Machine Learning for PV System Operational

Fault Analysis: Literature Review?

Tarikua Mekashaw Zenebe1,2 , Ole-Morten Midtgård1,3 , Steve Völler1,4 , and

Ümit Cali1,5
1
Norwegian University of Science and Technology, Trondheim, Norway
2
[email protected]
3
[email protected]
4
[email protected]
5
[email protected]

Abstract. This review paper aims to discover the research gap and
assess the feasibility of a holistic approach for photovoltaic (PV) system
operational fault analysis using machine learning (ML) methods. The
analysis includes the detection and diagnosis of operational faults in
a PV system. Even if standard protective devices are installed in PV
systems, they fail to clear various faults because of low current during low
mismatch levels, high impedance fault, low irradiance, etc. This failure
will increase the energy loss and endanger the PV system’s reliability,
stability, and security. As a result of the ML method’s ability to handle
a non-linear relationship, distinguishing features with similar signatures,
and their online application, they are getting attractive in recent years
for fault detection and diagnosis (FDD) in PV systems. In this paper, a
review of literature on ML-based PV system FDD methods is provided.
It is found that considering their simplicity and performance accuracy,
Artificial Neural networks such as Multi-layer Perceptron are the most
promising approach in finding a central PV system FDD. Besides, the
review paper has identified main implementation challenges and provides
recommendations for future work.

Keywords: Ensemble learning · Fault detection and diagnosis · Machine

learning · PV system fault · Transfer learning.

1 Introduction
Owning to the various advantages PV system can provide, the global market
for PV has been increasing sharply. According to [1], the cumulative globally
installed capacity in 2019 increased to about 627 GW. Assuming a medium
scenario where cases like COVID-19 pandemic considered, [14] estimated the
total global installed PV generation capacity to exceed 1.2 TW by 2022. In
addition, the price of electricity from a PV system is constantly decreasing [20].
This all shows a promise for further increase in the PV market in the coming
years.
?
Supported by NTNU
 Nomenclature

AC Alternative Current LST M Long Short Term Memory

AN N Artificial Neural Network M IM O Multiple Input Multiple Output
CN N Convolutional Neural Network ML Machine Learning
DA Discriminant Analysis M LP Multi-layer Perceptron
DC Direct Current M P P T Maximum Power Point Tracker
DL Deep Learning OCP D Over Current Protection Devices
DT Decision Tree RBF Radial Basis Function
DW T Discrete Wavelet Transform
RF Random Forest
EL Ensemble Learning
SAP V S Stand Alone PV System
F DD Fault Detection and Diagnosis
G Irradiance at Array SCADA Supervisory Control and Data Ac-
GCP V S Grid Connected PV System quisition
GF DI Ground Fault Detection and Inter- SOC State of Charge
ruption ST C Standard Test Condition
I Current SV M Support Vector Machine
IM P P Current at Maximum Power Point T Module Temperature
IGBT Insulated Gate Bipolar Transistor TL Transfer Learning
KELM Kernel Based Extreme Learning Ma- V Voltage
chine VM P P Voltage at Maximum Power Point

With the increase in PV system deployment for electricity production, en-

suring the system’s reliability, stability, and safety is crucial. However, despite
the advancement in technology and implementation of standards such as the
National Electric Code (NEC) article 690 [33], still, faults are problems for the
efficient and effective operation of a PV system. Because there are various con-
ditions where the protective devices fail to clear fault on time. For example,
according to NEC, the fuse rating should be greater than 2.1 times the short-
circuit current at standard test condition (STC) in PV system [31]. However, if
line to line (LL) fault occurred at low mismatch and high impedance level, the
fuse will not be able to clear the fault as the current will not be enough to blow
the fuse [33]. In addition, due to the blocking diodes, which prevent the string
from back-feed current, the protective devices may fail to interrupt the fault
current even under STC. Moreover, these diodes may fail and lead to serious
damage [12].
As mentioned in [27], the annual energy loss due to various faults might go
up to 18.9%. This reported power loss is very significant as the efficiency of a
typical PV cell range between 15-21%. Unless the faults cleared on time, they
might also cause additional damages to other property in case of fire. Therefore,
detecting and clearing the fault on time is an indispensable solution to mitigate
these losses while ensuring the reliability and security of a PV system.
Up to now, several techniques have been created for FDD in a PV system.
However, the demand for techniques which is simple and cheap, can handle non-
linear nature of the PV modules, can be remotely applied and can differentiate
features with similar signature are the primary motivation to move to data-
driven methods like that of machine learning (ML) [27] for many researchers in
recent years.
If ML is used to analyze the fault in a PV system, as much as possible, there
should be a holistic method that is used to detect and diagnosis at least all the
 ML for PV System Operational Fault Analysis 3

most frequent and dangerous faults. Nevertheless, most of the literature focused
on PV array fault. Besides, only one paper tried to implement the ML method in
a programmable logic device based on the author’s knowledge. Thus, this paper
aim at answering the reason behind. To the best of the author’s knowledge, this
review paper is the first to review literature, keeping in mind the feasibility of a
holistic fault analysis approach for a PV system specifically for standalone PV
system (SAPVS) as well as categorizing and analyzing FDD into methods based
on ML and deep learning, ensemble learning, and transfer learning.
The paper is organized as follows: After providing a summary of review pa-
pers, the first part of section 2 gives detailed information about various faults
commonly occurring in PV system components. Then, the second part of section
2 provides a comprehensive literature review on PV FDD. Whereas, section 3
presents and discuss all the findings. Finally, the paper concludes by summariz-
ing the main findings and providing some recommendations.

2 Literature Review

Pillai et al. [33] provided a comprehensive literature review on PV faults and

advanced detection techniques. The paper tried to review literature, including
all PV faults. However, most of the discussion focused on PV array faults. Mellit
et al. [31] presented very detailed information about PV faults, including FDD
methods. But similar to [33], the main focus of the paper was on PV array
faults. [40] reviewed papers on the role of artificial intelligence on modeling,
sizing, control, fault diagnosis, and output estimation of PV systems. Whereas,
Li et al. [24] reviewed recent work specifically applying Artificial Neural Network
(ANN) and hybrid ANN for FDD based on the fault they analyzed, the type
and amount of data they used, their model’s configuration, and its performance.
Besides, they highlighted the major challenges and prospects of the methods. [16]
is among the papers dedicated to explaining the PV system faults in a wider
spectrum. Fault detection methods on grid-connected PV system (GCPVS) were
studied comprehensively in [26]. Contrary to most of the review papers, the
current paper focuses only on the advanced data-driven approach that of ML.
This section will discuss different PV system faults classified based on com-
ponents and the various ML methods used in PV system FDD.

2.1 PV System Fault

To design an efficient and effective fault detection and diagnosis method, it is

necessary to know about the character of each fault, including their protection
challenges [33].
SAPVS comprises of PV array, inverter, battery, charge controller, MPPT,
connection wires, and other additional protection and safety devices.

PV Array Fault Some of the PV array faults are discussed below.

4 T. Zenebe et al.

Open Circuit Fault (OC) OC fault is an intentional disconnection of a closed-

loop that results in interruption of current flow due to breakage of the cable
that connects two strings, any object falling on panels, loose connection between
two points, or an accidental disconnection at a current-carrying conductor [4].
In addition, broken cells, physical breakdown of cable joints, loose connections,
and aged power cables near terminal may lead to OC fault [27]. Due to the
presence of a bypass diode, current flow will be kept even if an OC fault occurs.
In addition, it results in a substantial power loss due to the reduction in voltage
in a string [27].

Line to Line Fault (LL) LL fault is an unintentional connection between lines

with different potential difference [4, 31, 41] due to cable insulation failures, me-
chanical damage, water ingress, D-junction box corrosion, and hot spots caused
by the back-sheet failures [31]. LL fault could lead to serious problems like fire
hazards in addition to degrading PV arrays lifetime. LL fault is very hard to
identify by the conventional protection devices such as Over Current Protection
Devices (OCPD) that is mainly: 1) as a result of the decrease in current in cases
of LL fault during high impedance and low mismatch level [31, 33], 2) due to the
presence of a blocking diode as it blocks back-fed current [12, 33], 3) as the pres-
ence of MPPT decreases the current to optimize the power output and difficult
to distinguish it from normal cases [12, 33], 4) its similarity with ground fault
[33], 5) as a result of low current at low irradiance values [33].

Ground Fault (GF) GF occurs when a current-carrying wire/cell/module con-

nected with a ground accidentally. It can be detected by Ground Fault Detection
and Interruption (GFDI) and Ground Fault Protection Devices in a normal sce-
nario. However, during high impedance cases, detection is challenging as the
current will be low. In addition, there are scenarios where it looks like SC fault
[33]. Thus, this fault also needs an efficient method to detect and distinguish it
from other faults.

Arc Fault (AF) AF is a fault where current flows in the air or dielectric outside
the conductor due to loose connection. It could be a series arc in case of a
connection between modules or a parallel arc in case of a closely placed conductor
at different potential differences [4, 33, 31]. On the contrary to other faults, arc
fault has little effect on the I-V or I-P characteristic of PV arrays. Nevertheless,
it leads to a severe distortion in the output current and voltage waveform [33].
Arc Fault Circuit Interrupters and Arc Fault Detectors are recommended for
clearing this fault. However, multiple of them have to be installed to clear the
fault correctly. Moreover, when they are installed at the inverter side, they fail
to protect the fault as attenuated arc signals reach them. Beside detecting arc
fault, identifying which arc fault is occurring is important as the measure taken
for one will increase the impact of the other [33].

Partial Shading (PS) In addition to decreasing and resulting in continuous fluc-

tuating PV output power [4, 6], PS facilitates the degradation of PV arrays [33].
 ML for PV System Operational Fault Analysis 5

Even it can lead to destruction due to fire hazards as a result of cell/module tem-
perature increase due to the dissipated energy [27]. Furthermore, as it results
in multiple peaks in I-V characteristic curve, identifying the maximum power
point by MPPT will be challenging [33]. Besides, unless a time factor is used, it
is hard to differentiate it from OC fault as their effect on power output charac-
teristic has similarity [27]. Furthermore, to mitigate the problem, bypass diodes
are installed at each module, but this will increase the installation cost [33].

Others In addition to the above main PV array faults others may include degra-
dation faults [18, 34], hot spot fault [39], fault in bypass diode which could be
OC or SC fault [33], and blocking diodes faults [30].

Solar Battery Fault The battery takes around 43% of the life cycle cost of
SAPVS [37]. As a result, it shall get attention, and a good working condition shall
be provided. The main faults that could happen in this PV component includes
external short-circuit fault [36], degradation fault [35], internal fault which could
be GF and SC fault [32], overcharging (over-voltage), undercharging (under-
voltage) and open circuit (total voltage to zero) [39]. The impact of those faults
in a battery may range from decreasing its performance, shorten its lifetime, and
increased maintenance cost to fire hazard explosion [32]. The lack of guidelines
on how to select fuse and circuit breaker is mentioned in [32] as one of the main
challenges in detecting internal faults. Moreover, the gradual change of current
and voltage of a battery makes detecting faults on time extremely difficult.

Inverter Fault Inverter faults may include the OC of switches, SC of switches,

filter failure, and gating failure [31]. For instance, the gate failure could be an
incipient fault of the Insulated Gate Bipolar Transistor (IGBT). IGBT is the
most critical component in an inverter. It is also one of the main reasons for
the failure of inverters. So if the incipient faults of the IGBT can be identified,
the reliability of the PV system can be enhanced. Nevertheless, a procedure is
needed to generate this fault to train and validate ML algorithms. Thus, Ismail
et al. [21] provided the way to generate this fault.

MPPT Fault An MPPT control system comprises various sensors to get ir-
radiance, temperature, current, and voltage measurements and an optimization
algorithm to search the maximum power point to operate the PV array and
boost the PV system yield. Thus, any error in any part of the MPPT will lead
to a wrong operating power point, which significantly decreases the PV sys-
tem’s output power. Sensor failure and lack of an efficient and effective MPPT
algorithm are the most common fault in MPPT [27, 29].

2.2 PV System Fault Detection and Diagnosis Methods (FDD)

In this paper, fault detection indicates the process of identifying a fault occur-
rence, while fault diagnosis comprises the process of finding the type of fault
6 T. Zenebe et al.

and localizing the occurrence. This section closely look at the literature on ML
application for FDD in PV systems by classifying them into methods based on
1) machine learning and deep learning, 2) ensemble learning, and 3) transfer
learning.

Methods based on Machine Learning and Deep Learning Similar find-

ings and a comprehensive explanation about ML and deep learning with respec-
tive of PV application can be found in the book chapter in [30]. PV system fault
has been detected and diagnosed using supervised machine learning such as Sup-
port Vector Machine (SVM), Naive Bayes (NB), k-Nearest Neighbors (KNN),
Random Forest (RF), Decision Tree (DT), Discriminant Analysis (DA), Radial
Basis Function (RBF), and deep learning like Multi-Layer Perceptron (MLP)
and Convolutional Neural Network (CNN).
Among other works [38] used SVM with a higher classification accuracy using
a climate corrected performance ratio. However, the paper fails to mention which
fault has been analyzed specifically besides it is a fault or normal operation.
Similarly Dong et al. [11] proposed an FDD method based on SVM while using
available SCADA (Supervisory Control and Data Acquisition) data. In addition,
as an input they used an index called anomaly detection index. However, this
paper focuses only on PV string faults.
In order to analyze multiple faults Hajji et al. [17] tried to include a feature
extraction and selection stage using principal component analysis. They have
tested various classifiers like KNN, RF, DA, NB, DT, and SVM to classify fault in
GCPVS. To evaluate the performance of the classifier, they have used additional
metrics. They all have achieved an accuracy greater than 96%. In addition, the
execution time of each classifier was evaluated. Relative to other papers, they
have included inverters, MPPT, and DC-DC converter in addition to PV array.
Nevertheless, battery fault is not analyzed.
In [23], the authors used a comparison between model and real system output
to identify fault occurrence for GCPVS. First, they have tested various linear and
nonlinear models of PV system. Then, they used ML techniques like KNN, DT,
SVM, and MLP to identify faults such as SC, OC, degradation, and shadowing.
MLP is found to be a suitable and more accurate ML algorithm. Basnet et
al.[5] has used MLP, to detect and classify LL and GF in GCPVS as well. They
could achieve 100% training accuracy. As input parameters, voltage, current,
irradiance, average temperature of each module, and weather conditions were
utilized. Despite getting a good accuracy, both papers [23, 5] application’s is
limited to PV array fault.
[10] is among the few papers which evaluate the ML algorithm’s performance
based on both accuracy and execution time. Faults like module SC, MPPT
fault, OC, PS, and degradation have been detected and classified using five
ML techniques: kNN, DT, SVM, and ANN. ANN resulted in higher accuracy
( 99.65% ) even though it took longer computational time. Nevertheless, for
generalization, ANN should be tested by incorporating other faults.
 ML for PV System Operational Fault Analysis 7

A cascaded Probabilistic neural network (PNN), due to its robustness to

noise, has been used in [15] to detect and classify a different number of module
SC and string OC faults. In addition, the result compared with a feed-forward
back-propagation ANN with both noisy and noiseless data. As input features,
they have utilized temperature, tilted irradiance, current (IM P P ), and voltage
(VM P P ) at MPP. The training data set is generated from a validated one diode
PV system model. Despite the effort made to bring a robust method, the lit-
erature focused on faults on a PV array and DC side of the GCPVS. With a
similar focus on GCPVS, DT was used in [25] for detecting PS, inverter, and
bypass diode failure. The authors has also used other methods for detecting
and classifying these faults. They achieved an average classification accuracy of
98.7%.
A kernel-based extreme learning machine (KELM) was used in [7] to detect
and classify degradation fault, OC, SC, and PS. Features that enable the identifi-
cation of faults are extracted after examining their impact on I-V characteristics.
Besides, the PV model was validated before using it to simulate the fault to gen-
erate a training and test data set. In addition to the simulation data, a real
laboratory PV array data set has been used. In general, even if a very efficient
and accurate method is devised here, the determination of I-V characteristics of
the PV array in an online scenario might be problematic.
[35] is one of the few papers which has focused specifically on faults in SAPVS.
They proposed a fault diagnosis method using MLP feed-forward neural network
to detect and classify faults. The faults include SC of PV module, OC of PV
module, and external SC of a battery where the fuse fails to clear them in low
irradiance condition. Though most papers entirely focus on PV array fault, this
paper included the battery and load fault. Only electrical measurements like
current and voltage are used for validation using experimental data from an
existing PV system in Algeria. They have achieved 96% test and 97.8% training
accuracy. Nevertheless, to consider as a valid method for complete SAPVS fault
analysis, it shall be verified including the missing other faults like inverters fault,
MPPT fault, and others.
Chine et al. [8] used a combination of threshold method and ML to detect
and classify eight faults, including SC, bypass diode fault, OC, connection fault,
shadow effect and etc. From ML, MLP and RBF have been compared. Extracted
attributes like the current, voltage, and peaks from I-V characteristics were
used as input. In addition to showing the method’s feasibility, this is the only
paper encountered that shows a prototype by implementing the ML in a Field
Programmable Gate Array (FPGA). However, they used simulation data for
training and testing the models. The other drawback of this paper is that they
also applied threshold method which can be very much dependent on system
parameters and the accuracy of the threshold limits.
The authors in [12] focused on the detection of LL fault in PV system under
high impedance fault and low mismatch condition, which is one of the cases where
protection devices fail to clear a fault. The authors used an SVM classifier that
is resistant to model error and computational efficiency, based on the features
8 T. Zenebe et al.

extracted by analyzing I-V characteristics of a PV array. As the methods were

validated for LL fault, further analysis is needed to apply it for multiple input
and output (MIMO) cases in SAPVS. Moreover, validation is needed with real
PV system data.
Instead of threshold method, Ahmad et al. [2] has used a combination of
transformation for feature extraction and ML algorithm for detecting and clas-
sifying PS condition as a modular fault, DC-DC converter switch SC, inverter
switch OC, inverter switch SC with LCL filter failure and gating circuit failure.
They used discrete wavelet transform (DWT) due to its less computational time
and complexity, as well as it enables us to work both in the time and frequency
domain. Whereas the ML algorithm is MLPNN. The data set was obtained from
a simulated a PV system. They can achieve an accuracy greater than 99%. How-
ever, the battery and MPPT fault is missing. Furthermore, other faults in PV
array such as LL has to be checked.
[3] used a hybrid features-based support vector machine (SVM) model in or-
der to detect and classify hot spot fault in PV array using infrared thermography
as an input image for the model. The model could detect and classify with 96.8%
and 92% training and testing accuracy, respectively. This paper is dedicated to
one fault only. For small-scale SAPVS using this individual method will not be
cost-effective.
Improper operation is One of the reasons for the short lifetime of solar bat-
teries. In addition to the available energy from solar or demanded load, to decide
whether the battery has to be charged or discharged, knowing the battery capac-
ity accurately is a determining factor. There are various statistical estimation
techniques, but recently SOC (state of charge) estimation using ML is getting
attractive as it exhibits non-linear input-output characteristics. [9] presented an
ML-based SOC estimation method for the most common solar battery, which
is a lead-acid battery. The proposed methods are based on a feed-forward neu-
ral network, a recurrent neural network, and an adaptive neuro-fuzzy inference
system. As an input feature for the model, voltage and current data were used.
The findings of this paper could be used for further studying FDD methods in
PV batteries. However, the paper did not mention how the training SOC data
is obtained.
In [39] internal resistance effect and overcharging problem in lead-acid bat-
tery in a PV system was detected using solar radiation data estimated from
satellite image analysis. The paper showed the impact of overcharging and inter-
nal resistance fault on the battery voltage and SOC. Even if the paper used the
ML for estimating the solar irradiation, from the finding, there is an indication
for using battery voltage to detect and classify faults in a battery. In another
study in [19], the author has used a long short-term memory (LSTM) recurrent
neural network for state prediction and fault prognosis for battery in an electric
vehicle. This approach could also be used for solar batteries from the knowledge
domain, and its finding is significantly important though its feasibility has to be
checked in a solar battery.
 ML for PV System Operational Fault Analysis 9

One of the challenges in using ML, especially in analyzing inverter fault in

a PV system, is the lack of methods that guide us in generating the faults in
simulation as the faults do not occur frequently, but they are responsible for the
majority of inverter failures. Thus, Ismail et al. [21] used a feed-forward back
propagation neural network to detect SC incipient faults by first modeling a way
to generate this fault for GCPVS. For using it in SAPVS, others fault has to
be incorporated, and the method has to be verified for its performance for other
PV system components fault.
In addition to supervised learning, unsupervised ML methods has been used
in PV system FDD. In [37] an internal fault detection for solar battery using
unsupervised ML algorithm based on anomaly detection method has been pro-
posed. The intuition for using unsupervised learning is whenever it is difficult
to obtain a labeled data set, which is the case in solar battery fault analysis for
using ML. The internal faults investigated are SC and GF. The data set was
generated using simulation of SAPVS using irradiance and temperature data
from Algeria. They have used readily available current and voltage data set. As
the work is only for internal fault, it is important to incorporate it/hybridize it
with other methods to identify other faults in SAPVS.

Methods based on ensemble learning (EL) [13], similarly to [12] the model
is based on I-V characteristics and focus on LL fault at different mismatch and
impedance level. However, here they used probabilistic ensemble learning model
comprising of SVM, NB, and KNN. For decision, the average of all the results
of the algorithm was used. They could achieve an average of 99% and 99.5%
for detecting and classifying LL fault. Moreover, they have evaluated the model
with simulation and experimental data set. In [22] EL method with DT, RF,
DA, etc., was used to detect PS and SC fault, but the focus is still PV array.
They have used electrical parameters as input features.

Methods based on Transfer Learning (TL) In order to detect and classify

PV system faults, in [4], the concept of transfer learning has been employed
by using a pre-trained AlexNetCNN for feature extraction and classification
to minimize the impact of low data set in the model performance. They also
proposed a deep 2-D CNN to extract 2-D scalograms generated from a PV
system. The authors analyzed faults like PS, LL, OC, high-impedance series
/arc fault, and faults in PS with the presence of MPPT. A detecting accuracy of
73.53% and a classification accuracy of 70.45% were achieved. They have noticed
the decreasing of performance as the number of class increase. They have used
deep learning in-depth and also made a comparison with classical ML models.
They handled MIMO data. Though it needs to be verified, the methods they
followed seem promising for SAPVS fault analysis. Even though it is only for
inverter’s fault in GCPVS, in [28], TL was also used to detect faults like SC, OC
using ResNet with an accuracy greater than 97%.
10 T. Zenebe et al.

3 Result and Discussion

As we can see from Table 1 the majority of the papers, greater than 80%, has
analyzed the fault in PV array. However, faults in an inverter and a battery have
also been investigated. Relatively, faults in GCPVS have got special attention
than SAPVS. SC, OC, and PS in PV array are the most investigated type of
fault using ML methods though faults like GF, AF, and LL, are the most severe.

Among other ML methods, SVM and MLP, in general, have been used exten-
sively to detect and classify faults in a PV system. For evaluating the models,
accuracy and confusion metrics are the most employed performance indices.
However, some have utilized their own metrics and execution time. Due to ML’s
random nature, it is very important to report performance after conducting a
reasonable number of model execution though it takes time.
Looking at the data source where the experimental PV system was installed,
Algeria took the lead, China and Korea take second place. It is also very impor-
tant to analyze the performance of the ML methods under different climatic and
geographical conditions before utilizing them. This is because the challenge for
the PV array and the battery is different depending on the geographical location.
For instance, while snow is a big problem in the polar region, dust, soiling, and
higher operating temperature are huge problems in the equatorial region.
Most of the papers depend on the input features which has been generated
from a simulated PV system. Whereas only a few have included experimental
data. This is because of the difficultly of setting up a PV system only for collect-
ing data. Furthermore, when an available PV system exists as the environmental
condition can not be controlled, it is tedious and time-consuming to generate
a data set that will enable the model to acquire a generalization capacity. Ir-
radiance, temperature, and major points from I-V characteristics are the most
utilized input features in case of a fault in a PV array. In comparison, current
and voltage data are used in case of a fault in a battery, inverter, MPPT, and
others. Electrical and meteorological data are mostly used in ML, whereas image
data are the most common input features for deep learning algorithms such as
CNN. However, recently as 1-D can be transformed to 2-D data, electrical and
meteorological data are also employed for deep learning algorithms in general.
Faults like arc fault that does not reflect its effect on I-V characteristics of
PV arrays, a method that includes the analysis of signal waveform (some kind of
transformation, for example, wavelet) which could show signal distortion effect,
might be an appropriate method to capture most of the faults in a PV system.
Moreover, in most papers, prepossessing of data like normalization has resulted
in better accuracy. Nevertheless, whenever this is not possible deep learning
models are efficient due to their capacity in extracting features automatically.
Even if major progress has been seen in the research area in using the ML
method for FDD in a PV system, only one paper has implemented the ML
method in prototype based on the literature review. Furthermore, so far, this
method is not commercialized. Thus, the authors have identified the following
main challenges.
 Table 1: Summary of reviewed literature on PV system FDD using ML methods

Reference Year Methods Data Set Fault identified Comment

Type Algorithm Input Data Simulated Real system Components Faults Accuracy Type of PV system
[28] 2021 TL ResNet-50 CNN I, V 3 5 Inverter Norma, PS and SC >97% GCPVS
[22] 2021 EL DT, RF, DA ... String I, V, P, T, G 5 3 PV array Norma, OC and SC >97% GCPVS
[18] 2021 ML SVR, GPR T, G, Ppeak 3 3 PV array OC, SC, LL, PS, Degradation - GCPVS
[38] 2020 ML SVM - - - - Normal and Faulty - -
[17] 2020 ML KNN, RF, DA, NB, DT, SVM - - - PV array OC, PS, Sensor >96% GCPVS
Inverter IGBT, Grid connection
[23] 2020 ML KNN, DT, SVM, ANN(MLP) - 3 3 PV array OC, PS, SC, Degradation - GCPVS
[5] 2020 ML KNN, DT, SVM, ANN(MLP) - 3 3 PV array OC, PS, SC, Degradation - GCPVS
[12] 2020 ML SVM, GA From I-V char 3 5 PV array LL 97% -
[3] 2020 ML SVM Image (Infrared thermography) 5 3 PV array Hotspot >92% SAPVS
[37] 2020 ML Anomaly detection - - - Battery Internal SC, GF - SAPVS
[13] 2020 EL SVM, NB, KNN I-V char 3 3 PV array LL >99 % SAPVS
[4] 2020 TL AlexNetCNN T,G, I-V char, 3 3 PV array LL, OC, AF, PS with MPPT >70.45 % SAPVS
Boost converter Imax , Vmax , Pmax
ML SVM, RF, LSTM, Bi-LSTM
[10] 2019 ML KNN, SVM, ANN - - - PV array Module SC, OC, PS, Degradation 99.65%(ANN) -
MPPT MPPT fault
[35] 2018 ML MLP I, V 3 3 PV array SC, OC >96 % SAPVS
Battery External SC
[2] 2018 ML MLP DWT voltage data 3 5 PV array PS >99.1% -
DC-DC converter Switch SC
Inverter Switch SC, OC with LCL filter failure
and Gating circuit failure
[27] 2018 ML KNN T, G, Vm pp, Im pp, and Pm pp 3 3 PV array OC, LL, PS 98.7% GCPVS
[11] 2017 ML SVM SCADA data 5 3 PV array PS, Hotspot - -
[21] 2017 ML ANN - 3 5 Inverter SC incipient - GCPVS
[25] 2017 ML DT - - - PV array PS,By pass diode failure >95.3 % GCPVS
Inverter failure
[15] 2017 ML PNN, ANN T, G, Imppt , Vmppt 3 3 PV array Module SC, String OC - GCPVS
ML for PV System Operational Fault Analysis

[7] 2017 ML KELM I-V char 3 3 PV array SC, OC, PS, Degradation - -
[8] 2016 ML MLP, RBF I, V, peaks from I-V char 3 5 PV array SC,bypass diode, OC, connection, PS - -
[39] 2014 ML - - - - Battery Internal resistance fault, overcharging - -
11

Note: - : Not given, 3: Is used, 5: Not used G: Irradiance, T: Temperature and for other abbreviations please refer to the document
12 T. Zenebe et al.

– Training, validation, and test data set that fit at least major fault in a PV
system, PV type and size are very rare to find.
– Even if most researchers have developed their own data set, most of them
are simulation data. Besides, in DL-based methods, gathering the image data
using a camera and drone is very expensive.
– Many measuring devices and sensors are needed due to the absence of a
proper method for effective input feature selection.
– There is a lack of knowledge on how to generate rare but severe faults.
– Selection of model configurations is done with try and error.
– The model devised so far does not have the modularity and generalization
capacity; as a result, ML model selection varies depending on fault type, the
size, and type of input data.
– Studies that guide integrating the methods with the existing protective de-
vices are not developed very well. Moreover, all the paper does not go in-
depth on how to clear the faults. Once the fault is classified, a method and
strategy are needed to coordinate it with protective devices for clearing the
fault automatically and/or convey the message to the operators for solutions.
– The model’s accuracy is variable as it depends on the data size, data quality,
and the number of input and output features.
– For comparing ML methods based on accuracy, cost, execution time, memory
usage, there are no standards or common testing platforms.

4 Conclusion and Recommendation

We found that SVM and MLP are the most utilized ML methods in recent liter-
ature. In addition, only a few literature used ensemble and transfer learning. As
input features, electrical, meteorological, and image data have been used. Fur-
thermore, the majority of ML techniques have resulted in an accuracy of greater
than 90%. Besides, PV array faults such as SC, OC, and PS are the most investi-
gated faults in a PV system. Challenges related to a data set, model configuration
selection, and integration of the ML method with the existing PV system are
identified. For SAPVS, it can be concluded that there is a lack of a holistic
approach for critical faults in its components. Therefore, extensive research is
required to see the implementation of those methods, and less investigated algo-
rithms have to be studied. Moreover, for efficient and effective research, sharing
of training, validation, and testing data set shall be encouraged.

Acknowledgements
We would like to thank our Electrical Power Engineering department at NTNU
for funding this work as part of a PhD. project.

Conflict of interest
The authors declare that they have no conflict of interest.
 ML for PV System Operational Fault Analysis 13

References
1. Snapshot of Global PV Markets - 2020 p. 20
2. Ahmad, S., et al.: Fault Classification for Single Phase Photovoltaic Systems using Ma-
chine Learning Techniques. In: 2018 8th IEEE India International Conference on Power
Electronics (IICPE). pp. 1–6. https://doi.org/10.1109/IICPE.2018.8709463
3. Ali, M.U., et al.: A machine learning framework to identify the hotspot
in photovoltaic module using infrared thermography 208, 643–651.
https://doi.org/10.1016/j.solener.2020.08.027
4. Aziz, F., et al.: A Novel Convolutional Neural Network-Based Ap-
proach for Fault Classification in Photovoltaic Arrays 8, 41889–41904.
https://doi.org/10.1109/ACCESS.2020.2977116
5. Basnet, B., et al.: An Intelligent Fault Detection Model for Fault Detection in Photovoltaic
Systems 2020, e6960328. https://doi.org/10.1155/2020/6960328
6. Bognár, A., et al.: An unsupervised method for identifying local PV shad-
ing based on AC power and regional irradiance data 174, 1068–1077.
https://doi.org/10.1016/j.solener.2018.10.007
7. Chen, Z., et al.: Intelligent fault diagnosis of photovoltaic arrays based on optimized kernel
extreme learning machine and IV characteristics. Applied Energy 204, 912–931 (2017)
8. Chine, W., et al.: A novel fault diagnosis technique for photovoltaic systems based on
artificial neural networks. Renewable Energy 90, 501–512 (2016)
9. Cho, T.H., et al.: Comparison of Intelligent Methods of SOC Estimation for Battery of
Photovoltaic System 9(9), 8
10. Da Costa, C.H., et al.: A Comparison of Machine Learning-Based Methods for
Fault Classification in Photovoltaic Systems. In: 2019 IEEE PES Innovative Smart
Grid Technologies Conference - Latin America (ISGT Latin America). pp. 1–6.
https://doi.org/10.1109/ISGT-LA.2019.8895279
11. Dong, A., et al.: Fault Diagnosis and Classification in Photovoltaic Systems Using SCADA
Data. In: 2017 International Conference on Sensing, Diagnostics, Prognostics, and Control
(SDPC). pp. 117–122. https://doi.org/10.1109/SDPC.2017.31
12. Eskandari, A., et al.: Autonomous Monitoring of Line-to-Line Faults in Photovoltaic Sys-
tems by Feature Selection and Parameter Optimization of Support Vector Machine Using
Genetic Algorithms 10(16), 5527. https://doi.org/10.3390/app10165527
13. Eskandari, A., et al.: Line-line fault detection and classification for photovoltaic
systems using ensemble learning model based on I-V characteristics 211, 354–365.
https://doi.org/10.1016/j.solener.2020.09.071
14. Europe, S.P.: Global market outlook for solar power/2020–2024. Solar Power Europe:
Brussels, Belgium (2020)
15. Garoudja, E., et al.: An enhanced machine learning based approach for failures detection
and diagnosis of PV systems. Energy conversion and management 151, 496–513 (2017)
16. Ghaffarzadeh, N., Azadian, A.: A comprehensive review and performance evaluation in
solar (PV) systems fault classification and fault detection techniques
17. Hajji, M., et al.: Multivariate feature extraction based supervised ma-
chine learning for fault detection and diagnosis in photovoltaic systems .
https://doi.org/10.1016/j.ejcon.2020.03.004
18. Harrou, F., et al.: Monitoring of Photovoltaic Systems Using Improved Kernel-Based
Learning Schemes 11(3), 806–818. https://doi.org/10.1109/JPHOTOV.2021.3057169
19. Hong, J., et al.: Fault prognosis of battery system based on accurate voltage
abnormity prognosis using long short-term memory neural networks 251, 113381.
https://doi.org/10.1016/j.apenergy.2019.113381
20. IRENA: Renewable power generation costs in 2018. International Renewable Energy
Agency, Abu Dhabi (2019)
21. Ismail, N., et al.: Short-Circuit Incipient Faults Detection from Single Phase PWM Inverter
using Artificial Neural Network 10, 10
14 T. Zenebe et al.

22. Kapucu, C., et al.: A supervised ensemble learning method for fault diagnosis in photo-
voltaic strings 227, 120463. https://doi.org/10.1016/j.energy.2021.120463
23. Lazzaretti, A.E., et al.: A Monitoring System for Online Fault Detection and Classification
in Photovoltaic Plants 20(17), 4688. https://doi.org/10.3390/s20174688
24. Li, B., et al.: Application of Artificial Neural Networks to photovoltaic fault detection and
diagnosis: A review 138, 110512. https://doi.org/10.1016/j.rser.2020.110512
25. Livera, A., et al.: Advanced Failure Detection Algorithms and Performance Decision Clas-
sification for Grid-Connected PV Systems. https://doi.org/10.4229/EUPVSEC20172017-
6BV.2.13
26. Madeti, S.R., Singh, S.N.: A comprehensive study on different types of
faults and detection techniques for solar photovoltaic system 158, 161–185.
https://doi.org/10.1016/j.solener.2017.08.069
27. Madeti, S.R., Singh, S.N.: Modeling of PV system based on experimental data for fault
detection using kNN method 173, 139–151. https://doi.org/10.1016/j.solener.2018.07.038
28. Malik, A., et al.: Transfer Learning-Based Novel Fault Classification Technique for Grid-
Connected PV Inverter. In: Mekhilef, S., et al. (eds.) Innovations in Electrical and
Electronic Engineering. pp. 217–224. Lecture Notes in Electrical Engineering, Springer.
https://doi.org/10.1007/978-981-16-0749-3-16
29. Mano, M., et al.: Classification and detection of faults in grid connected photovoltaic
system. Int. J. Sci. Eng. Res 7(4), 149–154 (2016)
30. Mellit, A.: Recent Applications of Artificial Intelligence in Fault Diagnosis of Photovoltaic
Systems. In: Mellit, A., Benghanem, M. (eds.) A Practical Guide for Advanced Methods
in Solar Photovoltaic Systems, pp. 257–271. Advanced Structured Materials, Springer
International Publishing. https://doi.org/10.1007/978-3-030-43473-1-13
31. Mellit, A., et al.: Fault detection and diagnosis methods for photovoltaic systems: A review.
Renewable and Sustainable Energy Reviews 91, 1–17 (2018)
32. Nailen, R.L.: Battery protection-where do we stand? IEEE transactions on industry ap-
plications 27(4), 658–667 (1991)
33. Pillai, D.S., Rajasekar, N.: A comprehensive review on protection challenges and fault
diagnosis in PV systems. Renewable and Sustainable Energy Reviews 91, 18–40 (2018)
34. Sabbaghpur Arani, M., Hejazi, M.A.: The Comprehensive Study of Electri-
cal Faults in PV Arrays 2016, 1–10. https://doi.org/10.1155/2016/8712960,
https://www.hindawi.com/journals/jece/2016/8712960/
35. Sabri, N., et al.: Intelligent fault supervisory system applied on stand-alone photovoltaic
system. In: 2018 International Conference on Applied Smart Systems (ICASS). pp. 1–5.
https://doi.org/10.1109/ICASS.2018.8651950
36. Sabri, N., et al.: Faults diagnosis in stand-alone photovoltaic system using artificial neural
network. In: 2018 6th International Conference on Control Engineering & Information
Technology (CEIT). pp. 1–6. IEEE (2018)
37. Sabri, N., et al.: Battery internal fault monitoring based on anomaly detection algorithm.
Advanced Statistical Modeling, Forecasting, and Fault Detection in Renewable Energy
Systems p. 187 (2020)
38. Shin, J.H., Kim, J.O.: On-Line Diagnosis and Fault State Classification Method of Pho-
tovoltaic Plant 13(17), 4584. https://doi.org/10.3390/en13174584
39. Tadj, M., et al.: An innovative method based on satellite image analysis to check fault in a
PV system lead–acid battery 47, 236–247. https://doi.org/10.1016/j.simpat.2014.06.010
40. Youssef, A., et al.: The role of artificial intelligence in photo-voltaic systems design and
control: A review 78, 72–79. https://doi.org/10.1016/j.rser.2017.04.046
41. Zhao, Y., et al.: Line-line fault analysis and protection challenges in solar photovoltaic
arrays. IEEE transactions on Industrial Electronics 60(9), 3784–3795 (2012)