch faults by a thermography-based temperature distribution pp. 828–838, May 2005, doi: 10.1109/TBME.2005.845241.
analysis,” IEEE Trans. Device Mater. Rel., vol. 14, no. 4, pp. 951–960, [44] D. J. Sebald and J. A. Bucklew, “Support vector machine techniques for
Dec. 2014, doi: 10.1109/TDMR.2014.2348195. nonlinear equalization,” IEEE Trans. Signal Process., vol. 48, no. 11,
[23] S. Kouro, J. I. Leon, D. Vinnikov, and L. G. Franquelo, “Grid-connected pp. 3217–3226, Nov. 2000, doi: 10.1109/78.875477.
photovoltaic systems: An overview of recent research and emerging PV [45] S. Alshareef, S. Talwar, and W. G. Morsi, “A new approach based on
converter technology,” IEEE Ind. Electron. Mag., vol. 9, no. 1, pp. 47–61, wavelet design and machine learning for islanding detection of distributed
Mar. 2015, doi: 10.1109/MIE.2014.2376976. generation,” IEEE Trans. Smart Grid, vol. 5, no. 4, pp. 1575–1583,
[24] S. B. Kjaer, J. K. Pedersen, and F. Blaabjerg, “A review of single- Jul. 2014, doi: 10.1109/TSG.2013.2296598.
phase grid-connected inverters for photovoltaic modules,” IEEE Trans. [46] Y. Wang, S. Wang, and K. K. Lai, “A new fuzzy support vector machine to
Ind. Appl., vol. 41, no. 5, pp. 1292–1306, Sep. 2005, doi: 10.1109/ evaluate credit risk,” IEEE Trans. Fuzzy Syst., vol. 13, no. 6, pp. 820–831,
TIA.2005.853371. Dec. 2005, doi: 10.1109/TFUZZ.2005.859320.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on June 17,2021 at 04:07:33 UTC from IEEE Xplore. Restrictions apply.
 8556 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 11, NOVEMBER 2017

Zhehan Yi (S’13) received the B.Sc. degree in Amir H. Etemadi (S’08–M’12) received the
electrical engineering from Beijing Jiaotong Uni- B.Sc. degree from the University of Tehran,
versity, Beijing, China, in 2012, and the M.Sc. de- Tehran, Iran, in 2005, the M.Sc. degree from
gree in electrical engineering from The George the Sharif University of Technology, Tehran, in
Washington University, Washington, DC, USA, 2007, and the Ph.D. degree from the University
where he is currently working toward the Ph.D. of Toronto, Toronto, Canada, in 2012, all in elec-
degree in electrical engineering. trical engineering.
He is currently a Graduate Research As- He is currently an Assistant Professor with
sistant with The George Washington University. the Department of Electrical and Computer
His research interests include distributed gener- Engineering, The George Washington Univer-
ations, power electronics, microgrid, and solar sity, Washington, DC, USA. His research inter-
PV systems. ests include distributed generations and power system dynamics and
control.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on June 17,2021 at 04:07:33 UTC from IEEE Xplore. Restrictions apply.