EEEP 2019 IOP Publishing

IOP Conf. Series: Earth and Environmental Science 467 (2020) 012083 doi:10.1088/1755-1315/467/1/012083

Fault diagnosis of PV array using adaptive network based

fuzzy inference system

Yutao Gan, Zhicong Chen1, Lijun Wu, Shuying Cheng and Peijie Lin
College of Physics and Information Engineering, Fuzhou University, 2 XueYuan Road,
350116 Fuzhou, China
1
Emai: [email protected]

Abstract. A new online intelligent fault diagnosis method is proposed for PV arrays in this
paper to improve the reliability and efficiency of PV systems. Firstly, a new seven-dimensional
fault feature vector is extracted from the raw data of dynamic operating points of PV arrays
including operating voltage, current, irradiance and temperature. Secondly, an optimized
adaptive network based fuzzy inference system (ANFIS) is proposed as the fault diagnosis
model. Lastly, the feasibility and superiority of the proposed ANFIS based fault diagnosis
model are tested by both Simulink based simulation and real fault experiments on a laboratory
PV system. Experimental results validate that the proposed ANFIS based method achieves a
high performance and is superior to conventional back-propagation neural network (BPNN)
based methods. The overall accuracy of the ANFIS based fault diagnosis model on the
simulation and experimental dataset is 99.9% and over 97% respectively.

1. Introduction
To address the increasingly severe energy crisis, environment pollution and climate change, the
installed capacity of PV systems is rapidly increasing in recent years. As the core of PV systems, PV
arrays are susceptible to damage due to the complex installation environment, which may affect the
power generation efficiency of PV array. Consequently, fault diagnosis of PV arrays is significant to
improve the reliability and efficiency of PV systems [1]. At present, there are many kinds of fault
diagnosis of PV arrays. From the perspective of data sources, they can be categorized into two types,
the methods based on infrared thermal images and the methods based on output characteristics of PV
arrays [1, 2]. The former can find the defects of the photovoltaic panel earlier when there is no major
defect in the output characteristics of the PV array. The operating cost and data collection cost of this
method are very high, which is not economic to wide-scale promotion. Data sources for diagnostic
methods based on the output characteristics of PV arrays include I-V curves[3, 4], time series graph
[5], and dynamic operating point voltage and current [1, 6]. The method based on I-V curve can
collect more PV array information and detect PV faults more comprehensively [4], but it needs offline
operation, which will interrupt the normal operation of PV systems. However, the fault diagnosis
method based on dynamic operating points does not need to interrupt the normal operation of the PV
systems. The time series graph signal has the advantage of high signal matching when the fault occurs
but can only find the change of the PV array from normal working conditions to faulty working
conditions. The dynamic operating points based method has a wider application scenario, because the
dynamic operating point based fault diagnosis method can diagnose the normal state and the stable

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
EEEP 2019 IOP Publishing
IOP Conf. Series: Earth and Environmental Science 467 (2020) 012083 doi:10.1088/1755-1315/467/1/012083

fault state, and the time series graph signal based method can't do anything for the fault that has
already become a steady state fault, because only the transients at the time of the fault can be detected.
In recent years, as the core of artificial intelligence, machine learning has been widely used in data
processing due to the advantages of being able to process large amounts of data quickly. It is applied
in the photovoltaic fault diagnosis method and is trained with labelled data, which has a very high
accuracy rate of fault diagnosis. In machine learning, artificial neural networks (ANN) have strong
self-learning ability and memory ability, and have strong mapping ability to the nonlinear output of
PV arrays [7]. However, the shortcomings of neural networks are the inability to explain their own
reasoning ability and reasoning basis, need a lot of data training, which will cause a big change in the
results when the data changes slightly. Fuzzy logic inference systems (FIS) can apply fuzzy sets and
fuzzy rules for reasoning. However, the FIS has no self-learning ability, and it is necessary to adopt
the rules of expert experience. The formulation of the rules of expert experience greatly limits the
application of the FIS, which greatly increases the difficulty of the work. The adaptive network based
fuzzy inference system (ANFIS) combines the advantages of both the FIS and ANN, which does not
require a large amount of training data and features fast training speed, fuzzy reasoning, and strong
anti-interference ability which it is very suitable for fault diagnosis of nonlinear systems [8]. The main
contribution of this study is to propose a new fault diagnosis method for PV array, which is based on
the ANFIS and the dynamic operating point current together with the temperature and irradiance of the
PV array. The proposed fault diagnosis method can continuously detect and diagnose the condition of
the PV array without interrupting the working state of the PV arrays and the proposed algorithm model
has better generalization performance and anti-interference performance.

2. Adaptive network based on fuzzy inference system for fault diagnosis of PV array
This section details the ANFIS based diagnosis model for detecting several common faults of PV
arrays, including line-line fault, open-circuit fault, partial shadow fault and degradation fault. Firstly, a
seven dimensional fault feature is extracted from the raw data as the input of the model. Secondly, the
ANFIS model structure is detailed. Lastly, the ANFIS based model is trained and tested by the K-fold
cross-validation method and the ergodic method.

2.1. Extraction of fault features

The proposed ANFIS based fault diagnosis is based on the raw monitored data of the operating point,
including the operating voltage of PV array, operating current of each PV string, ambient temperature,
and ambient irradiance. In order to establish high performance diagnosis model, the fault features
should be carefully designed and extracted from the raw data. In this study, as defined in Eqs. (1) to
(7), seven fault features are proposed, including the normalized PV array voltage (Vn), array current
(In), array power (Pn), slope of the operating point (Sn), current dispersion rate (Cn), ambient
temperature (Tn), and incident irradiance (Gn). In order to eliminate the influence of the environment
condition, these seven fault features are all normalized by the corresponding simulation data.

(1)
(2)
(3)
(4)
(5)
(6)
(7)
* , ( )- ⁄ ( ⁄ )+ (8)
, ( )- (9)

2
EEEP 2019 IOP Publishing
IOP Conf. Series: Earth and Environmental Science 467 (2020) 012083 doi:10.1088/1755-1315/467/1/012083

where, Vmpp is the monitored operating array voltage and Vop is the corresponding therectical value
defined in Eq. (8); Impp is the monitored operating array current and Iop is the corresponding therectical
value defined in Eq. (9); μ is the mean of the string current, and σ is the standard deviation of string
current; Ga is the monitored incident irradiance, and Gstc is the irradiance under standard test condition
(STC) , i.e., 1000 W/m2; Ta is the monitored ambient temperature, and Tstc is the temperature under
STC, i.e., 25℃; n of the ideal factor of the solar cell in the PV modules; Ns is the number of solar
cells in series in a PV string; Np is the number of solar cell strings in parallel in the PV array; α is the
short-circuit current temperature coefficient, and β is the open circuit voltage temperature coefficient;
Ut is the thermal voltage; Vstc and Istc are the maximum power point (MPP) voltage and current in the
STC condition.

2.2. Adaptive network based fuzzy inference system based fault detection and diagnosis model
Using the features selected in section 2.1 as the input of fault diagnosis model, the ANFIS based fault
diagnosis model is established for photovoltaic fault diagnosis. The model has seven inputs and one
output, which can be used to predict the class of the unlabelled data. The brief process for establishing
the fault diagnosis model is illustrated in Figure 1.

Simulate PV array faults and measure X2 X1

the Impp, Vmpp , irradiance , temperature
and PV model parameters
Layer1

Extract and normalized the feature V n,

In, Pn, Sn, Cx, Gn, Tn W2 W1 Layer2

M2 M1 Layer3
Randomly divide the samples into
training and testing samples X1 X1
f2 f1 Layer4
X2 X2
Using the optimized ANFIS algorithm to
Layer5
train and obtain the fault diagnosis
model output

Figure 1. Brief flowchart of the ANFIS Figure 2. Structure of adaptive network based
based fault diagnostic model. fuzzy inference system.

The ANFIS is a fuzzy Sugeno model put in the framework of adaptive systems to facilitate learning
and adaptation [9], which consists of 5 layers in total as shown in Figure 2. The output of the nodes in
the first layer is:
( ) (10)
( ) (11)
μAi is fuzzy membership function, taking the Gaussian membership function as an example:
, ( ) ⁄ - (12)
where ai and bi are the parameters of the membership function, governing the gauss shaped
functions accordingly, i represents the number of membership function of each inputs.
The output of the nodes in the second layer is:
( ) ( ) (13)
The output of each node represents a regular excitation strength.
The output of node in the third layer is:
( ) (14)

3
EEEP 2019 IOP Publishing
IOP Conf. Series: Earth and Environmental Science 467 (2020) 012083 doi:10.1088/1755-1315/467/1/012083

Each node represents the intensity of the stimulus normalization.

The output of node in the fourth layer is:
( ) (15)
Each node indicates the output of each rule. pi, qi, and ri are the design parameters that are
determined during the training process.
The fifth layer calculates the sum of the fourth layer output nodes.

2.3. Training of the fault diagnosis model

Validated
ANFIS model
Valid- and select
set the optimal
ANFIS Optimal Training
Final
K-fold model parameter the
Train model
cross of ANFIS ANFIS
samples ANFIS
validation Travering the model model
number and the
Train- type of
set membership
functions of the
ANFIS model

Figure 3. Flowchart of parameter optimization for the ANFIS model.

Since the adaptive network based fuzzy inference system model for PV array fault diagnosis
belongs to a supervised self-learning network, the main parameters are the number and the type of
membership functions. The procedure to train the ANFIS model is shown as below and illustrated in
Figure 3:
Step 1: The training set is further randomly divided into a training subset and a verification subset
by K-fold cross-randomization;
Step 2: Traverse the number and the type of membership function of the ANFIS model in a
specified range, so as to establish different ANFIS model;
Step 3: Verify the accuracy of the trained models with the verification subset, and select the model
with the highest accuracy to determine the number and the type of membership functions;
Step 4: Use the training subset to train the ANFIS model with optimal parameters to obtain the
final ANFIS fault diagnosis model.

3. Simulation study

Figure 4. PV array simulation with fault condition.

4
EEEP 2019 IOP Publishing
IOP Conf. Series: Earth and Environmental Science 467 (2020) 012083 doi:10.1088/1755-1315/467/1/012083

To verify the proposed fault diagnosis method, as illustrated in Figure 4, the normal status and six
types of common faults of a laboratory PV array are simulated in Simulink to collect the data samples
with the irradiance varying from 100-975W/m2 (at the step of 25 W/m2) and with the temperature
varying from 25-70 °C (at the step of 2.5 °C). The simulated fault includes normal condition (N), line-
line fault of string level with one module difference in the same string (LL1) (simulated by connecting
a 0.001 ohms resistance between modules), line-line fault of array level with two modules difference
(LL2) at two difference PV string (simulated by connecting a 0.001 ohms resistance between modules),
open-circuit fault on one string (OC) (simulated by connecting a series resistor of 40000 ohms into the
negative of the PV array), partial shading fault (PS) (simulated through setting the irradiance gains
K=0.5), degradation fault of PV array level (D-A) (simulated by connecting 4 ohm resistor to the
negative output of the PV array), degradation fault of PV string level (D-S) (simulated by connecting 4
ohm resistors between a string of PV and the negative of the PV array). Based on the collected raw
data and the parameters of photovoltaic models, the fault features, including Vn, In, Pn, Sn, Cx, Gn, Tn ,
are extracted to form the data samples to train the photovoltaic fault diagnosis model. The number of
environment conditions for each fault is 684, and a data sample is collected for every environment
conditions. Therefore, the final dataset contains 4788 data samples in total. The dataset is randomly
divided into a training set and a test set. The training set accounts for 70% of the total data samples,
and the test set accounts for 30% of the total data samples. The training dataset are further divided into
training sets and verification sets by 5-fold crossover to search for optimal parameters of ANFIS. In
order to avoid contingency, each parameter combination runs 100 times and the result is averaged, and
then select the best combination of parameters as the optimal parameters for ANFIS. In this
experiment, based on the data of Simulink, the results of finding the optimal parameters of ANFIS by
K-fold are shown in Table 1.

Table 1. The classification accuracy of traversal the type and the number of membership function in
100 times 5-fold crossover test for the simulation data sample.
Number of MF 3 4 5
Type of MF Training Testing Training Testing Training Testing
accuracy% accuracy% accuracy% accuracy% accuracy% accuracy%
Trimf 99.32 99.16 99.75 99.58 99.81 99.76
Trapmf 99.26 99.24 99.24 99.18 99.52 99.40
Gbellmf 99.66 99.62 99.78 99.75 99.87 99.84
Gaussmf 99.63 99.62 99.78 99.72 99.88 99.83
Pimf 98.62 98.55 99.24 99.16 99.50 99.42
Dsigmf 99.31 99.20 99.77 99.70 99.88 99.83
Psigmf 99.31 99.26 99.76 99.70 99.87 99.82

Table 2. Comparison of classification accuracy of BPNN and ANFIS models in different types
of faults for the simulation data sample.
BPNN ANFIS
Training Testing Training Testing
Item accuracy% accuracy% accuracy% accuracy%
OC 100.00 100.00 100.00 100.00
LL1 100.00 100.00 100.00 100.00
LL2 98.90 98.69 100.00 100.00
PS 98.93 98.62 100.00 99.91
D-S 92.92 91.06 99.04 99.25
D-A 99.78 99.55 100.00 100.00
N 100.00 100,00 100.00 100.00

5
EEEP 2019 IOP Publishing
IOP Conf. Series: Earth and Environmental Science 467 (2020) 012083 doi:10.1088/1755-1315/467/1/012083

According to the results of Table 1, we find that the result of parameter combination which is
Gbellmf and each input 5 membership function are the best combination parameter of fault diagnosis
using ANFIS. Therefore, we take this parameters combination as the optimal parameter of ANFIS
model. In order to make the results more stable and reliable, the fault diagnosis model is trained and
tested for 100 independent times. To verify the superior performance of the ANFIS based model, the
common BPNN model is used for the comparison in this experiment. The BPNN model is trained for
the same 100 independent times. The average results of the two methods are shown in Table 2.

4. Experimental study
In this section, a 2kw small grid-connected photovoltaic system installed on the roof of the North
Building of the College of Physics and Information Engineering of Fuzhou University is used to
experimentally verify the proposed photovoltaic fault diagnosis model. The same faults as detailed in
previous section 3 is simulated on the laboratory PV array, and the data collection for each the fault
condition is performed continuously for 3 hours with the sampling rate of 1 Hz. The collected raw data
include the PV array Immp, Vmmp, PV array temperature Ta and environmental irradiance Ga. Therefore,
10800 raw data samples are collected for each fault condition, and thus there are 75600 raw data
samples in total. And then, the fault features are extracted from these raw data to form the final data
samples. The training and testing of the PV fault diagnosis model is the same as the simulation model
as discussed and detailed in section 3. The obtained experimental results are listed in Table 3. Because
the experimental data samples are noisy and the simulation data samples are ideal, the training
accuracy and the test accuracy of the experimental data is slightly less than those of the simulation
data. However, the experimental accuracy is still relatively high, which demonstrates that the PV fault
diagnosis model is effective for real PV array. As shown in Table 4, the comparison results show that
the overall accuracy of the ANFIS based fault diagnosis model is better than the BPNN based model.

Table 3. The classification accuracy of traversal the type and the number of membership function in
100 times 5-fold crossover test for the experimental data samples.
Number of MF 3 4 5
Type of MF Training Testing Training Testing Training Testing
accuracy% accuracy% accuracy% accuracy% accuracy% accuracy%
Trimf 90.02 90.06 91.25 91.29 92.81 92.76
Trapmf 90.15 90.14 91.24 91.18 92.52 92.40
Gbellmf 92.25 92.26 93.46 93.47 96.84 96.95
Gaussmf 92.27 92.17 94.48 94.42 97.88 97.68
Pimf 91.06 91.05 92.04 92.06 95.46 95.42
Dsigmf 91.09 91.02 92.17 92.16 94.38 94.34
Psigmf 91.02 91.06 92.17 92.20 94.36 94.35

Table 4. Comparison of classification accuracy of BPNN and ANFIS in different types of faults for
the experimental data samples.
BPNN ANFIS
Training Testing Training Testing
Item accuracy% accuracy% accuracy% accuracy%
OC 97.63 99.84 98.54 98.34
LL1 99.97 99.98 99.99 99.99
LL2 80.51 80.11 94.02 96.55
PS 89.42 94.86 98.53 98.91
D-S 90.63 86.01 90.01 89.91
D-A 99.44 99.85 99.89 99.91
N 99.97 99.96 99.97 99.98

6
EEEP 2019 IOP Publishing
IOP Conf. Series: Earth and Environmental Science 467 (2020) 012083 doi:10.1088/1755-1315/467/1/012083

5. Conclusions
This paper proposes a PV array fault diagnosis method based on the adaptive network based fuzzy
inference system for online detection of PV array line-line faults, open circuit faults, degradation faults
and partial shadow faults, which is based on the operating voltage, current, temperature and irradiance
of PV arrays. Seven features are extracted from these raw data to eliminate the influence of
environmental factors. In order to obtain the best diagnostic results, the averaging ANFIS parameters
and K-fold cross validation are used to search the best parameters of the model. The PV array faults
are simulated on a Simulink based model and a real laboratory PV system to obtain massive data
samples, so as to verify the accuracy of PV array fault diagnosis model. The results show that the
proposed ANFIS based PV fault diagnosis model has superior performance in terms of the accuracy,
reliability and generalization performance in comparison to the BPNN based model. The overall
accuracy of the proposed model based on the simulation data is 99.9%, while the overall accuracy for
the model based on the experimental data is over 97%.

Acknowledgments
The authors would like to acknowledge the financial supports in part from the National Natural
Science Foundation of China (Grant Nos. 61601127 and 61574038), the Fujian Provincial Department
of Science and Technology of China (Grant Nos. 2019H0006 and 2018J01774), and the Foundation of
Fujian Provincial Department of Industry and Information Technology of China (Grant No. 82318075).

References
[1] S R Madeti and S Singh 2017 Monitoring system for photovoltaic plants: A review Renewable
and Sustainable Energy Reviews 67 1180-1207
[2] M K Alam, F H Khan, J Johnson and J Flicker 2013 PV faults: Overview, modeling, prevention
and detection techniques, in Control and Modeling for Power Electronics (COMPEL), 2013
IEEE 14th Workshop on 1-7 IEEE
[3] Z Chen, L Wu, S Cheng, P Lin, Y Wu and W Lin 2017 Intelligent fault diagnosis of
photovoltaic arrays based on optimized kernel extreme learning machine and IV
characteristics Applied Energy
[4] Z Chen, Y Chen, L Wu, S Cheng and P Lin 2019 Deep residual network based fault detection
and diagnosis of photovoltaic arrays using current-voltage curves and ambient conditions
Energy Conversion and Management 198, 111793
[5] X Lu et al. 2019 Fault diagnosis for photovoltaic array based on convolutional neural network
and electrical time series graph Energy Conversion and Management 196 950-965
[6] Z Chen et al. 2018 Random forest based intelligent fault diagnosis for PV arrays using array
voltage and string currents Energy Conversion and Management 178 250-264
[7] M Dhimish, V Holmes, B Mehrdadi and M Dales 2018 Comparing Mamdani Sugeno fuzzy
logic and RBF ANN network for PV fault detection Renewable Energy 117
[8] K Polat and S Güneş 2007 An expert system approach based on principal component analysis
and adaptive neuro-fuzzy inference system to diagnosis of diabetes disease Digital Signal
Processing 17(4) 702-710
[9] M Kuai, G Cheng, Y Pang and Y Li 2018 Research of Planetary Gear Fault Diagnosis Based on
Permutation Entropy of CEEMDAN and ANFIS Sensors (Basel) 18(3) Mar 5

7
Reproduced with permission of copyright owner. Further reproduction
prohibited without permission.