Received 26 May 2024; revised 20 September 2024; accepted 5 November 2024.

Date of publication 13 November 2024; date of current version 22 November 2024.

Digital Object Identifier 10.1109/OAJPE.2024.3497880

A Robust Multi-Modal Deep Learning-Based

Fault Diagnosis Method for PV Systems
SHAHABODIN AFRASIABI 1 (Member, IEEE), SARAH ALLAHMORADI1 (Member, IEEE),
MOUSA AFRASIABI2 (Member, IEEE), XIAODONG LIANG 1 (Senior Member, IEEE),
C. Y. CHUNG 3 (Fellow, IEEE), AND JAMSHID AGHAEI 4 (Senior Member, IEEE)
1 Department of Electrical and Computer Engineering, University of Saskatchewan, Saskatoon, SK S7N 5A9, Canada
2 Cyient, Vaasa, 65101 Ostrobothnia, Finland
3 Department of Electrical Engineering, The Hong Kong Polytechnic University, Hong Kong, China
4 School of Engineering and Technology, Central Queensland University, Rockhampton, QLD 4701, Australia

CORRESPONDING AUTHOR: X. LIANG ([email protected])

This work was supported in part by the University of Saskatchewan, Saskatoon, Canada.

ABSTRACT In this paper, a robust, multi-modal deep-learning-based fault identification method is pro-
posed for solar photovoltaic (PV) systems, capable of detecting a wide range of faults at PV arrays, inverters,
sensors, and grid connections. The proposed method combines residual convolutional neural networks
(CNNs) and gated recurrent units (GRUs) to effectively extract both spatial and temporal features from raw
PV data. To enhance the proposed model’s robustness and accuracy, a probabilistic loss function based on the
entropy theory is formulated. The proposed method is validated using both experimental data obtained from a
PV emulator-based test system and simulation data, achieving over 98% accuracy in fault identification under
various noise conditions. The results indicate that the proposed model outperforms conventional CNN- and
MSVM-based methods, demonstrating its potential in providing precise fault diagnostics in PV systems.

INDEX TERMS Convolutional neural networks (CNNs), fault identification, feature extraction, gated neural
networks (GNNs), information theory, loss function, multi-modal deep neural network, photovoltaics.

I. INTRODUCTION - The DC side fault detection is essential for the safe

W ITH recent advances on solar photovoltaic (PV)

technologies, the worldwide PV capacity grew by
approximately 395.3 GW from 2007 to 2017 [1]. PV systems
operation of a PV system [5], but DC faults often fail to
be detected by the conventional protection systems [6],
which may be escalated into fire hazards [7]. IEC 62109-
generate clean energy, can operate in standalone or grid- 1 and IEEE Std. 1547 provide guidelines for PV systems
connected mode, and work well individually or with other safety, but DC side faults are not covered [8]. Therefore,
energy sources. The PV system protection plays an essential developing advanced DC side fault detection methods
role to ensure secure and reliable operations [2]. However, for PV systems is urgently needed [9].
several issues may cause malfunction of the PV protection To overcome these challenges, a comprehensive frame-
system, including work for the PV system fault detection is developed in this
- Faults have a low current magnitude in a PV system. paper.
- The maximum power point tracking (MPPT) system can PV fault identification methods in the literature can be
reduce the fault current magnitude and significantly alter categorized into: 1) model-based methods, 2) measurement
fault characteristics. analysis-based methods, 3) infrared (IR) thermography cam-
- PV outputs are nonlinear and depend on varying factors, era and unmanned aerial vehicle (UAV)-based methods,
such as the meteorological data. 4) data-driven techniques, and 5) hybrid methods. The char-
- Noises may seriously affect fault detection in PV sys- acteristics, benefits, and limitations of these methods are
tems, but noises are either neglected or only Gaussian shown in Table 1.
noises are considered in previous studies [3], while most Model-Based Methods compare measurements of a real PV
noises do not follow the Gaussian distribution function system with the model-generated values, and discrepancies
in power systems [4]. exceeding a pre-set threshold indicate potential faults. The
2024 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 11, 2024 583
TABLE 1. Definitions, Advantages and Disadvantages of fault detection methods for PV systems in the literature.

accuracy of model-based methods are affected by the model, system fault diagnosis [13], [19], [20], but cannot fully learn
thresholds, and noises [4]. An improved Kalman filter and the abnormal and normal patterns based on time-varying signals
Thevenin equivalent resistance are examples for fault identi- with different locational features across a PV system [21].
fication methods in PV systems [4], including short-circuit Two potential solutions would be: 1) extract and select sup-
faults, open-circuit faults [5], and partial shading faults. plementary features, and 2) replace classic machine learning
Measurement Analysis-Based Methods, where abnormal by advanced deep learning.
conditions are identified by analyzing measurement param- Hybrid Methods, where two or more techniques are com-
eters in the time or frequency domain. The ground fault bined to enhance the effectiveness of the PV system fault
detection in PV arrays using the time-domain analysis is identification. For example, in [22], the principal component
developed in [10]; the arc occurrence in PV arrays is detected analysis (PCA) is used with random forest to identify open-
using the improved empirical wavelet transform in [6]; the circuit faults, islanding, and partial shading of PV systems.
wavelet transformation is used to detect PV faults in [11]. In [23] and [24], a sequence of voltages and currents is
These methods do not rely on the PV system’s mathematical combined as feature extraction and fed into a ladder neural
model, but rely on the predefined thresholds for fault detec- network and t-distributed stochastic neighbor embedding (t-
tion, and are very sensitive to noises. SNE) with the learning vector quantization (LVQ) to detect
IR and UAV-Based Methods, where abnormalities faults considering Gaussian noises. The combination of the
are detected through continuous temperature monitoring, sequential forward search, the sequential backward search,
as faults in PV systems can lead to thermal imbalances and support vector machine (SVM), naïve Bayes, and the logistic
temperature variations. An infrared thermographic camera regression is used in [25] to identify PV array faults. The
can be used to identify faults in PV systems by remotely infrared thermography and SVM are combined for the PV
monitoring their thermal conditions (no direct electrical mea- system’s fault identification in [26]. Machine learning is fur-
surements). The ther employed alongside IR-based techniques to detect PV
UAVs, such as drones equipped with various sensors, can array faults in [25]. While the hybrid methods may produce
also be used to remotely detect faults in PV systems. In [17], accurate results in some cases, their performance can be
the infrared cameras are employed to identify faulty PV significantly affected by noises, leading to the increase of
modules. In [12], different UAV-based cameras are used to information redundancy [27].
identify PV array faults. The UAV-, IR- and sensor-based data Deep neural networks realize features directly from mea-
are used in [14] to identify faults in large-scale PV power surement data, and can be broadly classified into four
plants. Although precise in the fault detection, IR and UAV- types [28]: deep encoders (DA), deep Boltzmann machines
based methods can be costly for large PV power plants. The (DBMs), recurrent neural networks (RNNs), and convolu-
infrared thermography and UAVs can only be used to monitor tional neural networks (CNN). PV array’s fault detection
changes in PV arrays, but don’t provide a comprehensive fault methods based on DA and DBMs are proposed in [29]
diagnosis system or cover the grid-connected PV infrastruc- and [30], respectively, but they are not effective for time-
ture, such as controllers, switches, etc. varying features. RNN-based networks, particularly gated
Data-Driven Techniques, where PV system faults can be RNNs, such as long short-term memory (LSTM) and gated
detected through machine learning based on historical data recurrent neural networks (GRNN), are good at time-varying
gathered by measurements and sensor readings [18]. Machine features, and have been applied in the PV system’s fault
learning methods include classic machine learning and deep detection [31], [32], but GRNNs cannot fully capture fea-
learning. Classic machine learning are often used in the PV tures from different components of a PV system [33]. CNN

584 VOLUME 11, 2024

Afrasiabi et al.: Robust Multi-Modal Deep Learning-Based Fault Diagnosis Method for PV Systems

TABLE 2. Summary of date-driven methods in fault identification a set of learning weights W : R0 →Rn to construct f (X ) and
of PV systems in the literature. accurately characterize the classification function. To find
these optimal learning weights, a loss function must be
assigned for the training process.

A. THE PROPOSED LOSS FUNCTION

The loss function is critical in the optimization process of
a training model. It quantifies the difference between the
predicted outputs and actual values, and serves as a measure
of the model’s performance. Traditionally, a loss function
aims to minimize the discrepancy between the predicted
outcomes and the ground truth labels to effectively align a
model’s prediction with its desired outcome. However, since
has been used to identify faults in PV systems [26], [34],
the fault identification of PV systems is involved with noises
it is powerful for recognizing spatial features, but cannot
and complex spatial-temporal features, a more sophisticated
fully recognize temporal features [35]. The comparison of
loss function is required. In this paper, a novel loss func-
different machine learning methods for the PV system’s fault
tion is proposed. The proposed loss function is based on
identification is summarized in Table 2. Among the limited
the information theory by incorporating entropy and mutual
PV system’s fault diagnosis methods in the literature that
information to quantify the uncertainty between the predicted
considers noises, most only consider Gaussian noises, which
and actual outputs. It allows the model to adapt to both Gaus-
don’t represent the majority noises in power systems.
sian and non-Gaussian noise, and improves the robustness of
To fill in these research gaps, a robust noise-immune deep
the model compared to the traditional loss functions, such as
neural network-based PV system fault diagnosis method is
cross-entropy.
proposed in this paper, which can handle both Gaussian and
In this paper, the entropy theory is used, and the entropy
non-Gaussian noises, and capture spatial-temporal features
function f ent is
of the data. The proposed method is validated using both
XA
simulation data and PV emulator experimental data [38] with f ent (A) = Prob (A = a) log [Prob (A = a)] (1)
a high level of Gaussian and non-Gaussian noises. The main a=1
contributions of this paper include: where A is a random variable, and Prob (A = a) is the prob-
• A multi-modal deep neural network model is proposed ability of a. Based on the information theory, the mutual
to accurately identify faults in PV systems for PV arrays, information of the two different random variables, A1 and A2 ,
inverters, feedback sensors, power point tracking con- is obtained as follows:
trollers on the DC side, grid anomalies, etc. XA1 XA2
fent (A1 , A2 ) = Prob (A1 = a1 , A2 = a2 )
• The deep neural network is developed to fully realize
a1 =1 a2
Prob (A1 = a1 , A2 = a2 )

spatial and temporal features by converting CNN into
log (2)
residual CNN, and then adding GNN to the network. Prob (A1 = a1 ) Prob (A2 = a2 )
• To improve the noise handling capability, a new proba-
Assume the three different random variables, A1 , A2 , and
bilistic loss function formulated using the information
A3 (A3 < A2 < A1 ), based on an informative measure,
theory is proposed to enhance the network’s learning
we have
capability and robustness, and allows the network to
operate without assumptions for the noise model. f ent (A1 , A2 ) ≤ f ent (A1 , A1 ) (3)
The rest of the paper is organized as follows: Section II
describes the principle of the designed deep network, includ- where the entropy of f ent (A1 , A1 ) is considered as the mutual
ing the proposed loss function and the training process. information function. According to (2), the optimal learning
The structure of the designed multimodal deep network is weight can be obtained based on the entropy function, so the
explained in Section III. Numerical results based on different network, f (X ), can be constructed.
noise models are given in Section IV. Section V concludes However, the entropy-based loss function is sensitive
the paper. to noises and cannot be generalized for the PV system’s
fault identification model. In the constructed f [W (X ) , X ],
the outputs are yc ; while the constructed function for a
II. TRAINING OF THE ROBUST DEEP NEURAL
data-driven machine  learning structure with noisy inputs
NETWORK
and outputs X ,´ Ý is f´ [W (X ) , X ]. Using (2) as the loss

To develop a data-driven fault identification model for PV
systems, a structure is first developed that can construct function, it is possible that f [W (X ) , X ] > f´ [W (X ) , X ].
a function f (X ) : R0 →Rn that fits a set of (X , Y ) = However, Eq. (2) is not a general loss function to train the
{(x1 , y1 ) , · · · , (xn , yc )} of n input/output data with c classes data-driven fault identification model, and can be highly
for various operations. A machine learning structure defines affected by noises. Thus, a general form of the entropy-based

VOLUME 11, 2024 585

loss function, f G−ent [W (X ) , X ], is adopted from [28] as B. THE TRAINING PROCEDURE
follows: In the training phase of the proposed fault identification
model, a four-step procedure is given below:
 
f G−ent (A1 , A2 ) = det ProbJa1 ,Ja2 (4)
Step 1: Initialize the network by assigning the ini-
where ProbJa1 ,Ja2 is the joint probability of the random vari- tial/constant values, such as the initial learning weights and
ables, A1 and A2 . iteration numbers.
Accordingly, for the random variables (A3 < A2 < A1 ), Step 2: Collect a set of data for the training process based
we have on a random process.
Step 3: For t < T ,the Adam algorithm is used to optimize
f G−ent (A2 , A3 ) the iteration t and find the optimal learning weights, then
= f G−ent (A1 , A2 ) |det (Prob (A3 = a3 |A1 = a1 ))| (5) the learning weights are updated accordingly. The process
continues until the iteration T .
Then, Step 4: The output is determined.
h i
f G−ent [W (X , X )] > f G−ent Ẃ (X , X ) The training process is summarized in Algorithm 1 below.
h  i
⇔ f G−ent W X , X́ Algorithm 1 Training Process
h i Input: The training set
> f G−ent Ẃ (X , X ) (6) Output: The set of learning weights w
Iterative training process
Therefore, the loss function is defined as follows: 1. Step 1: Initialization
h
i h i Step 1.1: Initialize all learning weights randomly
f loss ProbJ W (X ) , X ′ = −6i=1 C
W C log W (X ) , X́ 2. Step 2: Dataset selection
Step 2.1: 70% of the data are selected as the training dataset
(7) 3. Step 3: Find optimal values
4. While t < T do
where the size of ProbJ depends on the class number c and 5. Step 3.1: Check the stopping criteria
W C is the weight for class c ( W C = f1c , where f c shows the 6. Step 3.2: Find the optimal learning weights by the Adam
frequency of class c). In this paper, c= 8, so the size of ProbJ algorithm
is 7 × 7. The log term regulates the scale of ProbJ . 7. Step 3.3: Evaluate the proposed entropy-based loss function in
Eq. (7).
The output function of the data-driven fault identification 8. Step 3.4: Save the optimal values
model is obtained as follows: 9. Step 3.5: Back to Step 3.1.
XK Prob End while
f out (yc , ỹc ) = f out Qk yk = c̃
 
k=1 K net
10. Step 4: Output
XK 1 Step 4.1: Output the optimal learning weights
W (Xk ) Qk yk = c̃
 
= (8) End
k=1 K
 
where Qk yk = c̃ is a binary function:
( III. STRUCTURE OF THE PROPOSED FAULT
  1 if yk replace in class c̃ IDENTIFICATION MODEL FOR PV SYSTEMS
Qk yk = c̃ = (9)
0 otherwise The proposed data-driven fault identification model for PV
systems focuses on the training procedure, the loss function,
The size of Qk is Qk ∈ R1×K , where K is the total number of and a strong network to fully capture spatial and temporal
samples at each record. The main challenge is how to demon- features of measurement data. The designed network consists
strate that the designed network based on (8) is independent of three parallel structures, and each structure includes a
of noises and the noise model. residual CNN, a GNN, and the dense and weight fusion
Proof: To obtain the optimal learning weights, the pro- blocks. Fig. 1 shows the proposed network in this study.
posed loss function in (8) should be minimized as follows: At the beginning of the training process, the input dataset
is Xinput = [VPV , IPV , VDC , Vf , If , Vabc , Iabc , Va , Ia , Vb , Ib ,
h   i
argmin f G−net W X̃ , Ỹ = argmin f G−net [W (X ) , Y ]
ω∈W ω∈W Vc , Ic ], where VPV , IPV , VDC , Vf , and If are the PV array
(10) voltage, the PV array current, the DC voltage of the boost
converter, the voltage frequency, and the current frequency,
The input and output of the noisy labeled data are depicted respectively. In addition, three-phase voltages are Vabc , and
by X̃ and Ỹ , respectively. Based on (10), we have three-phase currents are Iabc . Va , Vb , Vc are per-phase volt-
h   i
f G−ent W X̃ , Ỹ = f G−net [W (x) , Y ] + κ (11) ages, and Ia , Ib , Ic are per-phase currents; they serve as the
model inputs and are normalized by
where the only difference is the constant value κ in (11).
xi − ximin
Therefore, the designed network is not affected by noises and Xinput = (12)
the noise model. xi − ximin
max

586 VOLUME 11, 2024

Afrasiabi et al.: Robust Multi-Modal Deep Learning-Based Fault Diagnosis Method for PV Systems

B. GNN
The output of Res-CNN is firstly converted from 2D signals
into 1D signals by a flattening layer, and is then fed into the
GNN (the second block). The GNN is made up of two key
components: the update gate and the reset gate. The update
gate keeps track of most significant features; the reset gate
discards less important features.

C. DENSE LAYER
A dense layer is added to each parallel component of GNN
to control its output, and aims to improve the network’s
training capability without the need of additional equipment.
It consists of multiple hidden layers.

D. WEIGHTED FUSION
The combination of CNN and GNN features is achieved
through a weighted fusion technique. After the CNN extracts
spatial features and the GNN captures temporal dependen-
cies, the outputs of both layers are combined through the
weighted fusion, where different importance is assigned to
each feature based on its contribution through the weight.
The weights are learned during training, ensuring an optimal
combination of spatial and temporal features for the fault
identification task. The weighted fusion determines the final
FIGURE 1. The structure of the proposed multimodal deep
network-based PV systems fault identification model.
output of the proposed model, which eventually defines the
condition of a PV system. The parallel structure enhances
the computational efficiency, reliability and accuracy of the
where xi represents the samples in the input dataset. To fully
proposed model. The output of the weighted fusion block is
capture spatial characteristics of a PV system, including
variations among different components and their positions, yout = W1 ⊙ f1 + W2 ⊙ f2 + W3 ⊙ f3 (15)
one-dimensional (1D) signals are transformed into two-
where the output of each part of the designed network and the
dimensional (2D) signals [29].
corresponding weights are f1/2/3 and W1/2/3 , respectively.
A. RESIDUAL CONVOLUTIONAL LAYER
E. NETWORK DESIGN
The proposed model is equipped with a residual convolu- The primary objective of this study is to develop a robust
tional neural network (Res-CNN) block. The transformation deep learning-based fault detection system for PVs, capable
of the original 1D signals into 2D matrices allows for the of identifying various faults that could impact the efficiency
extraction of both spatial and temporal features, which is and safety of PV operations in the presence of Gaussian and
essential for capturing complex interrelationships in multi- non-Gaussian noises. The data used in this paper is gener-
variate time-series data generated by PV systems. 2D CNNs ated from a PV emulator experimental test system and from
can be applied after such transformation, which are better simulation to ensure that the proposed method is validated in
suited to detect both local and global fault patterns compared both controlled experimental and simulation environments.
to 1D CNNs. The 2D measurement data in different parts of The input data consists of the voltage, current, and frequency
a PV system are the input of Res-CNN. Res-CNN consists of signals sampled at a rate of 10 µs, including per-phase and
multiple convolutional layers with the following output: three-phase voltage and current signals from both DC and
 
yCn = f Act wCn ⊗ Xinput + βl,z (13) AC sides of a PV system. Given the variety and complex-
l,z
ity of all types of faults, the proposed method focuses on
l,z , and βl,z are outputs of the activation
where yCn , f Act , wCn detecting the most critical faults of PV systems. Seven fault
function (ReLU). The symbol ⊗ depicts the convolution oper- types are considered: Inverter Faults, Feedback Sensor Faults,
ator. Residual mapping can improve the learning capability of Open-Circuit PV Array Mismatch, Partial Shading PV Array
convolutional layers. The output of Res-CNN considering the Mismatch, Boost Converter Controller Faults, Power Point
residual mapping is Tracking (MPPT/IPPT) Faults, and Grid Anomalies; plus the
normal operation condition. These faults may not be detected
yRCn = f map + Xinput
res
(14)
by conventional SCADA or hardware-based systems due to
where f map res
and Xinput are the mapping function and its input, their subtle nature, particularly under low-noise conditions or
respectively. partial system failures.

VOLUME 11, 2024 587

 There are 13 time-varying input variables involved in the (S, 128). The GRU can handle long-term dependencies
proposed fault identification model for PV systems. These efficiently.
input variables are divided into four groups: 1) 3-phase Step 5: Dense layers. The output of GNNs #1, #2, and #3
voltages V3φ and per phase voltages Vφ1, φ2 ,φ3 ; 2) 3-phase are controlled by the dense layers. The resulting outputs are
current I3φ and per phase currents Iφ1, φ2 ,φ3 ; 3) volt- the vectors, (S, 64), (S, 128), and (S, 64), respectively.
age/current frequency Vf /If , and DC link voltage VDC ; Step 6: Decision making. The outputs of each parallel part
4) PV array voltage/current VPV /IPV . These measurements are fused by the weight fusion block, and a PV system’s
collectively form the input dataset denoted as Xi = healthy or faulty condition can then be determined. To address
V3φ , I3φ , Iφ1, φ2 ,φ3 , Vf /If , VDC , VPV /IPV , representing the overfitting, the model includes the dropout layers with a
PV system’s electrical characteristics. dropout rate of 0.3, and L2 regularization is applied to the
The proposed model can be implemented in the following network layers. To improve the computational efficiency, the
six steps. The parallel Res-CNNs and GNNs are used to model uses multiple parallel GPUs to speed up training and
process the inputs. The Res-CNN block, comprising two testing.
convolutional layers with residual mapping, captures spa-
tial features. GNNs effectively characterize the time-varying IV. NUMERICAL RESULTS
behavior of the measured signals, and address the temporal In this section, the proposed fault identification model is
dimension. After undergoing the necessary processing, the validated subjected to noises and disturbances. The analysis
outputs are sequentially flattened using dense layers. In the was conducted in Python on a PC with an Intel Core i5-4200U
final step, the outputs are combined by a weight fusion block processor. The CNN [36] and the modified support vector
to determine the condition of a PV system. machine (MSVM) [13] are used for comparison.
Step 1: Initialization
Step 1.1: The dataset is standardized using (12). A. METRICS
Step 1.2: The normalized data are converted from 1D Three metrics are chosen for the performance evaluation in
signals to 2D signals. this paper, the accuracy (E r ), the specificity (E S ), and the
Step 2: Res-CNN block. The Res-CNN block is the first positive predictivity (E Pr ), as defined below:
of the three parallel blocks. Each Res-CNN block consists
TN + TP
of two convolutional layers, and the residual mapping is Er = (16)
conducted at the end of the second convolutional layer. Each TP + FP + FN + TN
TN
residual CNN block consists of two convolutional layers with ES = (17)
a kernel size of 3 × 3 and 5 × 5, respectively. The first layer FP + TN
TP
has a stride of 2, and the second layer has a stride of 1. The E Pr = (18)
activation function used is ReLU, which is applied after each FP + TP
convolutional layer. A residual connection is added at the The specificity and positive predictivity represent the model’s
end of each CNN block to improve learning efficiency by correct detection rate. These metrics are defined based on a
skipping one or more layers during training. confusion matrix composed of four main elements: the true
Step 2.1: In Res-CNN#1 and Res-CNN#2, the first con- positive (TP), the true negative (TN), the false positive (FP),
volution layer transforms the input data of PV systems to and the false negative (FN) [37].
a set of tensors with size (S, 4, 1, 1, 512) (S is the number
of samples). 15% of the input is dropped using a dropout B. DATASETS
technique. The size of the second, third, and fourth layers To develop the proposed data-driven fault identification
are (S, 2, 1, 1, 512), (S, 4, 1, 1, 256), and (S, 2, 1, 1, 256), method, three types of datasets can be used: 1) simulation
respectively. data, 2) PV emulator generated experimental data, and 3) the
Step 2.2: In Res-CNN#3, four different tensors are gen- real-world data. In this paper, the first two types of datasets
erated with the size of (S, 5, 1, 1, 1024) , (S, 5, 1, 1, 1024), are used to validate the proposed method. The details of the
(S, 2, 1, 1, 1024), and (S, 2, 1, 1, 1024) by the first, second, simulation data are given in the Subsection G.
third, and fourth convolutional layers, respectively. The PV emulator generated experimental dataset used in
Step 3: Flatten. The outputs from Step 2 are flattened by this paper is adopted from [38]. Fig. 2 shows the experimental
Dense#1, Dense#2, and Dense#3, and are converted into ten- set-up used to create various faults in the PV system and
sors with the size of (S, 4, 256) , (S, 4, 512), and (S, 2, 128), acquire this dataset [38]. It consists of a boost converter,
respectively. an inverter, a DSpace environment, a programmable Chroma
Step 4: GNN. The GNN layers take the output from grid emulator, a programmable Chroma solar array emulator,
the residual CNN blocks (the flattened 2D spatial data) and an AC load. The emulated faults are generated using
and process it as a sequence, capturing the temporal rela- industry-standard PV equipment, such as a programmable
tionship in the PV system data. The flattened outputs of solar array emulator and a grid emulator, to ensure the real-
GNN #1, #2, and #3 characterize various measured time- ism of the faults and align them with real-world operational
varying signals, and convert them into the vectors of shape, challenges in PV systems.

588 VOLUME 11, 2024

Afrasiabi et al.: Robust Multi-Modal Deep Learning-Based Fault Diagnosis Method for PV Systems

1) THE INVERTER FAULT (A)

Inverter faults occur due to the component degradation, over-
heating, or electrical faults within the inverter circuitry, and
affect the energy conversion process, the power flow, and
voltage regulations. Although with self-diagnostic capabili-
ties, inverters are often subjected to temperature variations
and physical stresses, leading to early degradation of power
switching devices and DC link capacitors, which can induce
incipient faults that may not be immediately detectable by
the inverter’s internal diagnostics system, and thus, needing
advanced fault diagnostic methods [39]. Inverter faults were
FIGURE 2. The PV emulator-based experimental set-up [38]. generated using MATLAB/Simulink by manipulating power
switches and the voltage input in the inverter model. The
TABLE 3. Different PV operations in dataset. resultant fault scenarios were then emulated in Fig. 2’s exper-
imental set-up [38].

2) THE FEEDBACK SENSOR FAULT (B)

Faulty sensors cause inefficient operations, safety haz-
ards, and the reduced system reliability, so their timely
fault detection ensures effective maintenance/repair. Con-
ventional protection may fail to detect sensor faults effec-
tively, as only significant data deviations are monitored,
so they often miss subtle changes, calibration drifts over
time, intermittent failures, and complex faults involving
interconnected components or environmental factors [40].
In experiments conducted in [38], feedback sensor faults were
Fault scenarios, such as the inverter anomalies and
replicated by modifying the system’s feedback signals using
open-circuit conditions in the PV array, were initially simu-
MATLAB/Simulink by introducing offset and scaling errors,
lated using MATLAB/Simulink, the simulated data were then
intermittent failures, and random noises into sensor readings,
accurately emulated using the Chroma grid and solar array
simulating various sensor fault scenarios. These simulated
emulators. The PV array’s output is simulated through the
faults were then emulated in the physical setup to collect the
programmable Chroma 62150H-1000S solar array emulator,
data that mimic the faulty sensor’s characteristics.
which can replicate a range of environmental conditions,
variations in solar radiation (G), and temperature (T) effects.
To replicate the grid behavior, the programmable AC source 3) THE OPEN-CIRCUIT PV ARRAY MISMATCH (C)
Chroma 61511 serves as a grid emulator. The control Open-circuit PV array mismatch occurs due to the perfor-
algorithm is implemented within the DSpace 1104 environ- mance differences in individual PV modules caused by manu-
ment, serving the dual purpose of the algorithm execution facturing defects, aging, shading, temperature variations, and
and data acquisition. The control is realized by integrating installation inconsistencies. Conventional detection methods,
the voltage oriented control (VOC) technique in tandem with such as visual inspections, electrical measurements, and ther-
the space vector pulse width modulation (SVPWM). This mal imaging, may fail to detect such faults due to their
amalgamation facilitates active and reactive power regula- subtle differences, changing environmental conditions, and
tion based on signals from the grid side. The output voltage the complexity of analyzing large PV arrays [8]. In Fig. 2’s
is synchronized with the grid voltage by the Phase Lock experiments, this fault is simulated using MATLAB/Simulink
Loop (PLL). An AC load is used in this setup, primarily by creating open-circuit conditions in the PV array model.
served for the protection, and also enabled faults injec- This involves selectively disconnecting modules and altering
tion [32]. The dataset is accessible by the following link: their electrical characteristics to represent the performance
https://data.mendeley.com/datasets/n76t439f65/1. variations. The Chroma solar array emulator then physically
This test case involves seven abnormal conditions (‘‘A’’- replicated these conditions [38].
‘‘G’’) and a healthy condition, ‘‘H’’ (Table 3 ). These faults
are selected based on their significant and profound impacts 4) THE PV ARRAY MISMATCH DUE TO PARTIAL
on the overall performance and reliability of a PV sys- SHEDDING (D)
tem, representing a diverse range of system-level faults to Partial shading on PV arrays from obstacles (trees or build-
demonstrate the robustness and effectiveness of the proposed ings) causes mismatches, as shaded panels produce less
model. electricity, affecting the entire series-connected array, and

VOLUME 11, 2024 589

thus, cause the reduced power output, potential hot spots
and mismatch losses. Detection methods of partial shading
include visual inspection, monitoring system alerts for low
outputs, infrared thermography to detect hot spots, and the
I-V curve analysis for the performance assessment, but they
may not be sufficient due to the intermittent nature of partial
shading and the limited system sensitivity [8]. In Fig. 2’s
experiments, this mismatch is created by selectively discon-
necting certain modules in the PV array model, and altering
electrical output characteristics of these modules to reflect the
performance variations, such as the reduced efficiency, due to
partial shading [38].

5) THE BOOST CONVERTER CONTROLLER FAULT (E)

A boost converter’s controller fault occurs due to wear and
tear, electrical surges, or manufacturing defects in com-
ponents that increase the DC voltage from PV panels.
Conventional protection for this fault includes the system
diagnostics and performance monitoring, but may fail to FIGURE 3. The PV output voltage under different faults in a PV
system.
detect complex/ intermittent faults [41]. This fault is sim-
ulated by adding anomalies in the control signals using
MATLAB/Simulink, including distorting the signal wave-
forms and altering the timing of the PWM signals.

6) GRID ANOMALIES (F)

Grid anomalies due to faults in the utility grid or sudden
changes in the load demand can affect stability and reliabil-
ity of grid-connected PV systems. Grid-tied inverters with
monitoring capabilities can detect these faults, but may fail
to capture transient or minor fluctuations [20]. Grid anoma-
lies are simulated with disturbances including voltage sags
(temporarily reducing the grid voltage below its nominal FIGURE 4. Loss vs. epoch.
value), voltage swells (temporarily increasing the grid volt-
age higher than its nominal value), and frequency fluctua-
generalization across different subsets of the data. Fault
tions [38].
scenarios incorporate randomized noise injections and partial
shading conditions, in addition to traditional faults. Note: The
7) POWER POINT TRACKING CONTROLLER FAULTS ON dataset’s labels impose certain limitations on the fault types
THE DC SIDE (G) the proposed model can detect, but the proposed method can
Faults in the power point tracking controller can hinder the provide accurate fault detection for label-defined faults with
system’s ability to maximize the power extraction from PV room for future improvement for more broaden coverage.
arrays, leading to suboptimal operations and the reduced To illustrate the optimization of the proposed model, a loss
energy yield. They are hard to detect due to intermittent or curve is presented in Fig. 4, showing the reduction in the loss
minor faults in the tracking algorithm [42]. In the experi- value over training epochs.
ments in [38], this fault is simulated by adjusting the MPPT Among the factors affecting fault identification for PV
algorithm, adding delays to the algorithm’s response, and systems, one influential factor is the noise. In this study,
incorporating sensors errors, which create suboptimal oper- the proposed fault identification method is validated using
ating conditions. the measured fault data with the simulated noises added,
The experimental dataset in [38] was collected with a providing a more realistic testing environment. The following
sampling frequency of 10µs. Fig. 3 shows the PV output three cases are defined using three different noise models:
voltage under different faults. 1) Case I - Gaussian noise at a signal-to-noise ratio (SNR) of
 For training, the input samples are organized by 20 dB, 2) Case II - a combination of three levels of Gaussian
Xinput , Ý , and ∀k = 1, 2, · · · , N . k is the sample point,
k k
noises (SNR = 20 dB, 30 dB, and 40 dB), and 3) Case III -
and N is the total number of samples (In this dataset, N = 64). Laplacian noise with a Chi-square probability distribution
70% data are for training, and 30% data are for testing. (the degree of freedom = 4). Fig. 5 shows the phase A
A five-fold cross-validation technique ensures the model’s currents with and without noises.

590 VOLUME 11, 2024

Afrasiabi et al.: Robust Multi-Modal Deep Learning-Based Fault Diagnosis Method for PV Systems

TABLE 4. Accuracy of PV fault identification for Case I.

TABLE 5. Accuracy of PV fault identification for Case II.

FIGURE 5. Phase A current signal (top), noise signals (middle),

and noisy phase A current signal with Gaussian mixture noises
(bottom).

FIGURE 7. Comparison of three metrics for fault identification

using the proposed method, CNN, and MSVM for Case II.

TABLE 6. Accuracy of PV fault identification for Case III.

FIGURE 6. Comparison of three metrics for fault identification

using the proposed method, CNN and MSVM for Case I.

C. CASE I: GAUSSIAN NOISES

Case I is defined under Gaussian noises at SNR = 20 dB. mean accuracy is 87.49%, and the MSVM’s mean accuracy
The fault identification accuracy for Case I using the pro- is 84.52%. The proposed model improves the accuracy of
posed model is compared with that using CNN and MSVM, CNN and MSVM by 13.01 % and 16.99%, respectively. Fig. 7
as shown in Fig. 6. The proposed model shows superior shows superior performance of the proposed model among
performance according to the three metrics for classification three metrics. In terms of E Pr , the proposed model improves
of the seven faults. Table 4 shows the mean and the standard the positive predictivity of CNN and the MSVM by 65.15%
deviation of the accuracy. The mean accuracy of the proposed and 82.99%, respectively.
model is 98.96%, showing its effectiveness in handling Gaus-
sian noises. E. CASE III: LAPLACIAN NOISES
Case III is defined under Laplacian noises. Table 6 shows
D. CASE II: GAUSSIAN MIXTURE NOISES the mean and standard deviation of the accuracy results in
Case II is defined as a combined three levels of Gaussian Case III. For the three metrics, the proposed model performs
noises (SNRs = 20 dB, 30 dB, and 40 dB). Table 5 and Fig. 7 with better accuracy and reliability than CNN and MSVM,
depict the results of fault classification using the proposed as shown in Fig. 8. The proposed model achieves a 99.02%
model, CNN and MSVM in Case II. In Table 5, the proposed mean accuracy, a notable improvement of the mean accuracy
model has a mean accuracy of 98.88%, while the CNN’s at 8.51% over CNN and 11.19% over MSVM.

VOLUME 11, 2024 591

 FIGURE 9. Comparison of the proposed loss function with three
conventional loss functions, CE, MCE, and (CE+MSE), for
Case II.

FIGURE 8. Comparison of three metrics for fault identification

using the proposed method, CNN and MSVM for Case III.

F. IMPACT OF THE LOSS FUNCTION

To demonstrate its efficiency, the proposed loss function is
compared with the conventional loss functions in multiclass
classification problems. The three existing loss functions in
the literature, Cross-entropy (CE), modified cross-entropy
(MCE) [29], and (cross-entropy + mean squared error)
(CE+MSE) [24], are selected for comparison. The results are
compared under two faults, inverter faults (A) and PV array
mismatch considering partial shedding (D). Fig. 9 shows
the results in terms of the three metrics. The proposed loss
function is more robust compared with the existing cross-
entropy loss-based loss functions.

G. ADAPTABILITY OF THE PROPOSED METHOD FIGURE 10. Three performance metrics comparison for the
To validate the proposed fault identification method using proposed fault identification method, CNN and MSVM using
the simulation data, the following four fault types in PV simulation data when subjected to Laplacian noises.
arrays are simulated: line-to-line (L-L) faults (connecting two
TABLE 7. Time of different methods used for PV systems fault
lines in the model), a single line to ground faults (shorting a identification.
line to the ground), open-circuit faults (disconnecting lines
in the PV array), and partial shading (adjusting irradiance
parameters for specific panels within the PV array model
to replicate the effect of physical shading on PV panels,
impacting their power output). The PV system simulation is
conducted using MATLAB/Simulink in a radial distribution
network at a voltage level of 25 kV, fed from an upper power H. COMPUTATIONAL COMPLEXITY
grid with a short-circuit capacity of 2,500 MVA. A 250 kV PV The computational time is an important factor for the PV
system with 10 series PV panels, a boost DC-DC converter system fault diagnosis, so the proposed model is assessed
and a voltage source inverter is included in the distribution in terms of the implementation time, along with CNN and
network. The simulation study involves 3,730 PV system MSVM-based models, for 500 runs in this paper. The min-
operating conditions [24]. 70% of the simulation data is used imum, average, and maximum time are shown in Table 7.
for training, and 30% for testing. A comparison of the three The MSVM-based model performs the fastest among the
metrics for the proposed method, CNN and MSVM using the three, but has the accuracy and reliability issues, making it
simulation data is shown in Fig. 10 subjected to Laplacian unsuitable for the PV system fault identification. The parallel
noises. The proposed method shows superior performance structure of the proposed model performs much faster than
compared to CNN and MSVM. the CNN-based model. The parallel structure means that the

592 VOLUME 11, 2024

Afrasiabi et al.: Robust Multi-Modal Deep Learning-Based Fault Diagnosis Method for PV Systems

required number of layers is reduced, so the computational [14] F. Grimaccia, S. Leva, A. Dolara, and M. Aghaei, ‘‘Survey on PV modules’
time is also less. common faults after an O&M flight extensive campaign over different
plants in Italy,’’ IEEE J. Photovolt., vol. 7, no. 3, pp. 810–816, May 2017.
[15] S. Afrasiabi, M. Afrasiabi, B. Parang, M. Mohammadi, S. Kahourzade, and
V. CONCLUSION AND FUTURE WORK A. Mahmoudi, ‘‘Two-stage deep learning-based wind turbine condition
In this paper, a robust deep neural network-based model is monitoring using SCADA data,’’ in Proc. IEEE Int. Conf. Power Electron.,
developed for fault diagnosis in PV systems. The designed Drives Energy Syst. (PEDES), Dec. 2020, pp. 1–6.
[16] S. Afrasiabi, M. Afrasiabi, M. A. Jarrahi, and M. Mohammadi, ‘‘Fault
model has the multi-modal structure, consisting of Res-CNN location and faulty line selection in transmission networks: Applica-
to fully capture spatial features and GRU to learn temporal tion of improved gated recurrent unit,’’ IEEE Syst. J., vol. 16, no. 3,
features of the PV outputs. To enhance the learning capability, pp. 5056–5066, Sep. 2022.
the proposed model includes three parallel parts. An entropy [17] S. Vergura, ‘‘Correct settings of a joint unmanned aerial vehicle and
infrared camera system for the detection of faulty photovoltaic modules,’’
theory-based loss function is formulated, leading to excellent IEEE J. Photovolt., vol. 11, no. 1, pp. 124–130, Jan. 2021.
accuracy and reliability in the presence of Gaussian and non- [18] S. Afrasiabi, M. Afrasiabi, B. Parang, M. Mohammadi, H. Samet, and
Gaussian noises. In case studies, the proposed model achieves T. Dragicevic, ‘‘Fast GRNN-based method for distinguishing inrush cur-
rents in power transformers,’’ IEEE Trans. Ind. Electron., vol. 69, no. 8,
more than 98% accuracy with various noises, much better pp. 8501–8512, Aug. 2022.
than that of CNN- and MSVM-based models. The proposed [19] Y. Yang, D.-C. Zhan, Y.-F. Wu, Z.-B. Liu, H. Xiong, and Y. Jiang, ‘‘Semi-
model is validated using the datasets generated from the PV supervised multi-modal clustering and classification with incomplete
emulator experimental test system and from the simulation. modalities,’’ IEEE Trans. Knowl. Data Eng., vol. 33, no. 2, pp. 682–695,
Feb. 2021.
In this paper, no field tests are conducted on real solar PV [20] K. Dhibi, M. Mansouri, K. Bouzrara, H. Nounou, and M. Nounou, ‘‘An
systems, and the proposed method is validated based on emu- enhanced ensemble learning-based fault detection and diagnosis for grid-
lated/simulated faults. The future work will include field tests connected PV systems,’’ IEEE Access, vol. 9, pp. 155622–155633, 2021.
for further validation. The next step also includes expanding [21] S. Afrasiabi et al., ‘‘Wide-area composite load parameter identification
based on multi-residual deep neural network,’’ IEEE Trans. Neural Netw.
the current model to include identifying fault severity or Learn. Syst., vol. 34, no. 9, pp. 6121–6131, Aug. 2023.
potential root causes. [22] K. Dhibi et al., ‘‘A hybrid fault detection and diagnosis of grid-tied
PV systems: Enhanced random forest classifier using data reduction and
REFERENCES interval-valued representation,’’ IEEE Access, vol. 9, pp. 64267–64277,
2021.
[1] S. Allahmoradi, S. Afrasiabi, X. Liang, J. Zhao, and M. Shahidehpour,
‘‘Data-driven Volt/VAR optimization for modern distribution networks: A [23] S.-Q. Chen, G.-J. Yang, W. Gao, and M.-F. Guo, ‘‘Photovoltaic fault diag-
review,’’ IEEE Access, vol. 12, pp. 71184–71204, 2024. nosis via semisupervised ladder network with string voltage and current
measures,’’ IEEE J. Photovolt., vol. 11, no. 1, pp. 219–231, Jan. 2021.
[2] D. S. Pillai, F. Blaabjerg, and N. Rajasekar, ‘‘A comparative evaluation of
advanced fault detection approaches for PV systems,’’ IEEE J. Photovolt., [24] S. Afrasiabi et al., ‘‘Photovoltaic array fault detection and classification
vol. 9, no. 2, pp. 513–527, Mar. 2019. based on T-distributed stochastic neighbor embedding and robust soft
[3] M. U. Saleh et al., ‘‘An overview of spread spectrum time domain reflec- learning vector quantization,’’ in Proc. IEEE Int. Conf. Environ. Electr.
tometry responses to photovoltaic faults,’’ IEEE J. Photovolt., vol. 10, Eng. IEEE Ind. Commercial Power Syst. Eur. (EEEIC/ICPS Europe),
no. 3, pp. 844–851, May 2020. Sep. 2021, pp. 1–5.
[4] S. Afrasiabi, S. Allahmoradi, M. Salimi, X. Liang, and C. Y. Chung, [25] A. Eskandari, J. Milimonfared, and M. Aghaei, ‘‘Fault detection and clas-
‘‘Machine learning-based condition monitoring of solar photovoltaic sys- sification for photovoltaic systems based on hierarchical classification and
tems: A review,’’ in Proc. IEEE Can. Conf. Electr. Comput. Eng. (CCECE), machine learning technique,’’ IEEE Trans. Ind. Electron., vol. 68, no. 12,
Sep. 2022, pp. 49–54. pp. 12750–12759, Dec. 2021.
[5] M. Kavi, Y. Mishra, and M. Vilathgamuwa, ‘‘DC arc fault detection [26] M. U. Ali, H. F. Khan, M. Masud, K. D. Kallu, and A. Zafar, ‘‘A machine
for grid-connected large-scale photovoltaic systems,’’ IEEE J. Photovolt., learning framework to identify the hotspot in photovoltaic module using
vol. 10, no. 5, pp. 1489–1502, Sep. 2020. infrared thermography,’’ Sol. Energy, vol. 208, pp. 643–651, Sep. 2020.
[6] W. Gao and R.-J. Wai, ‘‘Series arc fault detection of grid-connected PV [27] A. Jamali Jahromi, M. Mohammadi, S. Afrasiabi, M. Afrasiabi, and
system via SVD denoising and IEWT-TWSVM,’’ IEEE J. Photovolt., J. Aghaei, ‘‘Probability density function forecasting of residential elec-
vol. 11, no. 6, pp. 1493–1510, Nov. 2021. tric vehicles charging profile,’’ Appl. Energy, vol. 323, Oct. 2022,
[7] M. K. Alam, F. Khan, J. Johnson, and J. Flicker, ‘‘A comprehensive Art. no. 119616.
review of catastrophic faults in PV arrays: Types, detection, and mitigation [28] S. Afrasiabi, S. Allahmoradi, M. Salimi, X. Liang, and C. Y. Chung,
techniques,’’ IEEE J. Photovolt., vol. 5, no. 3, pp. 982–997, May 2015. ‘‘Nonparametric maximum likelihood probabilistic photovoltaic power
[8] D. S. Pillai and R. Natarajan, ‘‘A compatibility analysis on NEC, IEC, and generation forecasting based on spatial–temporal deep learning,’’ in Proc.
UL standards for protection against line–line and line–ground faults in PV IEEE Can. Conf. Electr. Comput. Eng. (CCECE), Sep. 2022, pp. 72–77.
arrays,’’ IEEE J. Photovolt., vol. 9, no. 3, pp. 864–871, May 2019.
[29] Y. Liu et al., ‘‘Fault diagnosis approach for photovoltaic array based on the
[9] W. Miao, Y. Luo, F. Wang, and C. Jiang, ‘‘Fault detection and location
stacked auto-encoder and clustering with I-V curves,’’ Energy Convers.
algorithm by voltage characteristics for PV system,’’ IEEE J. Photovolt.,
Manage., vol. 245, Oct. 2021, Art. no. 114603.
vol. 13, no. 6, pp. 968–978, Nov. 2023.
[30] C. Tao, X. Wang, F. Gao, and M. Wang, ‘‘Fault diagnosis of photovoltaic
[10] S. Roy, M. K. Alam, F. Khan, J. Johnson, and J. Flicker, ‘‘An irradiance-
array based on deep belief network optimized by genetic algorithm,’’ Chin.
independent, robust ground-fault detection scheme for PV arrays based
J. Electr. Eng., vol. 6, no. 3, pp. 106–114, Sep. 2020.
on spread spectrum time-domain reflectometry (SSTDR),’’ IEEE Trans.
Power Electron., vol. 33, no. 8, pp. 7046–7057, Aug. 2018. [31] M. Alrifaey et al., ‘‘Hybrid deep learning model for fault detection and clas-
[11] B. P. Kumar, G. S. Ilango, M. J. B. Reddy, and N. Chilakapati, ‘‘Online fault sification of grid-connected photovoltaic system,’’ IEEE Access, vol. 10,
detection and diagnosis in photovoltaic systems using wavelet packets,’’ pp. 13852–13869, 2022.
IEEE J. Photovolt., vol. 8, no. 1, pp. 257–265, Jan. 2018. [32] J. Van Gompel, D. Spina, and C. Develder, ‘‘Satellite based fault diagnosis
[12] P. B. Quater, F. Grimaccia, S. Leva, M. Mussetta, and M. Aghaei, ‘‘Light of photovoltaic systems using recurrent neural networks,’’ Appl. Energy,
unmanned aerial vehicles (UAVs) for cooperative inspection of PV plants,’’ vol. 305, Jan. 2022, Art. no. 117874.
IEEE J. Photovolt., vol. 4, no. 4, pp. 1107–1113, Jul. 2014. [33] M. Afrasiabi, J. Aghaei, S. Afrasiabi, and M. Mohammadi, ‘‘Probability
[13] F. Harrou, A. Saidi, Y. Sun, and S. Khadraoui, ‘‘Monitoring of photovoltaic density function forecasting of electricity price: Deep Gabor convolu-
systems using improved kernel-based learning schemes,’’ IEEE J. Photo- tional mixture network,’’ Electric Power Syst. Res., vol. 213, Dec. 2022,
volt., vol. 11, no. 3, pp. 806–818, May 2021. Art. no. 108325.

VOLUME 11, 2024 593

[34] B. K. Karmakar and A. K. Pradhan, ‘‘Detection and classification of MOUSA AFRASIABI (Member, IEEE) received
faults in solar PV array using Thevenin equivalent resistance,’’ IEEE J. the B.Sc. degree in electrical power engineering
Photovolt., vol. 10, no. 2, pp. 644–654, Mar. 2020. from the University of Guilan, Rasht, Iran, in 2008,
[35] M. Mostafanezhad, M. Mohammadi, S. Afrasiabi, M. Afrasiabi, J. Aghaei, the M.Sc. degree in electrical power engineering
and C. Y. Chung, ‘‘Data-driven small-signal and N-1 security assess- from the K. N. Toosi University of Technology,
ment considering missing data,’’ IEEE Trans. Power Syst., vol. 39, no. 2, Tehran, Iran, in 2011, and the Ph.D. degree in elec-
pp. 2587–2597, Mar. 2024. trical power engineering from Shiraz University,
[36] X. Li, Q. Yang, Z. Lou, and W. Yan, ‘‘Deep learning based module defect Shiraz, Iran, in 2020. His current research interests
analysis for large-scale photovoltaic farms,’’ IEEE Trans. Energy Convers.,
include energy management, time series forecast-
vol. 34, no. 1, pp. 520–529, Mar. 2019.
ing, machine learning, power system dynamic
[37] M. Pournabi, M. Mohammadi, S. Afrasiabi, and P. Setoodeh, ‘‘Power
system transient security assessment based on deep learning considering analysis, electric machine design, and power system probabilistic analysis.
partial observability,’’ Electric Power Syst. Res., vol. 205, Apr. 2022,
Art. no. 107736. XIAODONG LIANG (Senior Member, IEEE)
[38] A. Bakdi, W. Bounoua, A. Guichi, and S. Mekhilef, ‘‘Real-time fault detec- was born in Lingyuan, Liaoning, China. She
tion in PV systems under MPPT using PMU and high-frequency multi- received the B.Eng. and M.Eng. degrees in elec-
sensor data through online PCA-KDE-based multivariate KL divergence,’’ trical engineering from Shenyang Polytechnic
Int. J. Electr. Power Energy Syst., vol. 125, Feb. 2021, Art. no. 106457. University, Shenyang, China, in 1992 and 1995,
[39] A. Malik, A. Haque, V. S. B. Kurukuru, M. A. Khan, and F. Blaabjerg, respectively, the M.Sc. degree in electrical engi-
‘‘Overview of fault detection approaches for grid connected photovoltaic neering from the University of Saskatchewan,
inverters,’’ e-Prime Adv. Electr. Eng., Electron. Energy, vol. 2, Jul. 2022,
Saskatoon, Canada, in 2004, and the Ph.D. degree
Art. no. 100035.
in electrical engineering from the University of
[40] S. R. Madeti and S. N. Singh, ‘‘A comprehensive study on different types of
faults and detection techniques for solar photovoltaic system,’’ Sol. Energy,
Alberta, Edmonton, Canada, in 2013.
vol. 158, pp. 161–185, Dec. 2017. From 1995 to 1999, she was a Lecturer with Northeastern University,
[41] D. Espinoza Trejo, E. Bárcenas, J. Hernández Díez, G. Bossio, and G. Shenyang, China. In October 2001, she joined Schlumberger (SLB), Edmon-
Espinosa Pérez, ‘‘Open- and short-circuit fault identification for a boost ton, and was promoted to a Principal Power Systems Engineer with this
DC/DC converter in PV MPPT systems,’’ Energies, vol. 11, no. 3, p. 616, world’s leading oil field service company in 2009. She worked with Schlum-
Mar. 2018. berger for almost 12 years until August 2013. From 2013 to 2019, she was
[42] T. Berghout, M. Benbouzid, T. Bentrcia, X. Ma, S. Djurović, and with Washington State University, Vancouver, WA, USA, and the Memorial
L.-H. Mouss, ‘‘Machine learning-based condition monitoring for PV sys- University of Newfoundland, St. John’s, NL, Canada, as an Assistant Profes-
tems: State of the art and future prospects,’’ Energies, vol. 14, no. 19, sor and later an Associate Professor. In July 2019, she joined the University
p. 6316, Oct. 2021. of Saskatchewan, where she is currently a Professor and the Canada Research
Chair in Technology Solutions for Energy Security in Remote, Northern,
and Indigenous Communities. She was an Adjunct Professor with Memorial
University of Newfoundland from 2019 to 2022. Her research interests
include power systems, renewable energy, and electric machines.
Dr. Liang is a Registered Professional Engineer in the Province of
Saskatchewan, Canada; a fellow of IET; and the Deputy Editor-in-Chief of
IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS.

SHAHABODIN AFRASIABI (Member, IEEE) C. Y. CHUNG (Fellow, IEEE) received the B.Eng.
received the B.Sc. degree in electrical engineering (Hons.) and Ph.D. degrees in electrical engineer-
from Semenan University, Semnan, Iran, in 2014, ing from The Hong Kong Polytechnic University,
the M.Sc. degree in electrical engineering from Hong Kong, China, in 1995 and 1999, respec-
the Shahid Chamran University of Ahvaz, Ahvaz, tively.
Iran, in 2017, and the Ph.D. degree in electrical He is currently the Head of Department and
engineering from the University of Saskatchewan, the Chair Professor of power systems engineer-
Saskatoon, Canada, in 2024. His research inter- ing with the Department of Electrical Engineer-
ests include power system dynamics, machine ing, The Hong Kong Polytechnic University. His
learning, state estimation, and power system research interests include smart grid technologies,
probabilistic analysis. renewable energy, power system stability/control, planning and operation,
computational intelligence applications, power markets, and electric vehicle
charging.

JAMSHID AGHAEI (Senior Member, IEEE)

received the B.Sc. degree in electrical engineering
from the Power and Water Institute of Technol-
ogy, Tehran, Iran, in 2003, and the M.Sc. and
Ph.D. degrees from Iran University of Science and
SARAH ALLAHMORADI (Member, IEEE) Technology, Tehran, in 2005 and 2009, respec-
received the B.Sc. degree in electrical engineering tively. He is currently a Full Professor with Central
from the University of Kurdistan, Kurdistan, Iran, Queensland University, Rockhampton, Australia,
in 2016, the M.Sc. degree (Hons.) in electrical with a research focus on smart grids, renewable
engineering from Tarbiat Modares University, energy systems, electricity markets, and power
Tehran, Iran, in 2020, and the Ph.D. degree system operation, optimization, and planning. He serves as an Associate
in electrical engineering from the University of Editor for IEEE TRANSACTIONS ON SMART GRID, IEEE SYSTEMS JOURNAL, and
Saskatchewan, Saskatoon, Canada, in 2024. Her IET Renewable Power Generation; and a Subject Editor for IET Generation,
research interests include smart grids, optimiza- Transmission and Distribution. In 2017, he was recognized as an Outstanding
tion, and machine learning applications in power Reviewer for IEEE TRANSACTIONS ON SUSTAINABLE ENERGY.
systems.

594 VOLUME 11, 2024