An intelligent method for fault diagnosis in photovoltaic systems

Full text

An Intelligent Method for Fault Diagnosis in
Photovoltaic Systems
Department of electrical engineering
Mohammed V University in Rabat
Ecole Mohammadia d'Ingenieurs
Rabat, Morocco
yassinechouay@research.emi.ac.ma
Mohammed Ouassaid
Department of electrical engineering
Mohammed V University in Rabat
Ecole Mohammadia d'Ingenieurs
Rabat, Morocco
ouassaid@emi.ac.ma
Abstract—This paper presents a diagnostic method based on
I–V characteristic analysis and Artificial Neural Networks
(ANN). A number of attributes are estimated using a simulation
model based on a set of working conditions such as solar
irradiance and module's temperature. The estimated attributes
are then compared with those obtained from the real PV array
measurements, which lead to the detection of possible faulty
operating conditions. After detecting the presence of possible
fault, tow algorithms are developed in order to isolate and
identify eight different types of faults. The simulation results
show that the proposed method can detect and classify the
different faults occurring in a PV array with a high accuracy.
Keywords—Photovoltaic, Fault detection, Fault diagnosis,
Fault localization, ANN.
I. INTRODUCTION
The use of photovoltaic (PV) systems is now increasing
rapidly all over the world. With this increase in PV systems,
the number of problems and failures is also increasing. The
importance of service and maintenance is therefore growing.
The need for optimal functioning of PV systems is linked to
the recent interest in fault diagnosis techniques, which are
today more and more developed in the literature.
Nowadays, many diagnostic techniques are developed for
possible fault detection in PV systems. Generally, fault
detection techniques for PV systems can be grouped as visual
(browning, discoloration, surface soiling, and delamination),
thermal (hot spot), and electrical (transmittance line diagnosis,
dark/illuminated current-voltage measurement, and Radio
frequency measurement) [1]. In this paper, we used an
electrical method to detect the investigated faults. Electrical
methods for fault detection in photovoltaic systems can be
grouped as:
- Methods that do not require climate data (solar irradiance
and modules temperature): in [2], the Earth Capacitance
Measurement (ECM) method was used as an electrical
technique for detecting where a photovoltaic module in a
string has been disconnected. In [3], The Time-Domain
Reflectometry (TDR) technique was used to detect the
disconnection of a PV string as well as the impedance
change due to degradation.
- Methods based on the evaluation of the current–voltage
characteristic (I–V characteristic): in [4], the (dI/dV)-V
characteristic was used to detect the partial shadow fault
in a PV array. In [5], a method based on the evaluation of
some current and voltage indicators was introduced as an
automatic fault detection for Grid-Connected Photovoltaic
(GCPV) systems;
- Methods based on the Maximum Power Point Tracking
(MPPT) approach: an automatic supervision and fault
detection procedure based on the analysis of the power
losses was proposed in [6]. This method leads to the
identification of three groups of faults (faulty string,
faulty module, and a group of different faults including
partial shadow, ageing, and MPPT failure). The technique
developed in [7] is based on the comparison between the
simulated and the measured string powers. This method is
even able to determine the number of open and short-
circuited PV modules in a string;
- Methods that use the Artificial Intelligence (AI): in [8] a
learning technique based on Expert Systems was
developed to identify two types of fault (shading effect
and inverter's failure). In [9], a method for the
identification of short-circuited PV modules is presented.
An ANN is used in order to classify different types of
faults occurring in a PV array in [10]. The ANN takes as
inputs the voltage and the current at maximum power
point, as well as the temperature of the PV modules. The
method presented in [11] uses tow algorithms. The first
algorithm is based on a signal threshold approach that
isolates faults having different combinations of attributes.
The second is an ANN used to distinguish between faults
having the same signature. This technique is able to detect
eight different types of faults. Different methods based on
the Takagie Sugeno Kahn Fuzzy Rule (TSKFRBS) have
been described in [12, 13].
In this paper, an automatic fault diagnosis method is
presented. It allows the detection and the localization of faults
occurring in: PV cells, PV modules, PV strings, and bypass-
diodes. The proposed technique is based on the analysis of a
set of attributes of the I–V characteristic. To expand the
number of detected faults the analysis is performed using two
different Algorithms:
978-1-5386-1516-4/17/$31.00 ©2017 IEEE
3rd International Conference on Electrical and Information Technologies ICEIT’2017
Yassine Chouay *

- Algorithm 1 uses the signal threshold approach based on
the I–V characteristic analysis, which identifies faults
affecting the I–V characteristic differently.
- Algorithm 2 consists of an ANN-based approach to
identify the faults that are characterized by the same
effect on the I–V characteristic.
II. FAULTS INVESTIGATED BY THE PROPOSED TECHNIQUE
Generally, faults occurring in a PV system are mainly
related to the PV array, the storage system, the inverter, and
the grid. The work presented in this paper intends to detect the
faults occurring in the PV array (DC side). According to Table
I, eight different faults are detected. These types of faults are
usually connected to: the failure of a solar cell or a PV
module, the degradation effect, a line disconnection, corrosion
and manufacturing defects, the presence of shadow, the effect
of soiling, and etc.
III. THE PROPOSED FAULT DIAGNOSIS TECHNIQUE
As shown in Fig. 1, the difference between the simulated
and the measured PV system output power (ΔP) is compared
with a threshold (Th) to decide whether the system is
functioning in faulty mode. Then, in order to identify and
determine the fault type, the main attributes in the I–V
characteristic of each string forming the PV array are analyzed.
The schematic of the developed fault diagnosis system is
depicted in Fig. 2.
A. Attributes identification
Before developing the fault detection algorithm, it’s
necessary to understand the effect of each fault on the normal
functioning. Therefore, a number of simulations have been
carried out considering both normal operation and different
fault conditions. The changes that affect the I–V characteristic
of a PV string are presented in Fig. 3. The analysis of these
simulations led to the identification of five situations of
changes:
- A reduction in the short circuit current (FIsc);
- A reduction in the open circuit voltage (FVoc);
- An increase or a reduction in the output current (FIm);
- An increase or a reduction in the output voltage (FVm);
- An increase in the number of MPPs in the I–V
characteristic (N). In case of partial shading the difference
between the simulated and the measured number of MPPs
(ΔN) is not null.
B. Algorithm 1 (Signal threshold based approach)
The first algorithm isolates the faults that have a different
combination of attributes. Firstly, the relative differences
between the simulated and measured PV string attributes are
compared with some predefined thresholds. The simulated I–
V characteristics are obtained from a Simulink model based
on the values of solar irradiance and module temperature.
The thresholds are calculated according to the maximum
TABLE I. DIFFERENT TYPE OF FAULTS OCCURRING IN A PV ARRAY
Locati-
on
Name
Symb-
ol
Module
Short circuit fault in any bypass diode, cell or
module.
F1
Inversed bypass diode, cell or module.
F2
Shunted bypass diode, cell or module.
F3
Open circuit fault in any cell or module.
F4
Cables
Connection resistance between PV modules.
F5
Array
Shading effect in the modules with normal
operation of other components of PV string.
F6
Shadow effect in a group of cells equipped by an
open-circuited bypass diode.
F7
Shadow effect in a group of modules equipped by
a connection fault.
F8
Fig. 1. Flow chart of the proposed fault detection technique.
Fig. 2. Schematic of the fault diagnosis technique.
Data from the
DAQ (G, T)
PV module parameters,
number of strings and
modules.
Electrical data from
DAQ (I
maes
, Vmaes)
Simulated PV model
(Isim, Vsim)
∆P=MPPsim-MPPmaes
|∆P|>Th
Normal
functioning
Start faults
isolation algorithms
No
Yes
Start
G
T
Ii(string)
P(array)
Non-faulty
PV array
Attributes calculation
Power
comparison
F8
F4
F6
F7
F5
F3
F2
F1
Real PV
array
D.A.Q
P(array)
ANN Algorithm
Vi(string)
Ii(string)
Vi(string)
3rd International Conference on Electrical and Information Technologies ICEIT’2017

error introduced by the measurement noise and the model
uncertainty:
- The sensors used to test the proposed diagnostic system
are supposed within the specifications required by the IEC
61724 standard [14] indicating a relative error of 1%, 1%,
and 2% for current, voltage, and power measurement,
respectively.
- The model uncertainty is related to sensor noise and the
manufacturing tolerance. According to [15], the
maximum error introduced by this uncertainty is
calculated by adding a dispersion parameter to the
simulation model parameters. The obtained relative errors
associated with current, voltage, and power are equal to
3.0%, 2.9%, and 5.9%, respectively.
Thresholds (Th, TCand TV) for the first algorithm are
obtained by adding errors introduced by both model and
measurement uncertainties:
(2% 5.9%)
(1% 3%)
(1% 2.9% )
mpp
Csc
Voc
Th P
TI
TV
u
u 
u 
Fig. 4 illustrates the flowchart of Algorithm 1 that allows
the isolation of six different situations: F4, F5, F6, F7, F8, and
a group of faults gathering F1, F2, F3, and F5.
C. Algorithm 2 (ANN based approach)
As shown in Fig. 4, the first algorithm cannot differentiate
between a group of faults including F1, F2, F3 and F5 because
they have the same fault signature. Furthermore, there are
some differences in the symptoms of the I–V characteristic
under these faults at the same conditions. Therefore, in order
to isolate these faults, an ANN model has been developed as
follows:
1) Selection of the input and output variables
The ANN architecture consists of three neurons in the input
layer corresponding to the ratio between the measured and the
simulated values of the open circuit voltage (RVoc), the
maximum power point current (RIm), and the maximum power
point voltage (RVm). One neuron in the output layer
corresponding to the fault class.
2) Selection of the network structure
The network used in this work is a Multilayer Perceptron
(MLP). The MLP structure consists of one hidden layer
containing 40 nodes as represented in Fig. 5. The transfer
function used in this architecture is the logarithmic sigmoidal
function.
3) Network training and test
The network is trained with the Levenberg-Marquardt (LM)
algorithm. The training data was generated from a set of
simulations of both normal and faulty operations for the four
Fig. 3.
I–V characteristics of a PV string in normal and faulty operations
: (a).
without shading
effect. (b). with shading effect.
Fig. 4. Block diagram of Algorithm 1.
Calculation of the attributes
strings amplitude
|∆N|> 0
FIsc=Isc_sim
Open
circuit F4
F
Im
>TC
F
Isc
>TC
FVoc>TV
Shadow
effect with
faulty bypass
diode F7
Group of faults:
-
short-
circuited diodes
F1
-
Inversed diodes F2
-
Shunted diodes F3
-Connection fault F5
Connection
fault F5
F
Im
>TC
Shadow effect
with connection
fault F8
Partia l shadow
without any fault on
bypass diode F6
Yes
Yes
Yes
Yes
Yes
Yes
No
No
No
No
No
No
ANN Algorithm
F
Vm
>TV
Start
End
No
Yes
3rd International Conference on Electrical and Information Technologies ICEIT’2017

faults. 80% of the patterns have been used for the training,
while 20% have been used for testing and validating the
model.
IV. SIMULATION RESULTS AND DISCUSSION
In order to test the effectiveness of this intelligent method a
simulation of both normal and faulty operation was carried out
in Matlab/Simulink environment. The detection program takes
attributes from two simulated PV systems, the first represents
the real PV array (contain different faults) and the other
represents the normal functioning.
A. Performance of the proposed technique
1) Performance of Algorithm 1
In order to verify the performance of the first Algorithm, a
PV array formed by one string of five MSX60 series-
connected PV modules having the electrical characteristics
illustrated in Table II has been considered, and different cases
of study have been examined in STC:
- Case 1: The connection resistance between two modules
was increased by 20 Ω;
- Case 2: one module was 50% shaded and all bypass
diodes open-circuited in this module;
- Case 3: two modules were shaded (the shading proportion
was 50% for the first module and 25% for the second);
The calculated thresholds, attributes and the Algorithm 1
decision are reported in Table III.
The simulation results show that the first algorithm is able
to localize and identify correctly the faults in the examined PV
array.
2) Performance of Algorithm 2
The network was trained using a set of data generated from
the simulation results of the four investigated faults, and with
reference to Fig. 6. (a) the minimum Mean Square Errors
(MSE) achieved while training, testing and validating the
model are 0.013, 0.032 and 0.0057, respectively. In order to
examine the effectiveness of the proposed ANN-based
approach, Fig. 6. (b) represents the classification confusion
Fig. 5. Schematic of the used ANN architecture.
40 Neurons
Faults:
F1, F2, F3, or F4
RVoc
RIm
RVm
TABLE III. ATTRIBUTES AND THRESHOLDS FOR THE CONSIDERED CASE
STUDIES
Th
T
C
T
V
23.7
0.152
4.11
Case
∆
P
FIsc
FIm F
Voc
FVm
∆N
Decision
1
181.3
0.08
1.31
0
31.41
0
F5
2
126.9
1.69
1.61
0.89
8.32
0
F7
3
103
0
0.7
0
14.2
2
F6
Fig. 6. (a). Evolution of the MSE for the MLP network. (b). Classification confusion matrix for the MLP network
(b)
TABLE II. SOLAR PV PANEL MSX60 PARAMET ERS AT 25◦C,
AM1.5,
AND 1000 W/M2.
Parameter
Value
Immp
3.5 A
Vmmp
17.1 V
Pmax
59.85 W
ISC
3.8 A
VOC
21.1 V
NS
36
A
1.3
(a)
3rd International Conference on Electrical and Information Technologies ICEIT’2017

matrix for the four considered faults. The green cells represent
the percentage of fault correctly classified, while the red ones
represent the wrong classifications.
The classification confusion matrix shows that 94.1% of the
training and testing samples were correctly classified, while
5.9% were not correctly classified. The false classifications
between F1 and F2 happened when the number of inversed
cells or bypass diodes was very low. The false classifications
between F3 and F5 occurred when the connected and shunt
resistances, had either a very low or a very high value.
B. Comparison
In order to confirm the effectiveness of the proposed
method, this one is compared with the method presented in [6]
that uses the power losses to detect and distinguish between
faults. The comparison between the attributes and the results
given by the two methods is presented in Table IV, for the
four different case studies. As can be noticed, the power losses
analysis method cannot identify the exact fault type.
V. CONCLUSION
This paper has proposed a fault diagnosis method is
conceived to detect and isolate eight faults occurring in a PV
array. Two different algorithms allow the isolation and
identification of faults. The first algorithm detects faults
having a different combination of attributes, whereas the
second use an ANN to discriminate between faults even if
they have the same signature. The simulation results
confirmed the performance of this technique by testing the
effectiveness of each algorithm individually. The results have
shown also that the proposed method detects and identify the
investigated faults with a high accuracy compared to other
methods.
REFERENCES
[1] B. Ando, S. Baglio, A. Pistorio, G.M. Tina, C. Ventura, “Sentinella:
smart monitoring of photovoltaic systems at panel level,”IEEE Trans.
Instrum. Meas. vol. 64, august 2015, 2199.
[2] T. Takashima, J. Yamaguchi, K. Otani, T. Oozeki, K. Kato, M. Ishida,
“Experimental studies of fault location in PV module strings,”Sol.
Energy Mater Sol. Cells 93, 2009, pp. 1079–1082 .
[3] L. Schirone, F.P. Califano, M. Pastena, “Fault detection in a
photovoltaic plant by time domain reflectometry,” Prog Photovolt Res
Appl, vol. 2, 1994, pp. 35–44.
[4] M. Miwa, S. Yamanaka, H. Kawamura, H. Ohno, H. Kawamura,
“Diagnosis of a power output lowering of PV array with a (-dI/dV)–V
characteristic,” in: Proceeding of IEEE 4th World Conference on
Photovoltaic Energy Conversion, Waikoloa, HI, 2006, pp. 2442–2445.
[5] S. Silvestre, M.A. da Silva, A. Chouder, D. Guasch, E. Karatepe, “New
procedure for fault detection in grid connected PV systems based on the
evaluation of current and voltage indicators,”Energy Convers. Manag.
vol. 86, 2014, pp. 241–249.
[6] S. Silvestre, A. Chouder, E. Karatepe, “Automatic supervision and fault
detection of PV systems based on power losses analysis,”Energy
Convers. Manag. vol. 51, 2010, pp. 1929–1937.
[7] N. Gokmen, E. Karatepe, B. Celik, S. Silvestre, “Simple diagnostic
approach for determining of faulted PV modules in string based PV
arrays,”Sol. Energy, vol. 86, 2012, pp. 3364–3377.
[8] Y. Yagi, H. Kishi, R. Hagihara, T. Tanaka, S. Kozuma, T. Ishida, M.
Waki, M. Tanaka, S. Kiyama, “Diagnostic technology and an expert
system for photovoltaic systems using the learning method,”Sol. Energy
Mat. Sol. Cells, vol. 75, 2003, pp. 655–663.
[9] S. Syafaruddin, E. Karatepe, T. Hiyama, “Controlling of artificial neural
network for fault diagnosis of photovoltaic array,”in: Proceedings of the
16th International Conference on Intelligent System Application to
Power Systems (ISAP), Greece, 2011, pp. 1–6.
[10] Z. Li, Y. Wang, D. Zhou, C. Wu, “An intelligent method for fault
diagnosis in photovoltaic array,”ICSC Part II CCIS, vol. 327, 2012, pp.
10–16.
[11] W. Chine a, A. Mellit, V. Lughi, A. Malek d, G. Sulligoi, A. Massi
Pavan, “A novel fault diagnosis technique for photovoltaic systems
based on artificial neural networks,”Renewable Energy, vol. 90, 2016,
pp.501–512.
[12] P. Ducange, M. Fazzolari, B. Lazzerini, F. Marcelloni, “An intelligent
system for detecting faults in photovoltaic fields”, in: Proceedings of the
11th International Conference on Intelligent Systems Design and
Applications IEEE, Cordoba, Spain, 2011, pp. 1341–1346.
[13] L. Bonsignore, M. Davarifar, A. Rabhib, G.M. Tinaa, A. Elhajjaji,
“Neuro-Fuzzy fault detection method for photovoltaic systems, Energy
Procedia,” vol. 62, 2014, pp. 431–441.
[14] W. Xiao, W.G. Dunford, A. Capel, “A novel modelling method for
photovoltaic cells”, in: Proceedings of the IEEE 35th, Annual Power
Electronics Specialists Conference (PESC), Germany, 2004, pp. 1950–
956.
[15] L. Bun, “Détection et localisation de défauts pour un système PV,” PhD
thesis, University of Grenoble, 2011.
TABLE IV. COMPARATIVE RESULTS BETWEEN THE PROPOSED METHOD AND THE POWER LOSSES ANALYSIS METHOD.
Case
Power losses analysis method
The proposed technique
Parameters
Decision
Parameters
Decision
Two shaded modules in the string.
R
C
=1, R
V
=1.53
Faulty module.
Algorithme 1:
∆N=1, FIm=0.01, TC=0.16
Partial shadow without any
fault on bypass diode.
Connection resistance between two
modules is increased.
R
C
=1.47, R
V
=1.54
Group of faults.
Algorithme 1:
∆N=0, FIsc =1.6, TC=0.16
Connectic fault.
Two short-circuited modules in the
string
.
R
C
=1, R
V
=1.5
Faulty module.
Algorithme 1:
∆N=0, FIsc =0, FIm =0.02, TC= 0.16
Group of faults.
Algorithme 2:
RVoc=0.66, RVm=0.66, RIm=0.99
Cells, diodes, or modules are
short circuited.
String with one inversed module.
R
C
=1, R
V
=1.59,
Faulty module.
Algorithme 1:
∆
N=0, FIsc=0, FIm=0.2,
TC=0.16, FVoc=65.8, TV=7.68
Group of faults.
Algorithme 2:
RVoc=0.66, RVm=0.62, RIm=0.97
Cells, diodes, or modules are
inversed.
RC: Current ratio, RV
: Voltage ratio.
3rd International Conference on Electrical and Information Technologies ICEIT’2017