Fault Detection in Solar PV Systems Using

Hypothesis Testing
1 Fouzi Harrou, 2,3 Bilal Taghezouit, Benamar Bouyeddou4,5 , 1 Ying Sun, 2 Amar Hadj Arab
2021 IEEE 19th International Conference on Industrial Informatics (INDIN) | 978-1-7281-4395-8/21/$31.00 ©2021 IEEE | DOI: 10.1109/INDIN45523.2021.9557582

1
King Abdullah University of Science and Technology, CEMSE Division
Thuwal, 23955-6900, Saudi Arabia, [email protected]
2
Centre de Développement des Energies, Renouvelables, CDER, 16340, Algiers, and Laboratoire
de Dispositif de Communication et de Conversion Photovoltaique,
3
Ecole Nationale Polytechnique Alger, 16200, Algeria.
4
Abou Bekr Belkaid University, STIC Lab., Department of Telecommunications, Tlemcen, Algeria
5
University of Saida-Dr Moulay Tahar, Department of Electronics, Faculty of Technology, Saida, Algeria

Abstract—The demand for solar energy has rapidly increased

throughout the world in recent years. However, anomalies in
photovoltaic (PV) plants can reduce performances and result in
serious consequences. Developing reliable statistical approaches
able to detect anomalies in PV plants is vital to improving
the management of these plants. Here, we present a statistical
approach for detecting anomalies in the DC part of PV plants
and partial shading. Firstly, we model the monitored PV plant.
Then, we employ a generalized likelihood ratio test, which is
a powerful anomaly detection tool, to check the residuals from
the model and reveal anomalies in the supervised PV array. The
proposed strategy is illustrated via actual measurements from a Fig. 1. Left: 383 kWp PV plant, California (fire hazard, 2009). Right: 1,208
9.54 PV plant. MWp PV plant, NC (fire hazard, 2011) [1].
Index Terms—Photovoltaic systems, fault detection, partial
shading, GLR test, statistical monitoring charts.
Accurate and early detection of faults in PV systems is
I. I NTRODUCTION becoming the backbone to facilitate developing efficient PV
systems that meet the desired requirements and specifica-
Faced with the great energy transition in the world, the tions [4], [5]. All over the years, various data-based and model-
electricity generation using renewable energy has marked based methods have been introduced to improve the detection
a fast-growing year-by-year. According to the International of faults in PV plants [3], [6]. For instance, In [7], a detection
Renewable Energy Agency, a capacity of 171 gigawatts (GW) procedure based on a double exponential chart is proposed to
was added in 2018 primarily by new solar and wind, and this monitor photovoltaic systems. In [8], an approach to monitor
capacity additions represent 84% of all new capacity installed in an automatic way the degradation of the PV system by
in 2018. This brings total renewable energy generation capac- analyzing the I-V features is introduced. Indeed, analyzing I-V
ity up to 2,351 GW at the end of 2018. characteristics for several values of irradiance and temperature
by comparing the measured I-V with the simulated one are
Failures in PV arrays are frequently challenging to avoid frequently used in the literature to detect and identify the
and may generate a loss of energy, shutdowns or even in failure appearing in the PV module, string or array (e.g.,
critical safety problems [1]–[3] as illustrated in Figure 1. As shading, short circuit, aging and degradation, and the dust
an example of critical failures in PV plants, Figure 1 are taken effect) [9]. In [10], the authors applied a neural network al-
respectively from two PV plants that were burned in the United gorithm for fault detection using I-V characteristics. However,
States (Bakersfield, CA, and Mount Holly, NC) in 2009 and it should be noted that the use of I-V characteristics requires
2011, respectively [1] due to undetected anomalies in their switching off the PV generator to collect measurements of
DC parts. Thus, continuous and accurate monitoring of PV the I-V characteristic, which limits the application of this
plants is crucial for the benefit of the PV project investments approach for online monitoring. In [11], a method coupling
and to reach the expected energy yield and maximize the criteria in the time and frequency domain is introduced to
PV system’s useful life. Indeed, a monitoring system is an detect parallel arc faults in the DC side of PV arrays. In
indispensable part to maintain optimal performance of PV [12], an approach based on a wavelet transform using the
systems, which allows to increase reliability, efficiency, service current at the exit of a PV panel is suggested to detect DC
lifetime performance and safety in PV plants and to reduce the side parallel arc fault. In [13], authors propose a classifier to
cost of operation and maintenance [4]. identify faults and shadows that occurred in PV based on two
new indicators related to radiation and ambient temperature. to sense the deviations between normal and abnormal features
In [14], an approach has been developed for fault detection and to better detect faults in PV systems. It should be pointed
localization if PV systems by merging the Gaussian kernel- out that univariate anomaly detection schemes (e.g., GLR) are
Fuzzy C Means algorithm for unsupervised clustering and a designed based on the hypothesis that data are uncorrelated.
probabilistic neural network to model fault process. In [15], However, measurements from the DC part of PV plants
an unsupervised strategy is introduced by using the one-diode are autocorrelated. To remedy this difficulty, the proposed
model and one-class SVM [15]. In one-class SVM, using fault- detection framework amalgamates a simulation model and a
free data the optimal decision hyperplane is determined to GLR test to detect anomalies in PV plants. First, a simulation
separate normal measurements from faulty data, and then used model is used for describing the evolution of MPP variables
for fault detection. The key characteristics of this approach are by using as input the ambient temperature and tilted solar
its flexibility to handle nonlinear and non-Gaussian data, and irradiance measurements. Then, the GLR is used for checking
it is assumption-free on data. In [16], several kernel-based the residuals generated from the simulation model to uncover
machine learning methods have been adopted for photovoltaic anomalies in PV plants. The capacity of the proposed scheme
plant monitoring. Many other machine learning algorithms is assessed via practical data from a 9.54 kWp PV array.
have been widely exploited in the diagnosis of photovoltaic
The following section describes the considered PV array.
systems including random forest [17], decision tree algo-
Then, Section III presents the simulation model. In Section IV,
rithm [18], and extreme learning machine [19].
the designed anomaly detection methodology is presented.
Statistical process control methods are considered as a pow- Section V checks the detection performance of the developed
erful tool for monitoring production systems. They are widely procedure. Lastly, Section VI offers some concluding remarks
used in several fields including manufacturing [20], health- of this study.
care [21], and traffic management [22]. Numerous univariate
II. PV SYSTEM DSCRIPTION
anomaly detection methods were designed in the literature
such as the cumulative sum schemes, and the exponential The fault detection method proposed in this study is as-
smoothing methods. However, these monitoring approaches sessed using measured data gathered from a grid-tied PV
are generally designed so that they are more or less sensitive to system in Algiers. The PV array consists of 90 modules
a specific range of changes [23]. For instance, Shewhart-type (Isofotron 106/12) with a total peak dc power of 9.54KWp.
charts are more suited to detect large changes, whereas other The PV plant is formed of a PV array of 90 panels. The PV
charts are more appropriate to uncover small changes [4]. In system is formed with three similar PV generators of 3.18
fact, CUSUM and EWMA methods may be set to be very kWp.
suitable for sensing small changes, but they lack efficiency Every PV generator is linked in a way to have the following
in detecting large changes. To cope with this challenge, one configuration: two parallel string of fifteen PV panels in series.
approach to get a desirable detection performance for different The PV generator is cabled through an electrical protection
ranges of changes is to employ the monitoring scheme based cabinet to the inverter (Fronius IG 30). The essential electrical
on a likelihood ratio test (GLR scheme) [24], [25]. The GLR features of the PV generator and inverter are given in Table I
has demonstrated good ability in detecting a wide range of and Table II, respectively.
settings making them advantageous for use in real applica-
tions [25]. A desirable characteristic of GLR schemes is their TABLE I
flexibility and ease to design; they need only the detection limit E SSENTIAL PROPERTIES OF PV PANEL AND GENERATOR .
to be defined [25], [26]. In recent years, deep learning-based
methods have received considerable attention in the diagnosis
of PV systems because of their extended performance and abil-
ity to learn complex features. Several deep learning algorithms
have been applied to improve the monitoring of PV systems,
such as a convolutional neural network [27], long short term
TABLE II
machine [28], and showed superior performance compared to E SSENTIAL PROPERTIES OF THE USED PV INVERTER .
the shallow learning methods.
The objective of this work is to propose an accurate anomaly
detector for supervising the DC part of a PV plant. To do so,
we proposed an efficient statistical scheme using a generalized
likelihood ratio (GLR) test. GLR detector, which is a statis-
tical hypothesis testing approach, provides desirable detection Figure 2 displays the monitoring system used for requiring
capability via maximization of the detection probability of and analyzing the collected measurements using the sensors.
anomalies for predefined false alarm probability. It has been Table III summarizes the features of the sensors utilized to
widely exploited in several industrial applications. Particularly, measure electrical and meteorological data from the PV array.
this approach benefits from the good ability of the GLR chart
 Fig. 2. Block chart of one PV sub-system.

TABLE III Rs and Rsh , which have a core role in in the prediction quality
E SSENTIAL PROPERTIES OF THE USED MONITORING SENSORS . of the SDM.

Fig. 3. Representation of SDM representing a PV cell.

Generally speaking, the designed anomaly detection proce-

dure is accomplished by doing the next steps: (1) compute
the parameters of SDM, (2) build and test a simulation
model of the entire PV system using PSIMT M /MatlabT M
co-simulation [30], and lastly evaluate residuals using the
This study focuses on designing a monitoring chart for GLR to uncover abnormal condition in the PV system under
uncovering anomalies in the DC part of a PV plant through monitoring.
the use of tilted irradiance and ambient temperature as input IV. GLR- BASED FAULT DETECTOR FOR PV SYSTEM
variables. MONITORING
III. PV ARRAY MODELING Here, the GLR detector and its application in monitoring
A PV solar cell is usually modeled using a single diode PV systems are briefed.
model (SDM) (Figure 3), and the equation representing the
A. GLR approach
I-V characteristics is [29]
The GLR test is an efficient statistical detector to decide be-
tween two composite hypotheses. It has been widely applied to
anomaly detection in time-series data in different. applications,
including air quality [25], strain detection in hospitals [16] ,
traffic congestion detection [26], and train positioning [24].
where I0 refers to the dark saturation current, q denotes the Importantly, a GLR-based chart can separate the null hypoth-
electronic charge, kB refers to the constant of Boltzmann esis H0 from the alternative one H1 based on only measured
and n denotes the diode’s ideality factor. There are fine five data. Let us consider the vector Y = [y1 , y2 , . . . , yn ] ∈ Rn is
parameters with unknown values in Equation (1), Iph , I0 , n, created using one of the following Gaussian distributions:
 The estimation of θ is expressed as θb = arg minkY −θk22 = Y .
θ
Concerning the likelihood ratio, L(Y ), it is re-written as,
1 
L(Y ) = 2 kY k22 .

(4)
σ
It is worth noticing that the detection threshold h(α), is defined
such that the desired probability of false alarms is a priori
fixed.
P0 (L(Y ) ≥ h(α)) = α. (5)

Overall, if the GLR statistic is within the control limit, then

the supervised PV array is operating normally; otherwise, there
is a potential anomaly.
B. Monitoring PV systems using GLR-based approach
Anomaly detection based on the GLR approach is realized
into three stages as summarized in Figure 5. First, the model
of the PV array is built using the training. Here, the ABC
method is utilized to find the optimized electrical parameters
of the SDM. Second, for testing unseen data, the build
model is applied for estimating the MPP power, current, and
voltage using measured temperature and irradiance. Residuals,
which are the dissimilarity between predicted values of MPP
variables from the model and the measured ones, are used to
detect anomalies. Finally, the efficiency of the PV array is
checked by applying the GLR scheme on the residuals from
Fig. 4. Principal steps to model a PV plant [30]. the model (Figure 5).

H0 = {Y ∼ N (0, σ 2 In )}

(1)
H1 = {Y ∼ N (θ 6= 0, σ 2 In )}.

Here, θ denotes the anomaly level, and σ 2 > 0 denotes the

variance.
Here, H0 denotes the null hypothesis that represents the
anomaly-free situation, and the alternative hypothesis, H1 , Fig. 5. Diagram of the proposed monitoring procedure.
describes the faulty case. The GLR scheme separates between
H0 and H1 based on the comparison of L(Y ) to the control V. R ESULTS AND DISCUSSION
limit, h(α). Now, the proposed anomaly detection approach will be
( sup fθ (Y ) applied to actual data given in Section II.
θ∈Rn
δ(Y ) = H 0 if L(Y ) = 2 log fθ=0 (Y ) < h(α) (2) Firstly, by using the best-so-far ABC optimization algorithm
H1 else. we identify the parameters of the PV module based on the
(I-V) curve measurements (Table IV). Next, we used the
Consequently, the GLR detection statistic, L(Y ), is defined computed parameters for predicting (I-V) curve via the SDM
as (Eq. (1)).
kY − θk22 kY k22
    
L(Y ) = 2 log sup exp − / exp − , TABLE IV
θ 2σ 2 2σ 2 E STIMATED PARAMETERS OF THE SDM.
1
exp − 2σ1 2 k Y − θ k22 is the pdf

where fθ (Y ) = n
(2π) 2 σ n
of Y . Then, (3) can be expressed as
1
  The built model is tested based on three-day profiles of
2 2
L(Y ) = 2
min kY − θk2 + kY k2 (3) temperature and irradiance as input to the model. Figure 6
σ θ
1 n o indicates that peak power measurements (Pmppmeas ) fit with
= b 2 + kY k2 .
kY − θk the estimated one (Pmppsim ).
2 2 2
σ
 Fig. 6. Gathered and predicted MPP power.

Fig. 8. Result of the GLR approach in the case of Fault #1.

A. Detection results
Herein, four types of scenarios are considered to evaluate
the abilities of the GLR-based approach: two short-circuit
faults, an open-circuit fault, and four PV modules temporary
shaded (Figure 7).

Fig. 9. Result of the GLR approach in the case of Fault #2.

Fig. 10. Result of the GLR approach in the case of Fault #3.

3) Example with temporary shading: In this scenario, we

test the capability of the GLR monitoring scheme when
Fig. 7. Four failures tested in this study. temporary shading happened (Figure 7, #4) in the time interval
from 150 to 250. Figure 11 shows the detection results of the
1) Short circuit faults: The results of the GLR approach GLR test. This result confirms the capacity of the GLR-based
based MPP power residuals when a short-circuit fault in five approach to detect temporary shading.
modules occurred are demonstrated in Figure 8. As expected,
one can see that the GLR statistic exceeded the control limit
indicating the presence of an anomaly.
Also, the performance of the proposed approach is verified
when short-circuit fault occurred in three modules as shown
in Figure 7 (Fault #2). The monitoring result in Figure 9 and
clearly confirms the capacity of the GLR approach to detect
this short-circuit fault.
2) Open circuit fault: This scenario aims to test the GLR
chart in the case of open-circuit faults (see fault #4 in Figure 7)
that happened from time point 300 to 400. The detection
Fig. 11. Result of the GLR approach in the case of Fault #4.
results are displayed in Figure 10. As shown in Figure 10,
the GLR approach detect successfully this fault.
 VI. C ONCLUSION [12] Silei Chen, Xingwen Li, and Jiayu Xiong. Series arc fault identification
for photovoltaic system based on time-domain and time-frequency-
In practice, photovoltaic plants are frequently exposed to domain analysis. IEEE Journal of Photovoltaics, 7(4):1105–1114, 2017.
anomalies that can reduce their productivity and safety. The [13] R Hariharan, M Chakkarapani, G Saravana Ilango, and C Nagamani.
present work deals with the fault detection problem in the DC A method to detect photovoltaic array faults and partial shading in pv
systems. IEEE Journal of Photovoltaics, 6(5):1278–1285, 2016.
side of PV systems using a model-based approach. Basically, [14] Honglu Zhu, Lingxing Lu, Jianxi Yao, Songyuan Dai, and Yang Hu.
the proposed detection scheme admits two stages: first, a Fault diagnosis approach for photovoltaic arrays based on unsupervised
reference simulation model is constructed for mimicking the sample clustering and probabilistic neural network model. Solar Energy,
176:395–405, 2018.
nominal evolution of the inspected PV array, and it is used [15] Fouzi Harrou, Abdelkader Dairi, Bilal Taghezouit, and Ying Sun. An
for residuals generation to uncover potential faults. After that, unsupervised monitoring procedure for detecting anomalies in photo-
the GLR scheme is implemented to verify the residuals for voltaic systems using a one-class support vector machine. Solar Energy,
179:48–58, 2019.
uncovering potential faults. Successful detection results have [16] Fouzi Harrou, Farid Kadri, Ying Sun, and Sofiane Khadraoui. Mon-
been achieved when using databases generated from a 9.54 itoring patient flow in a hospital emergency department: Arma-
Kwp PV system. The results testify the sensitivity of the GLR- based nonparametric glrt scheme. Health Informatics Journal,
27(2):14604582211021649, 2021.
based scheme in detecting short-circuit, open-circuit faults, [17] Zhicong Chen, Fuchang Han, Lijun Wu, Jinling Yu, Shuying Cheng,
and temporary shading. Peijie Lin, and Huihuang Chen. Random forest based intelligent fault
diagnosis for PV arrays using array voltage and string currents. Energy
ACKNOWLEDGEMENT Conversion and Management, 178:250–264, 2018.
[18] Rabah Benkercha and Samir Moulahoum. Fault detection and diagnosis
based on c4. 5 decision tree algorithm for grid connected pv system.
Solar Energy, 173:610–634, 2018.
[19] Zhicong Chen, Lijun Wu, Shuying Cheng, Peijie Lin, Yue Wu, and
Wencheng Lin. Intelligent fault diagnosis of photovoltaic arrays based
on optimized kernel extreme learning machine and iv characteristics.
Applied energy, 204:912–931, 2017.
[20] Jose A Romagnoli and Ahmet Palazoglu. Introduction to process control.
CRC press, 2005.
[21] Farid Kadri, Fouzi Harrou, Sondès Chaabane, Ying Sun, and Christian
Tahon. Seasonal ARMA-based SPC charts for anomaly detection: Appli-
cation to emergency department systems. Neurocomputing, 173:2102–
2114, 2016.
[22] Zeroual Abdelhafid, Fouzi Harrou, and Ying Sun. Road traffic den-
R EFERENCES sity estimation and congestion detection with a hybrid observer-based
[1] B. Brooks. The bakersfield fire: A lesson in ground-fault protection. strategy. Sustainable Cities and Society, 46, 2019.
SolarPro, 4(2), 2011. [23] Fouzi Harrou, Ying Sun, Muddu Madakyaru, and Benamar Bouyedou.
[2] Mohammed Khorshed Alam, Faisal Khan, Jay Johnson, and Jack Flicker. An improved multivariate chart using partial least squares with continu-
A comprehensive review of catastrophic faults in PV arrays: types, ous ranked probability score. IEEE Sensors Journal, 18(16):6715–6726,
detection, and mitigation techniques. IEEE Journal of Photovoltaics, 2018.
5(3):982–997, 2015. [24] F. Harrou, L. Fillatre, and I. Nikiforov. Anomaly detection/detectability
[3] Dhanup S Pillai, Frede Blaabjerg, and Natarajan Rajasekar. A compar- for a linear model with a bounded nuisance parameter. Annual Reviews
ative evaluation of advanced fault detection approaches for pv systems. in Control, 38(1):32–44, 2014.
IEEE Journal of Photovoltaics, 9(2):513–527, 2019. [25] Fouzi Harrou, Lionel Fillatre, Michel Bobbia, and Igor Nikiforov. Sta-
[4] Elyes Garoudja, Fouzi Harrou, Ying Sun, Kamel Kara, Aissa Chouder, tistical detection of abnormal ozone measurements based on constrained
and Santiago Silvestre. Statistical fault detection in photovoltaic systems. generalized likelihood ratio test. In 52nd IEEE Conference on Decision
Solar Energy, 150:485–499, 2017. and Control, pages 4997–5002. IEEE, 2013.
[5] Bilal Taghezouit, Fouzi Harrou, Ying Sun, Amar Hadj Arab, and [26] Fouzi Harrou, Abdelhafid Zeroual, and Ying Sun. Traffic congestion
Cherif Larbes. Multivariate statistical monitoring of photovoltaic plant detection based on hybrid observer and glr test. In 2018 Annual
operation. Energy Conversion and Management, 205:112317, 2020. American Control Conference (ACC), pages 604–609. IEEE, 2018.
[6] Siva Ramakrishna Madeti and SN Singh. A comprehensive study on [27] Xiaoxia Li, Qiang Yang, Zhuo Lou, and Wenjun Yan. Deep learning
different types of faults and detection techniques for solar photovoltaic based module defect analysis for large-scale photovoltaic farms. IEEE
system. Solar Energy, 158:161–185, 2017. Transactions on Energy Conversion, 34(1):520–529, 2018.
[7] Bilal Taghezouit, Fouzi Harrou, Ying Sun, Amar Hadj Arab, and [28] Albert Yaw Appiah, Xinghua Zhang, Ben Beklisi Kwame Ayawli,
Cherif Larbes. A simple and effective detection strategy using double and Frimpong Kyeremeh. Long short-term memory networks based
exponential scheme for photovoltaic systems monitoring. Solar Energy, automatic feature extraction for photovoltaic array fault diagnosis. IEEE
214:337–354, 2021. Access, 7:30089–30101, 2019.
[8] Masaki Miwa, Sanshiro Yamanaka, Hajime Kawamura, Hideyuki Ohno, [29] Wei Zhou, Hongxing Yang, and Zhaohong Fang. A novel model for pho-
and Hideaki Kawamura. Diagnosis of a power output lowering of tovoltaic array performance prediction. Applied energy, 84(12):1187–
pv array with a (-di/dv)-v characteristic. In 2006 IEEE 4th World 1198, 2007.
Conference on Photovoltaic Energy Conference, volume 2, pages 2442– [30] Fouzi Harrou, Bilal Taghezouit, and Ying Sun. Robust and flexible
2445. IEEE, 2006. strategy for fault detection in grid-connected photovoltaic systems.
[9] Yihua Hu, Bin Gao, Xueguan Song, Gui Yun Tian, Kongjing Li, and Energy Conversion and Management, 180:1153–1166, 2019.
Xiangning He. Photovoltaic fault detection using a parameter based
model. Solar Energy, 96:96–102, 2013.
[10] W Chine, A Mellit, Vanni Lughi, A Malek, Giorgio Sulligoi, and
A Massi Pavan. A novel fault diagnosis technique for photovoltaic
systems based on artificial neural networks. Renewable Energy, 90:501–
512, 2016.
[11] Chuxuan He, Longhua Mu, and Yijian Wang. The detection of parallel
arc fault in photovoltaic systems based on a mixed criterion. IEEE
Journal of Photovoltaics, 7(6):1717–1724, 2017.