Solar Energy 181 (2019) 137–147

Contents lists available at ScienceDirect

Solar Energy
journal homepage: www.elsevier.com/locate/solener

Study of joint temporal-spatial distribution of array output for large-scale T

photovoltaic plant and its fault diagnosis application
Honglu Zhua,b, Haizheng Wangb, Dahai Kangc, Lizhong Zhangc, Lingxing Lub, Jianxi Yaoa,b,
⁎
Yang Hud,
a
State Key Laboratory of Alternate Electrical Power System with Renewable Energy Sources and the School of Renewable Energy, North China Electric Power University,
Beijing, China
b
School of Renewable Energy, North China Electric Power University, Beijing, China
c
Beijing Power Concord Technology Co., Ltd., China
d
School of Control and Computer Engineering, North China Electric Power University, Beijing, China

A R T I C LE I N FO A B S T R A C T

Keywords: The temporal-spatial distribution of photovoltaic (PV) array output in large-scale PV plant has following fea-
PV power generation tures: (i) In the DC (direct current) side of PV plant, the outputs from different arrays display a substantial
Fault diagnosis correlation with each other. (ii) The dynamic difference for a PV array is covered by the random fluctuation of
Joint temporal-spatial characteristic PV output. (iii) It is difficult to explain the occurrence and evolution process of PV faults with operation data
Fuzzy logic
only. Currently, most fault diagnosis methods for the PV arrays do not take advantage of the temporal-spatial
distribution information contained in the operation data. To solve these above problem, a new fault diagnosis
method using the spatio-temporal distribution characteristics of photovoltaic array output is proposed. Here, the
temporal fluctuation and spatial distribution characteristics of PV array output under different fault conditions
were analyzed. The spatio-temporal composite function is constructed, and then used to set fuzzy rules for fault
diagnosis of the PV array. Finally, an example is used to verify the effectiveness of the proposed method, and the
results show that the method can describe the characters of spatial and temporal distribution of PV output under
faults conditions, and it can effectively classify different faults.

1. Introduction between PV modules to identify fault modules. Madeti and Singh

(2017b) used signals of sensors mounted on each PV module to locate
The global installed capacity of PV power generation exceeded the fault module and identify the fault type. Such methods need huge
397GW in 2017, with an additional installed capacity of 94GW, an amount of manpower and external sensors which greatly increase the
increase of 32% over the same period of 2016. The equipment in DC cost of PV plant maintenances and need more financial support. All the
side of the PV plant is large in number and complex in structure. They above reasons limit the use of such strategy in large-scale PV plant. The
are affected by the working environment, self-aging and other factors, second one is based on reference model, Harrou et al. (2018) analyzed
which may result in different faults. If the fault is not detected in time, the deviation between actual data and calculated data for fault diag-
the power generation efficiency of PV plant can be greatly reduced and nosis. Zhe and Amir (2016) built a standard reference model by support
even seriously endanger the safety of PV plant. Therefore, reliable vector machine (SVM) and executed fault diagnosis by deviation cal-
monitoring of the operation status of PV arrays and detection of the culation. Chine et al. (2017) established a standard reference model by
fault is necessary for the operation of modern PV plant. BP neural network and also performed fault diagnosis through devia-
Many approaches have been proposed for fault diagnosis of PV tion evaluation. Such methods can effectively determine the fault types
plant, and there are mainly three types of strategies for DC side fault by deviation analysis, but the threshold setting depends on expert
diagnosis (Madeti and Singh, 2017a; Mellit et al., 2018). The first one knowledge which limits its wider applicability. The third one is realized
relies on infrared thermal imaging or external sensors. Guerriero and via intelligent classification algorithm. Garoudja et al. (2017a) pro-
Daliento (2017) used infrared thermal camera to detect differences posed a probabilistic neural network (PNN) classifier for PV fault

⁎
Corresponding author.
E-mail addresses: [email protected] (H. Wang), [email protected] (D. Kang), [email protected] (L. Zhang),
[email protected] (L. Lu), [email protected] (J. Yao), [email protected] (Y. Hu).

https://doi.org/10.1016/j.solener.2019.01.083
Received 23 August 2018; Received in revised form 12 January 2019; Accepted 26 January 2019
0038-092X/ © 2019 International Solar Energy Society. Published by Elsevier Ltd. All rights reserved.
H. Zhu et al. Solar Energy 181 (2019) 137–147

Fig. 1. The configuration of a large-scale PV plant.

diagnosis. Zhu et al. (2018) proposed unsupervised sample clustering value can effectively find the fault PV array. However, the PV plant
method and established probabilistic neural network model. working in real environment is a nonlinear time-varying object, it is
Ramakrishna and Singh (2018) used KNN method to detect and classify assuredly helpful to well know the characteristics information of the
the fault as well as locate the faulted array. Ramos et al. (2018) used fault evolution process. And the performance deviation among different
fuzzy clustering techniques with automatic learning for on-line PV arrays is not fully considered, so the paper try to characterize the
faults detection. Rabah and Samir (2018) used decision tree algorithm performance deviation using spatial distribution information.
to detect and diagnose the faults in grid connected PV system. Liao et al. For the analysis of the time series characteristics of a PV array, Ji
(2017) proposed an improved particle swarm optimization (IPSO) al- et al. (2017) used time series information to build the model of PV
gorithm for fault classification using fault sample data. Chen et al. array. Harrou et al. (2018) established a reference time series model
(2017) used the extreme learning machine (ELM) and Zhao et al. (2018) based on the historical data of PV array and the model output is com-
used a fuzzy clustering method to identify faults. The intelligent clas- pared with the real output for fault diagnosis. Garoudja et al. (2017b)
sification strategy avoids the fault modeling complexity but requiring built the time series model with exponentially weighted moving-
high-quality fault data for training. average (EWMA) method to obtain the predicted value and the
The above three methods are widely used, but the dispersion in- threshold for different fault, and the residual error between actual value
formation between PV arrays and the information contained in the PV and predicted value was used for fault diagnosis. Zhong et al. (2015)
array output time series are not effectively used. The setting of in- directly used the differences between the characteristics of a historical
dicators and thresholds for different faults is the common method in PV time series and a real time series for fault diagnosis. Most of the time
fault diagnosis system (Pillai and Rajasekar, 2018). Guoxuan et al. series information is used to build the PV output model, but the mod-
(2014) executed the fault detection based on the dispersion rate of PV eling accuracy of the model is difficult to meet the requirements of fault
array current. Han et al. (2018) compared the current of different ar- diagnosis. The using of deviations for time series to analyze fault
rays by the Hampel identification method to find outlier data and locate characteristics is an idea of this paper.
the fault array. Zhao et al. (2013) set the thresholds of fault arrays by The DC side of a large-scale PV plant is composed of large numbers
comparing the current among different arrays and the thresholds were of PV arrays. There is a strong correlation between the output of dif-
continuously updated. Liu et al. (2017) used the spatial information ferent arrays, and the abnormal outputs of PV array caused by faults are
between the array level and the global level array to diagnose array covered by the fluctuations of outputs. The output of PV array has time-
fault. The fault diagnosis methods using discretization rate or threshold sequence characteristic, and the fault information hidden in massive

138
H. Zhu et al. Solar Energy 181 (2019) 137–147

time series data is difficult to get. The using of the time series and level now, and it makes fault diagnosis of PV array possible;
spatial information between PV arrays is a new idea for PV array fault 3. The outputs of different PV arrays are of strong correlation, and such
diagnosis. In this work, a DC side fault diagnosis method based on tremendous similarity data is difficult to analyze.
temporal-spatial distribution of PV array output is proposed. Main in-
novation of this paper is to propose a method to extract the spatial and Through the above analysis, the array current in the actual PV
temporal distribution features of PV array under different fault condi- power plant can reflect the operating state of different arrays. The array
tions. The proposed method effectively shows the dynamic character- voltage in one combiner box is the same, so the array current is used in
istics and the spatial dispersion characteristics of different arrays, and the analysis of the paper, and the voltage of the parallel connected PV
how to feature the temporal-spatial distribution of PV output under arrays cannot be used in the fault diagnosis in large-scale PV plant.
different faults is the core of this paper. The rest of the paper is as
follows. Section 2 studies the spatial and temporal characteristics of PV 2.2. Temporal and spatial distribution characteristics of PV array current
array output in a large-scale PV plant. In Section 3, the temporal-spatial
distribution function of the PV array is established and the temporal- To directly show the deviation of array current output, the reference
spatial distributions of PV array current are analyzed. Section 4 pro- current is introduced and compared with the actual current value. The
poses a fault diagnosis method for large-scale PV plants. In Section 5, reference current Im is the theoretical maximum power current of the
the experimental verification of the proposed method is carried out. PV array. It can be calculated as (Liu et al., 2015):
Finally, Section 6 summarizes the major innovations of this work.
G
Im = Im − ref [1 + a (T − Tref )]· ,
2. Temporal and spatial distribution characteristics of PV arrays Gref (1)
2
G is the measured solar irradiation intensity, Gref = 1000 W/m is the
2.1. Operation monitoring of large-scale PV plant reference solar irradiation intensity, T is the measured PV module
temperature, Tref = 25 °C is the reference PV module temperature, Im-ref
Fig. 1 shows the configuration of a large-scale PV plant in Shandong is the optimal operating point current for a PV module under standard
Province, China. The installed capacity of the PV plant is 40 MW, in- conditions. The temperature compensation coefficient a changes with
cluding about 130,000 PV modules, with more than 8000 PV arrays. the temperature and irradiance, but within the real range of PV plant
The power plant consists of 553 intelligent PV combiner boxes and 74 operational temperature and irradiance, a remains around 0.0003 (Yu
inverters. et al., 2018), so formula (1) can be simplified as follows:
Fig. 2 shows the output current and voltage of 9 arrays in the PV
G
plant for 18 days. From Fig. 2, it can be seen that the outputs from Im = Im − ref · .
different arrays are highly correlated and they fluctuate with changes of Gref (2)
the weather condition, which increases the difficulty of fault diagnosis. The current data of a array for one year is used for the fitting for-
In actuality, the data from large-scale PV plants are even more complex mula (2), and the calculated results can be used as the reference cur-
than the data shown in Fig. 2, in order to detect the fault of the PV plant rent. The impact of temperature to the operation current (the maximum
from such a huge amount of data, it is necessary to extract the fault power point current) of the array can be ignored. Fig. 3 shows the fit-
characteristics from the time series information and spatial difference of ting result, and the operation current can be fitted as
PV arrays. The features of the large-scale PV plant monitoring are
summarized as follows: ⎧ Im = 0.0077·G - 0.009, ifG ⩾ 5(W/m2)
.
⎨
⎩mI = 0, if0 ⩽ G < 5(W/m2) (3)
1. PV arrays connected to the same combiner box are parallel struc-
ture, so the voltage data of different arrays collected to one com- For the analyzed PV plants are equipped with a total of 74 inverters,
biner box is the same. For the fault diagnosis of a PV array, the and each inverter is connected with 112 arrays. To get the spatio-
available data is the PV array current; temporal distribution characteristics of the array current, the operation
2. The monitoring resolution of most PV plants can reach the array data of 30 arrays were randomly selected for analysis. Fig. 4 shows the

Fig. 2. The current (a) and voltage (b) of different PV arrays. Fig. 3. The linear relationship between solar irradiance and the current.

139
H. Zhu et al. Solar Energy 181 (2019) 137–147

Fig. 4. The current of different arrays on the same day (a), and current of one array on different days (c). The difference between actual current and reference current
of different arrays on the same day (b), and the difference of the same array on different days (d).

Fig. 5. The distribution of time component (a) and spatial component (b).

current of the selected arrays in 21 days, and the temporal and spatial radiation and shows strong random fluctuation characteristics;
distribution characteristics of the array current can be summarized as 4. The uncertainty from strong random fluctuations and the different
follows: performance of PV arrays impedes algorithm implementation for
array current analysis and fault diagnosis.
1. The currents between the different arrays are highly similar, but
there are differences among different array for the mismatch caused So how to description the spatio-temporal distribution character-
by received solar irradiation, performance degradation and failure istics for the output of PV arrays in large-scale PV plant is an effective
losses of different array; way to realize fault diagnosis under the current monitoring conditions
2. As the distance of different arrays increases, the deviation among of PV plant.
them tends to increase;
3. The output current of a PV array is mainly affected by the solar

140
H. Zhu et al. Solar Energy 181 (2019) 137–147

3.2. Spatio-temporal distribution characteristics of PV array output

Fig. 5 shows the time component and spatial component of three PV

arrays in normal condition for five days. Figure a. shows the distribu-
tion of the time component, and figure b. shows the distribution of the
spatial component. The time component of fault free PV arrays’ output
fluctuates with time, and they are almost the same for different arrays.
The spatial components of PV arrays’ output are in the range of −0.3 to
0.2, with some differences between the spatial components of different
arrays’ output.
Fig. 6 shows the joint spatial-temporal distribution of the three PV
arrays. The joint spatial-temporal distribution of the array under
normal operation condition is on the right side of the origin of the
coordinate system. Due to the difference in the output of the array and
the volatility output of the array, there are fluctuations in the dis-
tribution.
As shown in Figs. 5 and 6, the time component can reflect the dy-
namic process of arrays output deviating from the reference value, and
the spatial component can represent distance deviating from the re-
ference value. Combining the two component can effectively extract
more information hidden in static data for the realization of PV array
Fig. 6. Spatio-temporal distribution of arrays current. fault diagnosis.

3. Description of the spatial-temporal distribution of photovoltaic 4. The fault diagnosis method

array output
4.1. Fuzzy logic
3.1. Construction of spatio-temporal composite function
Fuzzy logic system is applied more and more in PV system. A brief
Suppose there are N data x1,x2, … ,xN and each piece of data consists overview of the recent publications on fuzzy logic system design is
of a spatial component and a time component. The kth data xk is re- presented by Suganthi et al. (2015). Dhimish et al. (2017) used fuzzy
presented as xk = [xk(s)|xk(t)], where xk(s) represents the spatial com- logic for PV fault diagnosis, but the method in this paper has high re-
ponent of the data and xk(t) represents the time component. Assuming quirements about fault data. The PV output shows nonlinearity and
that the dimension of spatial component is r and the dimension of time uncertainty, and fuzzy logic is suitable to it (Zadeh, 2008). So, in this
component is q, then the expression of the kth data can be obtained as paper, fuzzy rules are combined with spatio-temporal composite func-
follows: tion. And there are four steps of fuzzy logic: fuzzification, knowledge
base, fuzzy reasoning, and defuzzify (Dhimish et al., 2018).
xk = [xk1 (s ), ...,xkr (s ) | xk1 (t ), ..,xkq (t )]. (4) Step 1. Fuzzification describes the conversion of input variables into
the fuzzification value, and expresses the value with a corresponding
The spatio-temporal composite function can be defined:
fuzzy set. The inputs are the external reference input, the system output
D = DT (xT , xS ) + DS (xT , xS )·i. (5) or state, and similar factors. The fuzzification operation is as follows:
x = fz (x 0), (10)
The time component DT is the real part of the composite function
and the spatial component DS is the imaginary part of the composite where x0 is the input variable of the input; x is the fuzzy set; fz re-
function. presents the fuzzification operator.
The temporal dimension function is: Step 2. The knowledge base includes the knowledge of specific ap-
plication areas. This is usually composed of two parts: a database and a
DT (xT , xS ) = |IT , S − Ipre, T , S| , (6) fuzzy rule base. The database mainly contains the membership func-
tions of various linguistic variables and fuzzy space grading numbers.
|·| is the arithmetic symbol indicating the absolute value; DT(xT,xS) is The rule base contains a series of control rules expressed by fuzzy lin-
the time component of array S at time T; IT,S is the current of array S at guistic variables. The fuzzy rules usually are determined from expert
time T; Ipre,T,S is the theoretical reference current of array S at time T: experience and knowledge. In the control process, rules are described
by:
Ipre, T , S = b·GT + c, (7)
IF (Satisfy a set of conditions) THEN (Get aset of conclusions). (11)
b, c are the coefficient obtained by fitting the historical data; GT is the
Step 3. Fuzzy reasoning is the key of fuzzy control, and it simulates
solar irradiation intensity at time T.
human reasoning based on fuzzy concepts. The reasoning process is
The spatial dimension function is:
based on the implication relations and reasoning rules of fuzzy logic.
DS (xT , xS ) = IT , S − I¯T , That is, if the input fuzzy value is: x is A and y is B, then according to
(8)
the rules the output fuzzy value z (represented by fuzzy sets C) is:
where DS(xT,xS) is the spatial component of S arrays at T time; ĪT is the C = (AandB ) ∘R, (12)
combiner box’s reference current at time T:
where ‘and’ is the connection operator; ○ is the synthetic operator; R is
n
∑S = 1 IT , S a combination of n fuzzy relations.
I¯T = , Step 4. Defuzzify is the process of selecting the single value that best
n (9)
represents the fuzzy set, and it is completed by inference. The most
n is the number of arrays in the combiner box. common method of defuzzification is the weighted average method

141
H. Zhu et al. Solar Energy 181 (2019) 137–147

Fig. 7. Flow chart of the proposed fault diagnosis method.

Table 1
The electrical characteristics of different faults.
Type of fault Open circuit voltage Short circuit current Optimal operating point voltage Optimal operating point current Array current

Abnormal aging Unchangeable Unchangeable Decline Decline A slight drop

Shadow Decline Unchangeable Decline Decline Decline then recovery
Open circuit Unchangeable 0 Unchangeable 0 Down to zero

142
H. Zhu et al. Solar Energy 181 (2019) 137–147

Fig. 8. Experimental platform and array failure settings.

(center of gravity method). The weighted average method uses the the actual voltage and the theoretical voltage of the array.
weighted average of uC (z) as the crisp value of the output fuzzy value z:
b
∫a zuC (z )dz 5. Case study
z 0 = df (z ) = b
,
∫ uC (z )dz
a (13) 5.1. Data acquisition
where z0 is the crisp value of z; df is the defuzzification operator; uC is
The following faults often occur in the DC side of PV power plants:
the membership function of the fuzzy set; the domain of z is [a, b].
open circuit, abnormal aging, short circuit fault, shadow, ground fault,
arc fault and bypass diode fault (Sun et al., 2017; Ali et al., 2017; Pillai
4.2. Fault diagnosis method
et al., 2018). The open circuit fault will cause the current of the array to
drop to 0, the shadow and abnormal aging fault will cause the current
The flow chart of the proposed fault diagnosis method is shown in
and voltage of the array to decline, and the ground fault and short
Fig. 7:
circuit fault will show the array current remains unchanged and the
voltage drops (Shimakage et al., 2011). The short circuit of bypass
(1) Data preprocessing: the environment data and electric data of PV
diode causes a voltage drop at the PV array, and the open circuit of
plant is pre-processed. Here, the quartile deviation method is used
bypass diode causes hot spots under shadow fault (Brooks et al., 2015).
to clean the abnormal data, and the sliding average method is used
But as discussed in Chapter 2, the voltage changes caused by various
to smooth the data.
faults are covered by the parallel connection of PV arrays, so only the
(2) Establishing the fuzzy rule base: the reference current is fitted using
faults shown in Table 1 are selected for analysis in this paper.
history data which includes solar irradiation intensity and arrays
Table 1 describes the electrical characteristics of different faults. In
current. And the spatial component and time component for each
order to illustrate the temporal and spatial distribution characteristics
array current are calculated by formula (6)–(9). The temporal and
of PV array output under different fault conditions, one PV array was
spatial distribution characteristics of different faults are used for the
selected (array A9 in the M28M366 combiner box), and different faults
training of the temporal and spatial membership functions.
are set in this array. The resolution of the data was 10 min. The settings
(3) Fault diagnosis: the array voltage and current are taken as the in-
of faults are presented as Fig. 8:
puts. The spatial and temporal component of the array current is
calculated by the spatio-temporal composite function. Takes spatial
component and time component of each array current as inputs, the 5.2. Spatio-temporal distribution for PV array output under fault conditions
fuzzy output value is given by fault model base. Then is the process
of defuzzification for the results, and the fault diagnosis of array Fig. 9 illustrates the spatio-temporal distribution of the PV array
current is completed by the proposed method. The voltage of PV output under different fault conditions. Figure a. shows the contrast
array just used to recognize abnormal of the array as the method between the fault array and the fault free array. There was fault free on
proposed by Garoudja et al. (2017b). And the abnormal array is the first day. On the second day, abnormal aging was set in the ex-
classified by threshold([LCL, UCL]) and the deviation(Vt) between perimental array. There was a shadow fault on the third day at noon.
On the fourth day, open circuit was set. Figure b. shows the spatio-

143
H. Zhu et al. Solar Energy 181 (2019) 137–147

Fig. 9. Comparison of output current (a), joint spatio-temporal distribution (b), time component distribution (c) and spatial component distribution (d) between fault
free array and fault array.

Table 2
Spatial component and time component distributions of different faults.
Type of fault Spatio-temporal distribution Time component Spatial component

Fault free Near the origin Fluctuation around 0.5 Fluctuation around 0
Abnormal aging Slight deviation from the origin Fluctuation around 1.25 Fluctuation around −1.5
Shadow Large deviation from the origin Fluctuation around 2 Fluctuation around −4
Open circuit Drastic deviation from the origin Fluctuation around 2.75 Fluctuation around −8

temporal distribution of different faults. Figure c. and figure d. show According to Table 2, the membership functions of each input
the change of the time component and the spatial component of the variable are generated, and outputK is set to [0,5], and its corre-
fault free and the fault array. Table 2 summarizes the temporal and sponding fuzzy sets and quantization domain are
spatial distribution characteristics of PV array output under different
{K } = {L, ZE , S, M , B}, (15)
fault conditions.
Fig. 9 and Table 2 shows the distributions of the spatial and time where ZE is normal operation, S is abnormal aging fault, M is shadow
components of PV array output under different fault conditions, sug- fault, B is open circuit, and L is unknown fault. The membership
gesting that it is feasible to diagnose faults based on the spatio-temporal function of the input and output is shown in Fig. 10.
distribution characteristics of the PV array output. A series of fuzzy rule groups are established based on the above
studies. The fuzzy rule table and the input/output relation diagram of
5.3. Fuzzy rules generating the fuzzy system are shown in Table 3 and Fig. 11 respectively.

The time component DT(xT,xS) and the spatial component Ds(xT,xS) 5.4. Verification
are taken as the input of fuzzy rules system, and the K is taken as the
output variable. The time component DT(xT,xS) and the spatial com- The array A9 in the M28M366 combiner box was set to different
ponent Ds(xT,xS) are set to [0,3] and [−9,0.5], respectively. The cor- fault conditions for 16 days (normal, abnormal aging, shadow occlu-
responding fuzzy sets and quantization domain are defined as follows. sion, and open circuit), and 1146 samples were obtained. The 16-day
{DS (xT , xS )/ DT (xT , xS )} = {ZE , S, M , B}, (14) experiment was divided into 4 groups, each group contains 4 days’ data,
as the fault setting of different groups is shown in Table 4.
where ZE is normal; S is small; M is middle; B is big. The accuracy of the proposed method for different fault is shown in

144
H. Zhu et al. Solar Energy 181 (2019) 137–147

Fig. 10. Membership function.

Table 3
Fuzzy logic rule table.
DS/DT ZE S M B

ZE Normal operation Normal operation Unknown fault Unknown fault

S Abnormal aging Abnormal aging Abnormal aging Unknown fault
M Unknown fault Shadow Shadow Shadow
B Unknown fault Unknown fault Open circuit failure Open circuit failure

aging degree, and with the increase of aging degree the uncertainty
brought by other factors decreases. For shadow faults, the accuracy of
fault diagnosis increases with the increase of the number of shadow
modules, but the accuracy decreases with low irradiance in the early
morning and late afternoon. The fault diagnosis accuracy of open cir-
cuit fault is higher than abnormal aging and shadow fault. But the
uncertainty of the diagnosis result is higher in the early morning and
late afternoon, so there will be a higher misjudgment rate.

6. Conclusion

This paper proposes a spatial-temporal composite function to de-

scribe the output feature for array in large-scale PV plant. And the
spatial-temporal composite function is combined with fuzzy rules for
the fault diagnosis of PV array. The method was verified by the ex-
periment in the actual PV plant. The main innovations of the paper are
as follows:
Considering the monitoring status of the PV plant, the proposed
Fig. 11. Fuzzy system input and output relationship diagram. method can detect different faults just using the array current, which
allowing the available data from actual PV plant to be used effectively.
Table 5. The Fig. 12 and Table 5 show the proposed method can detect This is the first study of the joint spatio-temporal characteristics of PV
different fault effectively. array output under different fault conditions. And the proposed spatio-
The statistical results in Table 5 show that: For abnormal aging temporal composite function can effectively describe the distribution
faults, the accuracy of fault diagnosis increases with the increase of features of array current under different fault conditions. More im-
portantly, the function can reveal the dynamic evolution process and

Table 4
Experimental setup for four groups of experiments.
Group number Day1: Normal operation Day2: Abnormal aging Day3: Shadow Day4: Open circuit failure

Group 1 – F1: 6 Ω resistor in series F2: 4 modules shaded at noon F3: Open circuit in the afternoon
Group 2 – F4: 8 Ω resistor in series F5: 4 modules shaded in the morning and afternoon F6: Open circuit in the morning
Group 3 – F7: 10 Ω resistor in series F8: 6 modules shaded at noon F9: Open circuit at noon
Group 4 – F10: 12 Ω resistor in series F11: 6 modules shaded in the morning and afternoon F12: Open circuit in the morning and afternoon

145
H. Zhu et al. Solar Energy 181 (2019) 137–147

Table 5
Fault diagnosis accuracy statistics.
Type of fault Total number of samples (Irradiation intensity > 0 W/m2) Misjudged sample number Accuracy rate

Normal operation 327 5 98.5%

F1: 6 Ω resistor in series 67 7 89.6%
F4: 8 Ω resistor in series 88 3 96.6%
F7: 10 Ω resistor in series 66 2 97.0%
F10: 12 Ω resistor in series 65 3 95.4%
F2: 4 modules shaded at noon 52 2 96.2%
F5: 4 modules shaded in the morning and afternoon 43 5 88.4%
F8: 6 modules shaded at noon 46 1 97.8%
F11: 6 modules shaded in the morning and afternoon 42 4 90.5
F3: Open circuit in the afternoon 25 0 100%
F6: Open circuit in the morning 27 1 96.3%
F9: Open circuit at noon 21 0 100%
F12: Open circuit in the morning and afternoon 42 1 97.6%

Fig. 12. Actual current distribution of the fault free array and fault array and the fault diagnosis results of group 1 (a), group 2 (b), group 3 (c), group 4 (d).

the static deviation of the arrays in large-scale PV plant. characteristics. Appl. Energy 204.
Chine, W., Mellit, A., Lughi, V., Malek, A., Sulligoi, G., Pavan, A.M., 2017. A novel fault
diagnosis technique for photovoltaic systems based on artificial neural networks.
Acknowledgement Renew. Energy 90, 501–512.
Dhimish, M., Holmes, V., Mehrdadi, B., Dales, M., 2018. Comparing Mamdani Sugeno
The authors would like to acknowledge the financial support of the fuzzy logic and RBF ANN network for PV fault detection. Renew. Energy 117.
Dhimish, M., Holmes, V., Mehrdadi, B., Dales, M., Mather, P., 2017. Photovoltaic fault
Research Funds from State Grid Corporation of China detection algorithm based on theoretical curves modelling and fuzzy classification
“SGHADK00PJJS1800072”. system. Energy 276–290.
Garoudja, Elyes, Chouder, Aissa, Kara, Kamel, Silvestre, Santiago, 2017a. An enhanced
machine learning based approach for failures detection and diagnosis of PV systems.
Reference
Energy Convers. Manage. 151, 496–513.
Garoudja, E., Harrou, F., Sun, Y., Kara, K., Chouder, A., Silvestre, S., 2017b. Statistical
Ali, M.H., Rabhi, A., Hajjaji, A.E., Tina, G.M., 2017. Real time fault detection in photo- fault detection in photovoltaic systems. Sol. Energy 150, 485–499.
voltaic systems. Energy Proc. 111, 914–923. Guerriero, P., Daliento, S., 2017. Automatic edge identification for accurate analysis of
Brooks, A.E., Cormode, D., Cronin, A.D., Kam-Lum, E., 2015. PV system power loss and thermographic images of solar panels. International Conference on Clean Electrical
module damage due to partial shade and bypass diode failure depend on cell behavior Power.
in reverse bias. Photovoltaic Specialist Conference. Guoxuan, F.U., Xue, L.I., Gao, X., 2014. Analytical method and application of current
Chen, Z., Wu, L., Cheng, S., Lin, P., Wu, Y., Lin, W., 2017. Intelligent fault diagnosis of discrete rate of PV power station based on junction box string. Power Syst. Clean
photovoltaic arrays based on optimized kernel extreme learning machine and I-V Energy 30 (11), 109–113.

146
H. Zhu et al. Solar Energy 181 (2019) 137–147

Han, F.C., Chen, Z.H.C., Wu, L.J., Yu, J.L., Lin, P.J., Cheng, S.Y., Wu, Y.B., 2018. A real- Ramos, Adrián Rodríguez, da Silva Neto, Antônio José, Orestes, L.S., 2018. An approach
time status monitoring and fault location system for photovoltaic arrays. J. Fuzhou to fault diagnosis with online detection of novel faults using fuzzy clustering tools.
Univ. (Nat. Sci. Ed.) 2018 (07), 1–9. Expert Syst. Appl. 113, 200–212.
Harrou, F., Sun, Y., Taghezouit, B., Saidi, A., Hamlati, M.E., Harrou, F., 2018. Reliable Shimakage, T., Nishioka, K., Yamane, H., Nagura, M., Kudo, M., 2011. Development of
fault detection and diagnosis of photovoltaic systems based on statistical monitoring fault detection system in PV system. Telecommunications Energy Conference. IEEE.
approaches. Renew. Energy 116. Suganthi, L., Iniyan, S., Samuel, A.A., 2015. Applications of fuzzy logic in renewable
Ji, D., Zhang, C., Lv, M., Ma, Y., Guan, N., 2017. Photovoltaic array fault detection by energy systems – a review. Renew. Sustain. Energy Rev. 48, 585–607.
automatic reconfiguration. Energies 699. Sun, J., Sun, F., Fan, J., Liang, Y., 2017. Fault diagnosis model of photovoltaic array based
Liao, Z., Wang, D., Tang, L., Ren, J., Liu, Z., 2017. A heuristic diagnostic method for a PV on least squares support vector machine in bayesian framework. Appl. Sci. 7 (12),
system: triple-layered particle swarm optimization–back-propagation neural net- 1199.
work. Energies 10 (2), 226. Yu, J.H., Xu, S.Z., Han, S.W., Li, X.B., Tan, C., Yu, C., Liu, Z., 2018. Analysis of tem-
Liu, Q., Zhao, Y., Zhang, Y., Kang, D., Lv, Q., Shang, L., 2017. Hierarchical context-aware perature characteristics and influence factors of solar cells and PV modules. Sol.
anomaly diagnosis in large-scale PV systems using SCADA data. International Energy 287 (3), 29–36.
Conference on Industrial Informatics. IEEE. Zadeh, L.A., 2008. Fuzzy logic. Computer 21 (4), 83–93.
Liu, X., Zhuo, F., Chen, Y., Xiong, L., 2015. Development of fast simulation models for Zhao, Q., Shao, S.H., Lu, L.X., Liu, X., Zhu, H.L., 2018. A new PV array fault diagnosis
photovoltaic generation system based on Simulink. Energy Conversion Congress and method using fuzzy C-mean clustering and fuzzy membership algorithm. Energies 11
Exposition. (1), 238.
Madeti, Siva Ramakrishna, Singh, S.N., 2017a. A comprehensive study on different types Zhao, Y., Lehman, B., Ball, R., Mosesian, J., 2013. Outlier detection rules for fault de-
of faults and detection techniques for solar photovoltaic system. Sol. Energy 158 tection in solar photovoltaic arrays. Applied Power Electronics Conference and
(Dec), 161–185. Exposition.
Madeti, Siva Ramakrishna, Singh, S.N., 2017b. Online modular level fault detection al- Zhe, H.Y., Amir, H.E., 2016. A novel detection algorithm for Line-to-Line faults in
gorithm for grid-tied and off-grid pv systems. Sol. Energy 157 (Nov), 349–364. Photovoltaic (PV) arrays based on support vector machine (SVM). 2016 IEEE Power
Mellit, A., Tina, G.M., Kalogirou, S.A., 2018. Fault detection and diagnosis methods for and Energy Society General Meeting (PESGM).
photovoltaic systems: a review. Renew. Sustain. Energy Rev. 91, 1–17. Zhong, J., Yan, Z., Wen, F., Zhu, H., Hucheng, L.I., Fan, H., 2015. A fault diagnosing
Pillai, Dhanup S., Rajasekar, N., 2018. A comprehensive review on protection challenges method in power transmission systems based on time series similarity matching.
and fault diagnosis in PV systems. Renew. Sustain. Energy Rev. 91, 18–40. Autom. Electr. Power Syst. 6, 60–67.
Rabah, B., Samir, M., 2018. Fault detection and diagnosis based on c4.5 decision tree Zhu, H.L., Lu, L.X., Yao, J.X., Dai, S.Y., Hu, Y., 2018. Fault diagnosis approach for pho-
algorithm for grid connected pv system. Sol. Energy 173, 610–634. tovoltaic arrays based on unsupervised sample clustering and probabilistic neural
Ramakrishna, M.S., Singh, S.N., 2018. Modeling of pv system based on experimental data network model. Sol. Energy 176, 395–405.
for fault detection using knn method. Sol. Energy 173, 139–151.

147