International Journal of Applied Power Engineering (IJAPE)

Vol. 11, No. 2, June 2022, pp. 134~144

ISSN: 2252-8792, DOI: 10.11591/ijape.v11.i2.pp134-144  134

A multi-scale dual-stage model for PV array fault detection,

classification, and monitoring technique

Siti Nor Azlina Mohd Ghazali, Muhamad Zahim Sujod

Department of Electrical Engineering, Faculty of Electrical and Electronics Engineering Technology, Universiti Malaysia Pahang,
Pahang, Malaysia

Article Info ABSTRACT

Article history: The output generated by photovoltaic arrays is influenced mainly by the
irradiance, which has non-uniform distribution in a day. This has resulted in
Received Apr 1, 2022 the current-limiting nature and nonlinear output characteristics, and
Revised May 1, 2022 conventional protection devices cannot detect and clean faults appropriately.
Accepted May 19, 2022 This paper proposes a low-cost model for a multi-scale dual-stage
photovoltaic fault detection, classification, and monitoring technique
developed through MATLAB/Simulink. The main contribution of this paper
Keywords: is that it can detect multiple common faults, be applied on multi-scale
photovoltaic arrays regardless of environmental conditions, and be beneficial
Discriminate analysis for photovoltaic system maintenance work. The experimental results show
Fault detection that the developed algorithm using supervised learning algorithms mutual
K-fold cross-validation with k-fold cross-validation has produced good performances in identifying
K-nearest neighbor six common faults of photovoltaic arrays, achieved 100% accuracy in fault
Random Forest detection, and achieved good accuracy in fault classification. Challenges and
Solar photovoltaic suggestions for future research direction are also suggested in this paper.
Support vector machine Overall, this study shall provide researchers and policymakers with a
valuable reference for developing photovoltaic system fault detection and
monitoring techniques for better feasibility, safety, and energy sustainability.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Siti Nor Azlina Mohd Ghazali
Department of Electrical Engineering, Faculty of Electrical and Electronics Engineering Technology
Universiti Malaysia Pahang
26600 Pekan, Pahang, Malaysia
Email: [email protected]

1. INTRODUCTION
Globally, power generation from solar photovoltaic (PV) systems is experiencing a significant
increase [1]. This increase has also led to risks associated with damage to PV system components, injury to
operators, and fire hazards to PV systems and buildings. Since PV output is nonlinear, conventional
protection devices (CPD) such as fuses and circuit breakers can detect faults and isolate faulty circuits only at
large fault currents and voltages. Therefore, better fault detection and monitoring techniques for PV systems
are needed [2] for better feasibility, safety, and energy sustainability. Recent studies have developed
advanced or intelligent fault detection and monitoring techniques for solar PV systems. The main ones are
model-based and IV power loss curve approaches, machine learning techniques, statistics-based techniques,
and output signal analysis techniques [3].
The model-based approach for detecting and identifying PV faults compares the expected data
obtained from the simulation process with data measured from an experiment or data collected from a PV
system [4]–[7]. This technique involves the least integration complexity with PV systems and requires low
implementation costs. However, most studies have found that the accuracy obtained is lower than other

Journal homepage: http://ijape.iaescore.com

Int J Appl Power Eng ISSN: 2252-8792  135

advanced PV fault detection methods. Machine learning (ML) techniques, on the other hand, have been the
most favorable method for detecting and diagnosing PV systems faults. This approach exploits artificial
intelligence with three main algorithms; supervised learning, semi-supervised learning, and unsupervised
learning in task completion [8]–[15]. Studies have proven this technique acquires high accuracy. Still, the
need for data acquisition systems and advanced computing system skills has made it complex and
challenging to integrate with PV systems and expensive to implement.
Meanwhile, statistical-based analysis mostly sets a threshold value and compares it with the actual
value measured in determining a PV system's normal or faulty state [13]–[16]. Earlier studies indicated that
approaches using mean differences or variances have better abilities in determining errors in the PV system.
Though, incorrectly setting the threshold limit can reduce method accurateness. Lastly, the output signal
analysis using the frequency-time domain to detect abnormalities in the sample in identifying faults in the PV
system has also attained high accuracies [17]–[20]. Nevertheless, it requires sophisticated tools to generate
the signal and making it the most expensive method.
Furthermore, most of the methods/techniques that have been developed in the previous study were
only to detect specific faults and did not provide fault location. Whereas finding the location of the fault is
always challenging and time-consuming for large-scale PV systems [21], [22]. Apart from developing
previous fault detection methods mostly only tested or evaluated on small-scale PV arrays/systems, they did
not examine the PV fault detection methods on the maintenance aspects. Studies have found a good
maintenance system is important for inspecting and performing corrective work because different incidents
or failures have different characteristics that require specialized competent people, different tools and
techniques to deal with and implement corrections [23].
Hence, in this paper, we developed the multi-scale dual-stage (MsDs) model for PV fault detection,
classification, and monitoring technique, which requires a low implementation cost, can detect multiple faults
with fault locations, and can be applied to all PV array scales, also useful for PV maintenance works. The
dual-stage algorithm comprises of fault detection algorithm at stage-1 and fault classification and location at
stage-2. The MsDs has employed supervised learning techniques of discriminate analysis (DA), k-nearest
neighbor (KNN), support vector machine (SVM), random forest (RF) in identifying the best algorithm which
produces the best accuracy.
The remaining part of this paper is organized as follows: i) Section 2 describes PV array modeling
and simulation processes; ii) After that, section 3 presents the proposed MsDs technique; iii) Then, section 4
provides the simulation and testing algorithm's results, discussion, and limitations; and iv) Finally, section 5
presents the conclusion and recommendation for future work direction.

2. PV ARRAY MODELING AND SIMULATION

2.1. Model and input data for solar cells
A one-diode model (ODM) is chosen in the study to develop PV array modeling because of its
advantages compared to the double-diode model. It has good accuracy for steady-state conditions and faults
analysis for PV systems. Further, ODM parameters for PV modules are available for most PV modules in the
market and are the most commonly chosen model by researchers [24]. The equivalent circuit for ODM and
its parameter is shown in Figure 1.

Figure 1. An equivalent circuit of a one-diode model with five parameters

By using Kirchhoff's law, the output current 𝐼 in (A) of the PV cell is formulated as given by (1),
(2), and (3), where the IL represents light-generated current, while the ID represents the diode current and 𝐼𝑠ℎ
represents the shunt resistance current.
A multi-scale dual-stage model for PV array fault detection, classification … (Siti Nor Azlina Mohd Ghazali)
136  ISSN: 2252-8792

𝐼 = 𝐼𝐿 − 𝐼𝐷 − 𝐼𝑠ℎ (1)

𝑉+𝐼𝑅𝑠
𝐼𝐷 = exp ( . 𝑞) − 1 (2)
𝐴𝑘𝑇

𝑉+𝐼𝑅𝑠
𝐼𝑠ℎ = (3)
𝑅𝑠ℎ

Where q is the electron charge (1.610-19 C), A is the diode ideal factor, T is the ambient temperature (K), V is
the solar cell voltage, and k is the boltzmann's constant (1.38 10-23 J/K), the polycrystalline silicon PV
modules model solartech energy ASC-6P-48-200 is taken for practical comparison. The value of input
parameters of open-circuit voltage (Voc), short-circuits current (Isc), series resistance (Rs), and shunt
resistance (Rsh) is obtained from the PV manufacturer's datasheet as in Table 1.

Table 1. Solartech energy ASC-6P-48-200 PV module parameters data

Parameter Symbol Value
Maximum Power Pmpp 199.998 W
Open Circuit Voltage Voc 30.12 V
Maximum Power Voltage Vmp 24.6 V
Short Circuit Current Isc 8.63 A
Maximum Power Current Impp 8.13 A
Light-generated current IL 8.6789 A
Diode saturation current Io 2.929 x 10-10 A
Diode ideality factor N 1.0136
Shunt resistance Rsh 210.82 Ω
Series resistance Rs 0.223 Ω
ISC Temperature coefficient α 0.06
VOC Temperature coefficient β -0.35999
Solar cell number in series n 48

2.2. Simulation procedure using MATLAB/Simulink

Modeling, simulation process, and development of PV array fault detection and classification
algorithm are by using MATLAB/Simulink. Using simulation data can produce a more precise algorithm. In
addition, it is useful for an unavoidable restriction in pandemic situations and the inevitable constraints with
the impossibility of external operational irradiation to obtain data from the actual working conditions of the
PV system. Six common faults or abnormal conditions of PV array, namely, degradation array (DF), open-
circuit fault (OCF), line-line fault (LLF), ground fault (GF), partial shading (PS) condition, and faulty
module (FM), were explored respectively in this study. A PV module consists of several solar cells with
identical parameters, as shown in Figure 2. Several PV modules/panels were then used to build PV arrays, a
modified version adopted in [8]. In this study, the small-scale PV array model was configured as five parallel
PV strings of six in a series (5*6) of PV modules, as presented in Figure 3.
The simulation process assumed PV array is the only source of the fault current, and there is no
overcurrent or overvoltage from external sources. A Simulink model of the I-V testing circuit configured was
to generate the I-V curves and simulated data (power, voltage, and current) from the PV array models, as
presented in Figure 4. This paper does not present PV array models for the six common fault simulations
individually to save space. The models are combined on one diagram for description, as shown in Figure 5.

Figure 2. MATLAB/Simulink of the one-diode model module with solar cell

Int J Appl Power Eng, Vol. 11, No. 2, June 2022: 134-144
Int J Appl Power Eng ISSN: 2252-8792  137

Figure 3. MATLAB/Simulink of 5*6 small-scale PV array model

Figure 4. I-V characteristics circuit in MATLAB/Simulink

Figure 5. Description of six PV array faults simulation model

These six PV array fault models were simulated and tested under standard test conditions (STC)
with radiation at 1000 W/m2 and a module temperature of 25 ºC. The simulation processes were carried out as:

A multi-scale dual-stage model for PV array fault detection, classification … (Siti Nor Azlina Mohd Ghazali)
138  ISSN: 2252-8792

i) Simulation of the LLF model was performed by short-circuiting two different potential points in the PV
array string. This simulation assumes that the fault impedance is zero, and the LLF at a large voltage
difference fault was considered.
ii) Simulation of the GF model was achieved by extending the LLF model by connecting to the ground to
create a fault current.
iii) Simulation of the PS model was carried out by setting PS Gain connected to PV modules to less than 1 to
reduce the irradiance value received by the module(s) to less than 1000 W/m2.
iv) Simulation of the OCF model was performed by adding an Rs to a PV array string, and Rs was set to
infinity.
v) Simulation of the FM model was accomplished by reversing the bypass diode of the solar cell.
vi) Simulation of the DF model was performed by adding Rs to the PV array and gradually increasing the
value of Rs.

2.3. PV array model validation

Figures 6(a) and 6(b) show the I-V and P-V curves generated from the simulation process under
STC for a small-scale (5*6) PV array model. The developed PV array model was validated by comparing the
simulation results of maximum power (Pmax), Voc, and Isc with the PV module manufacturer's datasheet
available in the market. This study chose solartech energy ASC-6P-48-200 PV module.

(a) (b)

Figure 6. Simulation process of (a) I-V curve of PV array model and (b) P-V curve of PV array model

It can be seen that the simulation results are closely matched with the datasheet, as shown in
Table 2. Therefore, this can be concluded that the proposed PV array model is accurate enough to predict the
performance of the PV array under normal and fault conditions in this study.

Table 2. Comparison of simulation results (small-scale PV model) with actual PV module datasheet
Parameters Solartech Energy ASC-6P-48-200 Simulated Data
Value of one module Total of 5*6 PV array model Value of one module Total of 5*6 PV array model
Pmax 199.988 W 5999.94 W 200 W 6000 W
Voc 30.12 V 180.72 V 30.12 V 180.72 V
Isc 8.63 A 43.15 A 8.63 A 43.15 A

3. MULTI-SCALE DUAL-STAGE FAULT DETECTION AND CLASSIFICATION ALGORITHM

3.1. Medium-scale and big-scale PV array model: modeling and simulation
Medium-scale and big-scale PV array models were constructed through MATLAB/Simulink by
adding panels in series and parallel strings, as shown in Figures 7 and 8. The data for input parameters; Voc,
Isc, Rs, and Rsh for the medium-scale and big-scale PV array models were also from the PV manufacturer's
datasheet, model solartech energy ASC-6P-48-200, as listed in Table 1. While, the simulation results of the
medium-scale and big-scale PV array models under STC is presented in Table 3.

Int J Appl Power Eng, Vol. 11, No. 2, June 2022: 134-144
Int J Appl Power Eng ISSN: 2252-8792  139

Figure 7. Model of medium-scale (10*30) PV array Figure 8. Model of big-scale (20*30) PV array

Table 3. Comparison of simulation results (medium-scale and big-scale PV model) with actual
PV module datasheet
Parameters Solartech Energy ASC-6P-48-200 Simulated Data of PV array model
Value of one Total of (10*30) Total of (20*30) Medium-scale Big-scale
module module module (10*30) (20*30)
Pmax 199.99 W 59.99 kW 119.99 kW 60 kW 120 kW
Voc 30.12 V 903.7 V 903.7 V 903.7 V 903.7 V
Isc 8.63 A 86.3 A 172.6 A 86.3 A 172.6 A

It can be seen that the value of Pmax, Voc, and Isc generated are almost the same as the datasheet of
solartech energy model ASC-6P-48-200, listed in Table 1. Hence, it can be concluded that the proposed
medium-scale and large-scale PV array models are precise enough to predict their performance under normal
and fault conditions in this study. The rest of simulation processes for medium and large-scale PV array fault
models (LLF, GF, PS, OCF, FM, and DF) were carried out with the same procedure as the small-scale PV
array model.

3.2. Fault detection and classification algorithm procedures

Figures 9(a) and 9(b) show the flowchart of the multi-scale dual-stage (MsDs) PV fault detection,
classification, and monitoring technique procedures. The MsDs procedure consists of stage-1 and stage-2. A
(𝑃𝑉_𝑛𝑜𝑓𝑎𝑢𝑙𝑡) represents the PV array no-fault model, and (𝑃𝑉_𝑓𝑎𝑢𝑙𝑡 𝑛 ) represents the PV array fault
models of DF, FM, GF, LLF, OCF, and PS. The flowchart of stage-1 describes the fault detection algorithm.
Due to the non-uniform PV output characteristics, a simple PV fault detection algorithm has been developed
to compare power, voltage, and current generated from the PV array fault-free model, which is higher than
the PV array fault model [25].
The parameter chosen for the testing of the detection algorithm is a difference in open-circuit
voltage (RVoc), a standard deviation of output power (stdP), and mean output voltage and current (µV & µI).
The fault detection algorithm was tested using four different supervised learning algorithms of DA, RF,
KNN, and SVM through MATLAB Simulink to acquire the best detection accuracy. Then, the testing
procedure of this fault detection algorithm was repeated and evaluated on medium and big-scale PV array
models built in this study to validate its practicality as a multi-scale PV array fault detection algorithm.

A multi-scale dual-stage model for PV array fault detection, classification … (Siti Nor Azlina Mohd Ghazali)
140  ISSN: 2252-8792

Data acquisition

Models
Modelling of PV array using MATLAB simulink developed:
PV module parameter: Isc, Voc, Iph, Rs, Rsh, N small,
medium and
big scale PV
PV no fault model PV fault model (PV_fault): array models
(PV_nofault) PS, GF, LLF, OCF, DF, FM
Simulation
and training
Simulation process to Simulation process to of algorithm
generate data generate data (simulation is are repeated
(G=1 kW/m2, T= 25˚ ) repeated for each fault model) for medium
and big
scales of PV
array models
Fault detection (stage-1):
Feature vector/input parameters: RVoc, µV, µI & stdP

YES: NO: Fault

No fault detected
Testing using four supervised ML algorithms:
DA, KNN, RF & SVM

Normal Fault classification

condition & location (stage-2):

(a)

Fault detection (stage-1):

YES: NO: Fault
No fault detected

Fault classification & location (stage-2):

Feature vectors/14 input parameters:
Pmax, Isc, Voc, Rs, RVoc, μV, μI, rmsV, rmsI varV, varI, stdP, stdV, stdI,

Evaluates using four supervised ML algorithms

(DA, SVM, RF, KNN)

Line-line
Normal Partial shading Open-circuit Ground fault Faulty Degradation
fault
condition (module) fault (string) (string) (module) (array)
(string)

(b)

Figure 9. The flowchart of the (a) MsDs fault detection algorithm (stage-1) and
(b) the MsDs fault classification and location algorithm (stage-2)

The flow chart for the stage-2 describes the testing algorithm procedure for the classification and
location of faults. The stage-2 procedure can proceed if any faults are detected at stage-1. In this study, the
algorithm of classification and location was tested and evaluated with the following processes:
i) The testing algorithm involved 15 data sets for 𝑃𝑉_𝑛𝑜𝑓𝑎𝑢𝑙𝑡 model and 15 data sets for each GF, LLF,
GA, OCF, DF, and PS 𝑃𝑉_𝑓𝑎𝑢𝑙𝑡 𝑛 models.
ii) Fourteen feature vectors/input parameters of Pmax, Isc, Voc, Rs, RVoc, µV, µI, root mean square voltage
and current (rmsV, rmsI), variance voltage and current (varV, varI), and standard deviations of power,
voltage and current (stdP, stdV, stdI) were selected for the testing algorithm because they have been
proven to produce good accuracy for PV system/array fault detection and classification [21], [26], [27].
iii) The testing algorithm used four ML algorithms, DA, RF, KNN, and SVM, to obtain the best algorithm
and produce the best classification accuracy.

Int J Appl Power Eng, Vol. 11, No. 2, June 2022: 134-144
Int J Appl Power Eng ISSN: 2252-8792  141

iv) The K-fold cross-validation method was adopted in the testing algorithm to optimize the parameter
chosen and improve the classification accuracy.
v) The testing algorithm procedure at stage-2 was repeated and evaluated on medium-scale and large-scale
PV array models to establish its feasibility as a multi-scale PV array fault classification and monitoring
algorithm.

4. RESULT AND DISCUSSION

4.1. Simulation results and analysis of fault detection algorithm
Figure 10 shows the I-V curves generated from the simulation process using MATLAB Simulink for
the small-scale PV array models that illustrates the relationship between the output voltage and the output
current yielded. From the figure, it can be observed that the I-V curves generated from the simulation of
𝑃𝑉_𝑛𝑜𝑓𝑎𝑢𝑙𝑡 model, and six 𝑃𝑉_𝑓𝑎𝑢𝑙𝑡 𝑛 models, having different characteristics as:
i) The ISC and the VOC remain unchanged for PS simulation while the Pmax decreases.
ii) For OCF simulations, the ISC and Pmax value decreases while the VOC remains unchanged.
iii) For the FM simulation, the ISC remains unchanged, while the VOC and the gradient of the end part of the I-
V curve also decrease.
iv) For the DF array simulation, the VOC value remains unchanged. But the ISC experienced a slight decrease,
and the overall slope of the I-V curve decreased.
v) For GF simulations, the VOC value increases, while other characteristics of I-V curves remain unchanged.
vi) For LLF simulations, the VOC decreases while the ISC remains unchanged, there is no significant change
in the remaining characteristics of I-V curves.

Figure 10. I-V curves of the fault and no-fault models small-scale PV arrays

Table 4 shows the testing results for the fault detection algorithm accuracies of the proposed small-
scale, medium-scale, and big-scale PV array models. It can be seen that the fault detection method based on
the RF algorithm has successfully acquired 100% accuracy for all PV array models. Other algorithms also
achieved good accuracies, with more than 96%.

Table 4. Fault detection accuracies (small, medium & big scales PV model) using four ML algorithms
Algorithm type Fault detection accuracy (%)
Small-scale PV model Medium-scale PV model Big-scale PV model
Discrimination Analysis 99 99 98
Random Forest 100 100 100
K-nearest neighbours 96 99 98
Support Vector Machine 97 99 100

4.2. Simulation results and analysis of fault classification algorithm

Table 5 presents the classification accuracy for small-scale, medium-scale, and large-scale PV array
models developed in this study. It can be seen that the classification accuracy based on the RF algorithm is
the highest compared to other algorithms, with an accuracy of more than 90% for medium and large scales
PV models and almost 80% for small-scale PV models. Meanwhile, the accuracy of the testing RF algorithm
A multi-scale dual-stage model for PV array fault detection, classification … (Siti Nor Azlina Mohd Ghazali)
142  ISSN: 2252-8792

for fault classification and location (module/string/array) for the six faults; DF, FM, GF, LLF, OCF, and PS
of the small-scale, medium-scale, and big-scale PV array models, can be seen in Table 6.

Table 5. Fault classification accuracies (small, medium & big scales PV model) using four ML algorithms
Algorithm type Fault classification accuracy (%)
Small-scale PV model Medium-scale PV model Big-scale PV model
Discrimination Analysis 82 90 86
Random Forest 78 90 93.3
K-nearest neighbours 48 53 55
Support Vector Machine 70 89 71

Table 6. Fault classification accuracies of six faults for small, medium, and big-scale PV model using RF algorithm
Fault Fault classification and location accuracy (%)
Small-scale PV model Medium-scale PV model Big-scale PV model
DF (array) 93.3 93.3 86.7
FM (module) 93.3 93.3 80.0
GF (string) 50.4 100 100
LLF (string) 53.3 100 93.3
OCF (string) 86.7 80.0 93.3
PS (array) 93.3 80.0 100

It can be seen that the fault classification method based on the RF algorithm has succeeded in
achieving high accuracy. Almost all fault types for all PV model scales achieve more than 90% classification
accuracy, and for DF and FM (large-scale), OCF and PS (medium-scale), and OCF for small-scale get more than
80% classification accuracy. Only the fault classification for GF and LLF (small scale) achieved low accuracy.

4.3. Discussion
This study has developed and simulated MsDs algorithms for PV array fault detection,
classification, and location via MATLAB/Simulink, which consists of stage-1 (fault detection algorithm), and
stage-2 (fault classification and location algorithm). Although the I–V curves generated from the simulation
process of the PV no-fault (𝑃𝑉_𝑛𝑜𝑓𝑎𝑢𝑙𝑡) model, and the six PV fault (𝑃𝑉_𝑓𝑎𝑢𝑙𝑡 𝑛 ) models have shared the
same characteristics of Voc, Isc, and Pmax (Figure 10). High accuracies were accomplished when the
developed fault detection algorithm was tested using four different supervised learning algorithms; DA, RF,
KNN, and SVM. The RF algorithm has achieved 100% accuracy for all scales of PV array models, as can be
seen in Table 4.
For fault classification at stage-2, the RF algorithm has again achieved high accuracy for medium-
scale and large-scale PV array models (more than 90%) compared to other algorithms, as shown in Table 5.
Only for small-scale PV array has produced modest accuracy. However, if we look at the accuracies of fault
classification and location for each 𝑃𝑉_𝑓𝑎𝑢𝑙𝑡 𝑛 model as presented in Table 6, the RF algorithm with the
combination of k-fold cross validation has delivered high accuracy for almost all PV array fault models
(more than 90%), except for GF and LLF on the small-scale PV array achieved the low accuracy values. This
might be due to the low discrimination power of ML algorithms in describing the faults, thus resulting in
poor performance. In summary, the proposed MsDs has the following research contributions over earlier
works [8], [28]:
i) It is low-cost and inexpensive modeling. The fault detection algorithm with the k-fold cross validation at
stage-1 has proven to detect multiple common faults; GF, LLF, PS, OCF, DF, and PS in PV arrays with
good accuracy and without interruption to the system.
ii) The classification and location algorithm with the k-fold cross validation at stage-2 can identify faults at
different locations; at the string, module, or array level, useful for large-scale PV systems/plants, and
achieved good accuracies.
iii) The study has proven that the developed algorithms are easy to execute and feasible to apply to all PV
array scales globally regardless of environmental conditions.
iv) A simple fault detection algorithm at level-1 is beneficial for preventive and predictive maintenance in
finding hidden faults in PV systems that CPD cannot detect. The hidden faults can reduce the system's
efficiency and cause worse circumstances such as fire hazards, injuries, and electric shocks to the PV
system operator.

Int J Appl Power Eng, Vol. 11, No. 2, June 2022: 134-144
Int J Appl Power Eng ISSN: 2252-8792  143

4.4. Limitations
This study has some limitations; the MsDs technique has been tested using simulated data only.
Accuracy may vary when this proposed algorithm is implemented on an actual operating PV system. Other
than that, the MsDs algorithm was tested using supervised learning algorithms, in which fully labeled data
were used. But for MsDs to be used for maintenance work with unlabeled data, the accuracy of the proposed
algorithm is not verified. Lastly, the algorithm for fault classification and location tested on small-scale PV
array models has shown moderate accuracies for GF (string) and LLF (string).

5. CONCLUSION AND FUTURE WORK DIRECTION

This study proposed a multi-scale dual-stage (MsDs) model for PV array fault detection,
classification, and monitoring technique that have demonstrated good accuracy. The MsDs consists of the PV
array fault detection algorithm at stage-1 and the PV array fault classification and location algorithm at
stage-2. The MsDs algorithms have been tested using four supervised learning algorithms; Discriminate
analysis (DA), K-nearest neighbor (KNN), Support vector machine (SVM), and Random Forest (RF),
together with k-fold cross-validation in finding the best algorithm that delivers the best accuracy. Further,
MsDs have also been evaluated on small, medium, and large-scale PV array models to ascertain their
feasibility on multi-scale PV arrays.
The simulation results have proved that the RF algorithm has accomplished the best accuracy for
both medium-scale and big-scale PV array models, with 100% accuracy for various faults (open-circuit fault,
degradation array, partial shading, faulty module, ground fault, and line-line fault) detection, and more than
90% accuracy for fault classification and location, excluding for model of a ground fault and line-line fault of
a small-scale PV array that produced low classification accuracy values. Overall, the simulation results have
justified the study's objectives to develop a low-cost model for PV arrays with various fault detection and
classification algorithms that can be implemented at various PV array scales and applicable for PV
maintenance works for better efficiency, reliability, and security of the PV system.
Nevertheless, some recommendations can be carried out for future work. It is recommended to
validate the proposed MsDs technique by testing developed algorithms using data from the real PV system.
This ensures the accuracy of the results obtained from the developed PV array model and the actual PV
system. Furthermore, the MsDs testing algorithm in this study applied supervised learning algorithms, in
which fully labeled data was used. Thus, the testing algorithms need to be evaluated on unlabeled data to
obtain more precise accuracy and verify the feasibility of MsDs for PV system maintenance work. Finally,
more training and testing need to be done on the classification and location algorithm for the small-scale PV
array models' line-line and ground faults to improve accuracy.

ACKNOWLEDGMENT
The author would like to thank the financial support of Universiti Malaysia Pahang (UMP) for the
financial support received under project number PGRS2003189.

REFERENCES
[1] BP, “Statistical Review of World Energy,” 2020 | 69th Edition, 2020.
[2] D. S. Pillai and N. Rajasekar, “A comprehensive review on protection challenges and fault diagnosis in PV systems,” Renewable
and Sustainable Energy Reviews, vol. 91, pp. 18–40, Aug. 2018, doi: 10.1016/j.rser.2018.03.082.
[3] A. Livera, M. Theristis, G. Makrides, and G. E. Georghiou, “Recent advances in failure diagnosis techniques based on
performance data analysis for grid-connected photovoltaic systems,” Renewable Energy, vol. 133, pp. 126–143, Apr. 2019, doi:
10.1016/j.renene.2018.09.101.
[4] M. H. Ali, A. Rabhi, A. El Hajjaji, and G. M. Tina, “Real Time Fault Detection in Photovoltaic Systems,” Energy Procedia, vol.
111, pp. 914–923, Mar. 2017, doi: 10.1016/j.egypro.2017.03.254.
[5] S. Fadhel et al., “PV shading fault detection and classification based on I-V curve using principal component analysis:
Application to isolated PV system,” Solar Energy, vol. 179, pp. 1–10, Feb. 2019, doi: 10.1016/j.solener.2018.12.048.
[6] A. Hazra, S. Das, and M. Basu, “An efficient fault diagnosis method for PV systems following string current,” Journal of Cleaner
Production, vol. 154, pp. 220–232, Jun. 2017, doi: 10.1016/j.jclepro.2017.03.214.
[7] M. Dhimish, V. Holmes, B. Mehrdadi, and M. Dales, “Diagnostic method for photovoltaic systems based on six layer detection
algorithm,” Electric Power Systems Research, vol. 151, pp. 26–39, Oct. 2017, doi: 10.1016/j.epsr.2017.05.024.
[8] Z. Chen, L. Wu, S. Cheng, P. Lin, Y. Wu, and W. Lin, “Intelligent fault diagnosis of photovoltaic arrays based on optimized
kernel extreme learning machine and I-V characteristics,” Applied Energy, vol. 204, pp. 912–931, Oct. 2017, doi:
10.1016/j.apenergy.2017.05.034.
[9] Z. Yi and A. H. Etemadi, “Line-to-Line Fault Detection for Photovoltaic Arrays Based on Multiresolution Signal Decomposition
and Two-Stage Support Vector Machine,” IEEE Transactions on Industrial Electronics, vol. 64, no. 11, pp. 8546–8556, Nov.
2017, doi: 10.1109/TIE.2017.2703681.
[10] Z. Chen et al., “Random forest based intelligent fault diagnosis for PV arrays using array voltage and string currents,” Energy
Conversion and Management, vol. 178, pp. 250–264, Dec. 2018, doi: 10.1016/j.enconman.2018.10.040.

A multi-scale dual-stage model for PV array fault detection, classification … (Siti Nor Azlina Mohd Ghazali)
144  ISSN: 2252-8792

[11] S. R. Madeti and S. N. Singh, “Modeling of PV system based on experimental data for fault detection using kNN method,” Solar
Energy, vol. 173, pp. 139–151, Oct. 2018, doi: 10.1016/j.solener.2018.07.038.
[12] Y. Zhao, R. Ball, J. Mosesian, J.-F. de Palma, and B. Lehman, “Graph-Based Semi-supervised Learning for Fault Detection and
Classification in Solar Photovoltaic Arrays,” IEEE Transactions on Power Electronics, vol. 30, no. 5, pp. 2848–2858, May 2015,
doi: 10.1109/TPEL.2014.2364203.
[13] M. Dhimish, V. Holmes, B. Mehrdadi, and M. Dales, “Comparing Mamdani Sugeno fuzzy logic and RBF ANN network for PV
fault detection,” Renewable Energy, vol. 117, pp. 257–274, Mar. 2018, doi: 10.1016/j.renene.2017.10.066.
[14] X. Lu et al., “Fault diagnosis for photovoltaic array based on convolutional neural network and electrical time series graph,”
Energy Conversion and Management, vol. 196, pp. 950–965, Sep. 2019, doi: 10.1016/j.enconman.2019.06.062.
[15] A. Haque, K. V. S. Bharath, M. A. Khan, I. Khan, and Z. A. Jaffery, “Fault diagnosis of Photovoltaic Modules,” Energy Science
& Engineering, vol. 7, no. 3, pp. 622–644, Jun. 2019, doi: 10.1002/ese3.255.
[16] E. Garoudja, F. Harrou, Y. Sun, K. Kara, A. Chouder, and S. Silvestre, “Statistical fault detection in photovoltaic systems,” Solar
Energy, vol. 150, pp. 485–499, Jul. 2017, doi: 10.1016/j.solener.2017.04.043.
[17] Z. L. X. Qing, X. Rong, J. Shengchang, Z. Lingyu, L. Yuan, “DC arc detection method based on electromagnetic radiation
characteristics,” High Voltage Technology, vol. 43, no. 9, pp. 2967–2975, 2017, doi: 10.13336/j.1003-6520.hve.20170831026.
[18] C. He, L. Mu, and Y. Wang, “The Detection of Parallel Arc Fault in Photovoltaic Systems Based on a Mixed Criterion,” IEEE
Journal of Photovoltaics, vol. 7, no. 6, pp. 1717–1724, Nov. 2017, doi: 10.1109/JPHOTOV.2017.2742143.
[19] B. P. Kumar, G. S. Ilango, M. J. B. Reddy, and N. Chilakapati, “Online Fault Detection and Diagnosis in Photovoltaic Systems
Using Wavelet Packets,” IEEE Journal of Photovoltaics, vol. 8, no. 1, pp. 257–265, Jan. 2018, doi:
10.1109/JPHOTOV.2017.2770159.
[20] N. K. Tumkur Jayakumar, M. U. Saleh, E. Benoit, J. Lacombe, M. Scarpulla, and C. Furse, “Fault Detection In PV Strings Using
SSTDR,” in 2018 USNC-URSI Radio Science Meeting (Joint with AP-S Symposium), Jul. 2018, pp. 9–10, doi: 10.1109/USNC-
URSI.2018.8602847.
[21] S. R. Madeti and S. N. Singh, “A comprehensive study on different types of faults and detection techniques for solar photovoltaic
system,” Solar Energy, vol. 158, pp. 161–185, Dec. 2017, doi: 10.1016/j.solener.2017.08.069.
[22] V. Carletti, A. Greco, A. Saggese, and M. Vento, “An intelligent flying system for automatic detection of faults in photovoltaic
plants,” Journal of Ambient Intelligence and Humanized Computing, vol. 11, no. 5, pp. 2027–2040, May 2020, doi:
10.1007/s12652-019-01212-6.
[23] S. N. A. M. Ghazali, M. Z. Sujod, and M. S. Jadin, “Forensic of Solar PV: A Review of Potential Faults and Maintenance
Strategies,” in 2021 International Conference on Engineering and Emerging Technologies (ICEET), Oct. 2021, pp. 1–6, doi:
10.1109/ICEET53442.2021.9659624.
[24] V. J. Chin, Z. Salam, and K. Ishaque, “Cell modelling and model parameters estimation techniques for photovoltaic simulator
application: A review,” Applied Energy, vol. 154, pp. 500–519, Sep. 2015, doi: 10.1016/j.apenergy.2015.05.035.
[25] B. Goss, I. R. Cole, E. Koubli, D. Palmer, T. R. Betts, and R. Gottschalg, “Modelling and prediction of PV module energy yield,”
in The Performance of Photovoltaic (PV) Systems, Elsevier, 2017, pp. 103–132.
[26] J. Zhang et al., “A reinforcement learning based approach for on-line adaptive parameter extraction of photovoltaic array
models,” Energy Conversion and Management, vol. 214, p. 112875, Jun. 2020, doi: 10.1016/j.enconman.2020.112875.
[27] P. K. Ray, A. Mohanty, B. K. Panigrahi, and P. K. Rout, “Modified wavelet transform based fault analysis in a solar photovoltaic
system,” Optik, vol. 168, pp. 754–763, Sep. 2018, doi: 10.1016/j.ijleo.2018.03.131.
[28] Y. E. Yağan, K. Vardar, and M. A. Ebeoğlu, “Modeling and Simulation of PV Systems,” IOSR Journal of Electrical and
Electronics Engineering (IOSR-JEEE), vol. 13, no. 2, pp. 1–11, 2018, doi: 10.9790/1676-1302030111.

BIOGRAPHIES OF AUTHORS

Siti Nor Azlina Mohd Ghazali was born in Kelantan, Malaysia, in 1978. She
received her B.Eng. degree in electrical engineering from the University of Mara Technology,
Malaysia, in 2002. Then she received her MSc in Energy Studies from the University of
Otago, New Zealand, in 2013. She is a member of the Board of Engineer Malaysia (BEM)
since 2004 and has been appointed as Professional Engineer in April 2018. Then, she has been
appointed as ASEAN Charted Professional Engineer in March 2019. She is currently working
towards her Ph.D. at the College of Engineering, Department of Electrical and Engineering,
Universiti Malaysia Pahang, Malaysia. Her current research interests include PV forensic
electrical, PV smart maintenance strategies, and PV smart fault monitoring system. She can be
contacted at email: [email protected].

Muhamad Zahim Bin Sujod was born in Selangor, Malaysia in 1976. He

received the B.Eng. degree and M.Eng. degree in Electrical & Electronics Engineering from
the University of Ehime, Ehime, Japan, in 2000 and 2002, respectively, and the Ph.D. degree
from the University Duisburg-Essen, Germany, in Power System Engineering, in 2014. He is
a member of the Board of Engineer Malaysia (BEM) since January 2004 and has been
appointed as Professional Engineer in January 2009. Currently, he is an Associate Professor at
the College of Engineering, Universiti Malaysia Pahang, Malaysia. His primary research
activities involve renewable energy systems (wind turbine and photovoltaic), energy
conversion, energy management, and electrical machines. He can be contacted at email:
[email protected].

Int J Appl Power Eng, Vol. 11, No. 2, June 2022: 134-144