Solar Energy 174 (2018) 762–769

Contents lists available at ScienceDirect

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

Fault detection in a grid-connected photovoltaic system using adaptive T

thresholding method
Mustapha Ammichea, Abdelmalek Kouadria, Laith M. Halabib, Amar Guichic, Saad Mekhilefb
a
Signals and Systems Laboratory, Institute of Electrical and Electronic Engineering, University M’Hamed Bougara of Boumerdes, Avenue of Independence, 35000
Boumerdès, Algeria
b
Power Electronic and Renewable Energy Research Laboratory (PEARL), Department of Electrical Engineering, Faculty of Engineering, University of Malaya, 50603 Kuala
Lumpur, Malaysia
c
Department of Electronic, University Mohamed Boudiaf, BP 166, 28000 M'Sila, Algeria

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

Keywords: In this paper, an adaptive monitoring scheme with Fuzzy Logic Filter (FLF) is developed and applied to monitor a
MWPCA Grid-Connected Photovoltaic System (GCPVS). This method is based on Principal Component Analysis (PCA) and
PCA Moving Window Principal Component Analysis (MWPCA). It is designed to generate adaptive thresholds for its
Grid-connected PV systems monitoring indices. The FLF filters the monitoring indices to reduce the number of False Alarms (FA) and in-
Process monitoring
crease the Fault Detection Rate (FDR). The application is carried out on the GCPVS of the Power Electronics and
Data-driven methods
Renewable Energy Research Laboratory (PEARL) of Malaya University. The proposed technique is compared
Adaptive thresholds
against PCA method in terms of FAR reduction. The detection ability of the adaptive thresholding with FLF
monitoring scheme is tested first on simulated faults then it is applied to detect a real abnormal behaviour. The
results show that the proposed method is effective in reducing the number of false alarms and in detecting
different types of faults with high accuracy.

1. Introduction techniques developed for a GCPVS monitoring. Many measured vari-

Photovoltaic (PV) energy is considered to be the third most im-
portant renewable energy source after the hydro and wind power
(Spataru et al., 2015). It can be used to generate heat and electricity and
it can replace the unsustainable energies (fossil fuels) (Hassani et al.,
2016).
The performance of a GCPVS and the energy production depends
explicitly on the system monitoring which aims to report information
about the energy potential and energy extracted. In addition, it permits
the detection of different types of faults that may occur (Madeti and
Singh, 2017, Harrou et al., 2018). Many methods have been developed
to monitor a GCPVS, for instance, Firth et al. (2010) based on high
resolution to investigate on a PV system performance and operational
faults where monitoring data is used to develop a new empirical model
of a GCPVS performance which then identifies the occurred faults.
Spataru et al. (2015) proposed a method which takes advantage of the
current-voltage (I–V) measurement for fault diagnosis. Kim (2016) used
wavelet transform for online fault detection in a grid-connected PV
system, Silvestre et al. (2016) developed a new approach based on OPC
technology for automatic supervision and remote fault detection in a
GCPVS. Dhimish and Holmes (2016) have reported some other
ables are involved in a GCPVS. These variables are mainly the solar
radiations, panels temperature, the wind speed, the solar panels current
and voltage values and the inverters current and voltage inputs/out-
puts, etc. Hence, a GCPVS can be seen as a multivariate process. Con-
sequently, Multivariate Statistical Process Control (MSPC) approaches
can be employed for fault detection purposes in this system.
MSPC methods have been widely used to monitor large-scaled dada
processes. They are also named as data-driven or historical data tech-
niques because only historical data are needed (Chiang et al., 2001).
Particularly, Principal Component Analysis (PCA) is a well known
multivariate statistical method (Ge and Chen, 2016; Liu et al., 2011;
Russell et al., 2000; Zhu et al., 2017). The application of this method to
a process monitoring raised some problems that led to extending it to
Dynamic (DPCA) and Kernal PCA to include the dynamics and the
nonlinear behaviours of the data in the statistical model (Deng et al.,
2017; Ku et al., 1995; Rato and Reis, 2013). Moving Window PCA
(MWPCA) and Recursive PCA (RPCA) were developed in order to track
the system’s changes in the operation setpoints (Rato et al., 2016;
Elshenawy et al., 2010). Combined methods, such as Moving Window
Kernel PCA (MWKPCA), were designed to deal with both nonlinear
behaviour and operation setpoint changes (Fazai et al., 2016). To

E-mail address: a.kouadri@univ-boumerdes.dz (A. Kouadri).

https://doi.org/10.1016/j.solener.2018.09.024
Received 23 March 2018; Received in revised form 7 July 2018; Accepted 12 September 2018
0038-092X/ © 2018 Elsevier Ltd. All rights reserved.
M. Ammiche et al. Solar Energy 174 (2018) 762–769

handle the dynamic behaviour in both steady and unsteady states, Lou Once the number of principal components, a, is selected, Xs be-
et al. (2017) proposed a two-step PCA. There are other techniques be- comes as
sides PCA have been also extensively developed and employed for s + E
Xs = X (4)
system monitoring such that Canonical Variate Analysis (CVA) (Russell
et al., 2000, Shang et al., 2017), Partial Least Squares (PLS) (Wold et al., Such that
2001) and Independent Canonical Analysis (ICA) (Zhong and Deng,
X ̂ T
s = TP (5)
2016).
The main problem that one can face when applying linear data- ∼∼T
E = TP (6)
driven techniques, such as PCA, to a process monitoring with fixed
∼ ∼
control limits is constantly the high number of false alarms. These false where P ∈ Rmxa , P ∈ Rm x(m − a) , T ̂ ∈ Rnxa and T ∈ Rn x(m − a)
alarms may be due to the presence of random noise in the system,
dynamics, nonlinear behaviours, the operation setpoint changes and the 2.1.2. The number of principal components
mathematical formulas by which the control limits are calculated. To In the literature, many methods are used for the number of principal
address this problem, adaptive thresholding technique with Fuzzy Logic components, a, selection. These methods mainly include the parallel
Filter (FLF) is developed. The monitoring scheme combines both PCA analysis, cumulative percentage of variances, scree plot and cross-va-
and MWPCA to result in an adaptive and robust technique to false lidation (Chiang et al., 2001, Sheriff et al., 2017). In this work, the
alarms. The use of PCA allows the construction of a statistical model parallel analysis criterion has been used. So, the number of retained
that properly fits the process. This model is kept constant during the principal components, a, is defined as the point at which the eigenvalue
monitoring task. The constructed model is then used to evaluate the profile of the original data crosses the eigenvalue profile of generated
monitoring indices of each new testing sample. A moving window of data assuming independent observations (Chiang et al., 2001).
fixed size allows the generation of the adaptive thresholds for the
evaluated monitoring indices after they have been filtered by FLF. Both 2.1.3. PCA-based fault detection
adaptive thresholds and FLF involve in the number of false alarms re- Two monitoring indices are associated with PCA-based fault de-
duction besides to the PCA performance enhancements. The application tection method, the Hotelling’s T2 and Q statistics
of the proposed method has been performed on the grid-connected PV (Ramahaleomiarantsoa et al., 2012). These indices are provided by
system of the Power Electronic and Renewable Energy Research
̂ −1P T̂ x

T 2 = xT PΛ (7)
Laboratory (PEARL) at Malaya University, Malaysia. The adaptive
scheme with FLF has been tested first to evaluate its detection ability of T 2
Q = (I −PP̂ ̂ ) x = xT (I −PP̂ ̂ ) x
T 2
(8)
random sudden changes, and incipient fault. These faults are associated
with short circuits, an abrupt increase or decrease of currents and The control limits, which correspond to each monitoring index, are
voltages, sensors failure, and slow temperature augmentation. The given by
monitoring scheme has been then used to detect a real abnormal op-
a (n−1)(n + 1)
eration of the GCPVS. The obtained results demonstrate the effective- Tδ2 = Fδ (a, n−a)
n (n−a) (9)
ness of the proposed technique in promptly and correctly detecting
different types of faults. 1
2 h0
The remaining sections of this paper are organized as follows. ⎛ Cδ 2θ2 h 0 θ h (h −1) ⎞
Qδ = θ1 + 1 + 2 0 20
Section 2 is dedicated to the mathematical formulation of the PCA- ⎜ θ1 θ1 ⎟
⎝ ⎠ (10)
based fault detection method. In Section 3, the proposed adaptive
thresholding scheme with FLF has been provided and explained. where n is the number of observations and a is the number PCs.
Section 4 shows the application results and discussions. The last section, Fδ (a, n−a) is an F-distribution of a, n-a degree of freedom evaluated at
Section 5, is restricted to some conclusions and findings. given confidence level (1−δ ) .
m

2. Backgrounds
2.1. Principal component analysis
2.1.1. Definition and mathematical formulation
Principal Component Analysis (PCA) is a dimensionality reduction
technique able to capture the most dominant variances in the data
(Chiang et al., 2001). PCA defines the principal component subspace
and the residual subspace by means of a linear transformation (Chiang
et al., 2001). Given a process data of m measured variables x1, x2, …, xm
and n observations. These data can be stored in a zero mean and unity
variance matrix Xs = [XsT1 , XsT2 , ⋯, Xsm
T
]. The linear transformation, that
projects the data onto the subspaces, is provided by the following
equation
T = Xs P
Where T is the score matrix. P is an m by m orthonormal matrix
consists of the covariance matrix of Xs eigenvectors. The covariance
matrix of Xs is given by the Singular Value Decomposition (SVD) as
cov (Xs ) = PT ΛP
Eq. (1) permits to rewrite Xs as
Xs = TPT
θi = ∑ λij , i = 1, 2, 3
j=a+1 (11)
2θ1 θ2
h 0 = 1−
3θ22 (12)
Cδ is the normal deviate corresponding to (1−δ ) percentile.
3. Proposed adaptive monitoring scheme with FLF
The performance of PCA based fault detection with fixed control
limit is poor in terms of false alarms rate reduction due to the random
noise in the process measured variables and the process dynamics. In
addition, the mathematical formulas used to develop the thresholds
always allow some normal samples to violet the limits which corre-
spond to the type I error. The false alarms rate with this technique can
(1)
be reduced by increasing the confidence interval, δ . However, high
values of δ decreases the PCA monitoring indices sensitivity and hence
some of the faults may not be detectable. The proposed monitoring
scheme aims to overcome this problem and improve the PCA based
fault detection technique. It is a combination of both PCA and Moving
(2)
Window PCA (MWPCA) schemes. The monitoring indices are being
evaluated using a fixed constructed PCA model while the adaptive
thresholds are being updated through a fixed length moving window.
(3)
The sensitivity enhancement is handled by the adaptive thresholds

763
M. Ammiche et al.
whereas the FAR reduction is handled by both the adaptive thresholds
and the FLF. The FLF filters the Q and T2 before making a decision
about the tested samples and then the filtered values are considered for
the adaption process of the thresholds.
3.1. Fuzzy logic filter
Fuzzy logic filters have been widely used to remove noise from
different measured signals (Ferahtia et al., 2012; Naso et al., 2006). In
this paper, the goal of the Fuzzy Logic Filter (FLF) is to remove the spike
noise from the monitoring indices to reduce the number of alarms. The
implemented FLF has been inspired from (Naso et al., 2006). The fuzzy
system inputs are u1, u2 and u3 while the fuzzy system output is K . For
instance, the vector Z = [ z1 z2 ⋯ z n ] is to be filtered, then the inputs,
in a specific moving window length of three samples, are given in Eqs.
(13)–(15)
u1 (j ) = zj−f
u2 (j ) = zj + 1−zj
u3 (j ) = zj + 2−zj + 1
where f represents the filtered sample of the last (j − 1)th moving
window
wj − 1 = [ zj − 1 ≡ f zj zj + 1]
These inputs will be transformed to linguistic variables. The eva-
luation of pre-determined fuzzy rules yields to obtain the membership
information K (j ) . Consequently, the filtered sample is obtained by
zjfiltered = f + K (j )(z j−f )
The output of the filter, z jfiltered , depends on the value of K (j ) .
For this work, the vector Z represents either the Q or T2.
3.2. Adaptive thresholds generation
The monitoring indices thresholds depend on the number of re-
tained principal components a and the number of observations (The
length of the sliding window) (refer to Eqs. (9)–(11)). However, with
the developed technique, these two parameters are kept unchanged.
The Q statistic threshold can be updated by using the linear projection
defined in equation Eq. (1). Within a selected window of fixed length,
the corresponding variances of the residual subspace from one time to
another serve to update the control limit. If wn is the moving window
∼ ⎩
over the data of length H and P ∈ Rm × (m − a) is the matrix of the ei-
genvectors that correspond to smallest eigenvalues, then the projected
data vector onto the residual subspace is given by
∼
Bj = wn∗Pj (18)
wn is first shifted and scaled using the mean and standard deviation
∼ ∼
values of the constructed PCA model. Pj is the jth vector in P where
j = 1, 2, ⋯, m−a
θi of Eq. (11) can be calculated by
m−a
θi = ∑ (var (Bj ))i i = 1, 2, 3
j=1
where var (·) stands for the variance value.
Using Eqs. (10) and (12) and after substituting the values of θi , the
new threshold of the index Q can be computed.
The control limit for the T2 statistic is updated by fitting the em-
pirical distribution over the window of the past evaluated values.
Fig. 1 represents the schematic diagram of the proposed adaptive
monitoring scheme with FLF.
In this figure, the offline phase includes two important tasks in
process monitoring which are the process training and the model va-
lidation. The process training is used to train the process and construct
Solar Energy 174 (2018) 762–769
a statistical model by determining the necessary parameters of the PCA
model (means, standard deviation, the number of principal components
and the initial thresholds). This task is performed on the training data-
set that is collected while the process is not exhibiting an abnormal
behaviour. Once PCA model is obtained, it is validated using a testing
data-set that is also acquired when the process is under its normal op-
eration. During the validation tasks, the window length, the FLF para-
meters (the membership function and the fuzzy rules) and the fixed
thresholds, that will be used when a fault is detected, are selected.
These parameters are determined according to the obtained FAR and
resulted in fault recovery delay for this data-set. The offline measure-
ments are taken from the inputs and the outputs of the GCPVS. At the
end of this phase, the adaptive thresholding scheme with FLF is ready to
monitor new testing samples. At each time a new observation is
available, this sample is shifted and scaled using the mean and standard
deviation of the constructed PCA model then its monitoring indices are
computed. These indices are passed through the implemented FLF to
(13)
obtain the filtered Q and T2 which will be fed to the decision block
(14) where they are compared with their initial thresholds. If the filtered
monitoring indices are greater than the initial thresholds, then the
(15) sample is considered to be faulty; as a result, an alarm is generated via a
binary variable, fault indicator. The adaptive thresholds are then set to
the predetermined fixed thresholds while the initial control limits are
set to the adaptive thresholds for the next sample monitoring. In the
(16) case where the filtered monitoring indices are both less than their as-
sociated initial thresholds, the fault indicator generates no alarm and
the adaptation mechanism will be activated to update the thresholds
where the new sample is included in the moving window and the oldest
one is excluded. The adaptation procedures are given by Eqs. (18), (19),
(17) (10), and (12).
The following case study represents a numerical example to illus-
trate how the adaptive thresholding scheme with FLF can be applied to
monitor multivariate processes.
3.3. Case study
Consider the following system
⎧ X1 = u1 + ε1
⎪ X2 = u1 + ε2
⎪
X3 = u1 + ε3
⎨
⎪ X 4 = u2 + ε4
⎪ X5 = u1 + 3u2 + ε5 (20)
εi is an independent white noise variable uniformly distributed over
[0.02, 2].
ui is a random variable which follows a normal distribution with
zero mean and 0.1 variance.
The training data-set is stored in the data matrix
X = [ X1 X2 X3 X 4 X5 ] ∈ R70 × 5 .
The mean and standard deviation values of the generated training
data-set are the vectors Mean ∈ R1 × 5 and Strd ∈ R1 × 5, respectively
(19) Mean = [0.9603 1.0724 0.9312 0.8945 0.8832] (21)
Strd = [0.6020 0.5859 0.5972 0.5943 0.6422] (22)
These two vectors are used to scale and shift the data matrix X to
zero mean and unity variance. The new matrix obtained throughout this
operation is denoted by Xs where its (5 × 5) covariance matrix can be
manipulated accordingly to Eq. (2) by two matrices Λ and P . These
matrices are respectively,
Λ = diag(1.3288 1.2142 1.0284 0.8500 0.5786) (23)
764
M. Ammiche et al. Solar Energy 174 (2018) 762–769

Fig. 1. The schematic diagram of the adaptive monitoring scheme with FLF.

Table 1 Using the parallel analysis criterion, three principal components are
Process description. retained. The corresponding eigenvectors of principal and residual
∼
Specifications Polycrystalline Monocrystalline Thin film subspaces P and P are given in Eq. (24).
The initial thresholds for 99% confidence level are
Number of modules 16 25 20
DC power per module (Wp) 125 75 135 Qthin = 7.9291 (25)
Voltage at max. power (V) 17.3 47.0 17.0
Current at max. power (A) 7.9 2.9 4.4 2
Tthin = 14.6636 (26)
DC/AC inverter capacity (W) 1600 1600 2500
Inverter max. DC voltage (V) 600 600 700
Inverter max. AC voltage (V) 220–240 220–240 220–240 A testing data-set, collected from the fault-free process, is employed
to validate this PCA model (Mean, Stdrd, number of principal compo-
nents), set the FLF parameters (the membership function parameters
⎡ ⎤ and the fuzzy rules), determine the appropriate length of the moving
⎢− 0.6019 0.4698 − 0.0454 0.1759 − 0.6197 ⎥ window and select the fixed thresholds. Suppose that the optimal
⎢− 0.0126 0.7517 0.3706 − 0.0384 0.5441 ⎥ window length that corresponds to the minimum FAR for the testing
P = ⎢− 0.3144 − 0.2752 0.6665 − 0.6049 − 0.1238 ⎥
⎢− 0.5199 − 0.3715 0.2382 data-set is 30 samples, then the last 30 observations of the training
0.6233 0.3829 ⎥
⎢ − 0.5181 − 0.0244 − 0.5997 − 0.4618 0.3976 ⎥ data-set initiate the window wn ∈ R30 × 5 . The new testing sample
⎢ ⏟ ⏟
∼ ⎥ xT = [1.9468 1.0006 2.0161 0.1681 1.6616] is scaled using the mean
⎣ P P ⎦ (24)
and standard deviation values. The standardized observation is

Monocrystalline Thin Film

Polycrystalline

Fig. 2. Solar panels.

765
M. Ammiche et al. Solar Energy 174 (2018) 762–769

Wind Speed Ambient

Module Sensor temperature
temperature sensor
sensor

Sensor box with

irradiance sensor

Poly-crystalline Mono-crystalline Thin film

PV module PV module PV module

INV 02 INV 03 Power injector

INV 01

AC line system

Local network Web box

Printer
CPU CPU
Computer 1 Computer 2

Fig. 3. Schematic diagram of grid-connected PV system.

Table 2 Testing Data-Set

Simulated Faults. Q Qth

Faults Fault description Number of Fault amplitude

faulty samples

Fault 1 Abrupt fault activated at sample 200 +1% 0

10
600
Fault 2 Additive random noise fault with 0 300 Between −15%
mean and 0.05 standard deviation and 17%
200 400 600 800 1000 1200 1400
activated at sample 103
T2 T2th
Fault 3 Drift fault activated at sample 700 500 From (0% to
with a slope of 1.5 × 10−4 9%)

Training Data-Set
0
Q Qth
10

1 200 400 600 800 1000 1200 1400

0.5 Fault Indicator
0
200 400 600 800 1000 1200 1400
0
Samples
10
Fig. 5. Testing data-set.
500 1000 1500 2000 2500
T2 T2th
xsT = [1.6388 − 0.1225 1.8165 − 1.2225 1.2120] (27)
The monitoring indices values associated with this observation are
Q = 6.1979 and T 2 = 2.1085. These values are passed through the de-
0
scribed FLF. The filtered values are then compared to the initial
10 thresholds. Suppose that this observation is normal, the adaptive
500 1000 1500 2000 2500 thresholds are generated as follows
1
0.5 Fault Indicator
0 1. Exclude the first observation from wn and include the new ob-
500 1000 1500 2000 2500
Samples servation.
2. Compute Bj ∈ R30 × 2 by Eqs. (18) and (24).
Fig. 4. Training data-set. 3. Compute the θi using Eq. (19) where i = 1, 2, 3.

766
M. Ammiche et al. Solar Energy 174 (2018) 762–769

Training Data-Set Fault 1

Q Qth Q Qth

0 0
10 10

500 1000 1500 2000 2500 200 400 600 800 1000 1200 1400
2 2 T2 T2 th
T T th

0 0
10 10
500 1000 1500 2000 2500 200 400 600 800 1000 1200 1400
Fault Indicator Fault Indicator
1 1
0 0
500 1000 1500 2000 2500 200 400 600 800 1000 1200 1400
Samples Samples

Fig. 6. Training data-set. Fig. 8. Detection of abrupt fault.

Testing Data-Set Fault 2

Q Qth Q Qth

0
10 0
10

200 400 600 800 1000 1200 1400 5 200 400 600 800 1000 1200 1400
10
T2 T2th T2 T2th

0
10 0
10
200 400 600 800 1000 1200 1400 200 400 600 800 1000 1200 1400
Fault Indicator Fault Indicator
1 1
0 0
200 400 600 800 1000 1200 1400 200 400 600 800 1000 1200 1400
Samples Samples
Fig. 7. Testing data-set. Fig. 9. Detection of random fault.

Table 3 Fault 3

FAR of adaptive thresholding scheme with FLF and PCA for the training and Q Qth

testing data-sets.
Method Adaptive thresholding scheme with FLF PCA

0
Statistics Q T2 Q T2 10

Training data-set 0 0.61 7.99 4.45 200 400 600 800 1000 1200 1400
Testing data-set 0 0 7.14 5.94 T2 T2th

The adaptive thresholds for the filtered Q will be computed by Eqs.

(10) and (12) after the θi substitution while the adaptive thresholds for 0
10
filtered T 2 will be computed empirically by estimating the distribution
200 400 600 800 1000 1200 1400
of a vector of the same length as the moving window wn where T 2 is Fault Indicator
included and the Hotelling of the first observation of wn is excluded. 1
0
The initial thresholds are set to the adaptive thresholds for the next 200 400 600 800 1000 1200 1400
testing sample. In the case where the observation is faulty, the obtained Samples
filtered monitoring values are greater than the initial thresholds, the Fig. 10. Detection of drift fault.
adaptive thresholds are set to the fixed control limits, an alarm is ac-
tivated by the fault indicator and the initial thresholds are set to the
adaptive thresholds for the next testing samples. At each new ob-
servation, the above steps are repeated.

767
M. Ammiche et al.
Real Abnormal Behaviour
Q Qth
0
10
500 1000 1500
T2
0 because the GCPVS is operating under its normal circumstances. To
10
500 1000 1500
Fault Indicator
1 Figs. 6 and 7. Form the fault indicator signal of these figures, it can be
0
500 1000 1500
Samples
Fig. 11. Detection of a real fault.
4. Application
4.1. Process description
The GCPVS from which the measurements are taken is installed at
the Power Electronic and Renewable Energy Research Laboratory
(PEARL) of Malaya University (Hasan and Mekhilef, 2017). This system
contains three inverters: INV 01, INV 02 and INV 03 which are con-
nected to 16 modules of polycrystalline silicon (POLY), 25 modules of
monocrystalline (MONO) and 20 modules of thin film, respectively.
Several sensors are installed within the system; these sensors measure
solar radiation, wind speed, panel and ambient temperatures, voltages
and currents at different stages of the GCPVS. This PV system is con-
nected to supply power to the laboratory building. The description of
the process is shown in Table 1 (Hossain et al., 2017). Fig. 2 represents
the three types of solar panels. The schematic diagram of the system is
shown in Fig. 3.
4.2. Application procedures
For monitoring purposes, two data-sets have been collected from the
GCPVS described above. The first data-set consists of m = 29 measured
variables and the n = 2766 observations with a sampling interval of 1
observation per 5 min. These data are used as a training data-set to
construct the PCA model (computing the mean, standard deviation, and
the number of PCs). These data are employed also to compute the initial
thresholds. Before applying PCA, the data have been processed where
all the observations exceeding μ ± 3σ are removed, such that μ and σ
represent respectively the mean and standard deviation values of each
measured variable. This data processing is required for constructing a
quite representative PCA model. The remaining data-set is the testing
data-set; it is used to validate the PCA model and appropriately de-
termine the length of the moving window as well as the FLF parameters
(the membership function parameters). This data-set consists of m = 29
measured variables and n = 1500 observations in which one sample is
recorded every 5 min. The fixed control limits, which are utilized when
a fault is detected, are determined by trial and error. The condition
under which these limits are selected is that they don’t affect the fault
recovery. To test the detection ability of the proposed method and
evaluate its monitoring performance, three types of faults have been
injected in the data: abrupt, random and drift faults. The simulated
faults description is shown in Table 2. 8 principal components have
been retained using parallel analysis technique. These components ex-
plain 79.53% of the total variance in the data. The best window length
is found to be 142 samples. The time required to complete the com-
putations and return the state of the monitored data for the proposed
Solar Energy 174 (2018) 762–769
algorithm is negligible.
4.3. Results and discussions
The resulted monitoring indices of the adaptive thresholding
scheme with FLF along with their corresponding adaptive thresholds for
2000 the GCPVS are shown in Figs. 4 and 5.
T2th
It can be seen from these figures that all the samples of the training
and testing data-sets are under the developed thresholds as it is de-
monstrated by the fault indicator signals. These results are obtained
illustrate the improvements achieved with the proposed method, the
2000 PCA technique is also applied to these data-sets. The monitoring indices
with their corresponding thresholds for the PCA are depicted in the
2000 seen that a large number of false alarms is resulted. These false alarms
affect the monitoring accuracy where the process status is not clearly
identified. The obtained FARs are tabulated in Table 3. The resulted
FAR in the Q is reduced to 0% for both training and testing data-sets.
The adaptive thresholding with FLF decreases also the FAR to 0% and
0.61% in the T2. According to the results, the detection accuracy has
been noticeably improved. These results confirm the ability of the
adaptive scheme with FLF to reduce the effect of false alarms in a
GCPVS monitoring.
The detection ability of the monitoring proposal has been tested on
simulated faults of Table 2. The results of the simulated faults detection
are shown in Figs. 8–10. The monitoring indices start violet the adap-
tive thresholds at the time index at which the fault is introduced sig-
nifying that an abnormality is occurred except for the drift fault, in
which the detection is delayed. This delay is due to the slope of the fault
that is taken very small to simulate the worst case. Q and T2 return
under the adaptive thresholds when the fault is disappeared. These
figures show that all the faults have been successfully and accurately
detected.
New data have been measured for online monitoring. The number of
observations of this set is n = 2465 sample recorded every 5 min. The
monitoring results are shown in Fig. 11. This figure illustrates that some
of the samples are above the adaptive thresholds which indicate the
occurrence of a fault. The fault indicator signal confirms the occurrence
of the abnormality since it switches from zero to one for some samples.
The occurrence interval of the fault, from this figure, appears to be
small; this is because the sampling interval is high.
5. Conclusion
In this paper, an adaptive monitoring scheme with fuzzy logic filter
has been applied to monitor a Grid-Connected Photovoltaic System
(GCPVS). The applied technique is able to detect different types of
faults that may occur in a GCPVS; hence it permits to increase the en-
ergy production with high quality. It serves also to avoid damages of
GCPVS equipment, economic losses and accidents. This method has
been presented and demonstrated how it can be used in monitoring. It is
a data-driven technique based on PCA and MWPCA that needs only
collected data from a GCPVS. The generated adaptive thresholds in-
crease the monitoring indices sensitivity besides the number of false
alarms reduction. The use of the filter involves also in the number of
false alarms reduction since it removes noise and outliers. The appli-
cation results show that no false alarm is signalled for the testing data-
set. Furthermore, a real fault has been successfully and accurately de-
tected.
References
Chiang, L.H., Russell, E.L., Braatz, R.D., 2001. Fault Detection and Diagnosis in Industrial
Systems. Springer, London.
Deng, X., Tian, X., Chen, S., Harris, C.J., 2017. Fault discriminant enhanced kernel
768
M. Ammiche et al.
principal component analysis incorporating prior fault information for monitoring
nonlinear processes. Chemom. Intell. Lab. Syst. 162, 21–34.
Dhimish, M., Holmes, V., 2016. Fault detection algorithm for grid-connected photovoltaic
plants. Sol. Energy 137, 236–245.
Elshenawy, L.M., Yin, S., Naik, A.S., Ding, S.X., 2010. Efficient recursive principal com-
ponent analysis algorithms for process monitoring. Ind. Eng. Chem. Res. 49, 252–259.
Fazai, R., Taouali, O., Harkat, M.F., Bouguila, N., 2016. A new fault detection method for
nonlinear process monitoring. Int. J. Adv. Manuf. Technol. 87, 3425–3436.
Ferahtia, J., Djarfour, N., Baddari, K., Kheldoun, A., 2012. A fuzzy logic-based filter for
the removal of spike noise from 2D electrical resistivity data. J. Appl. Geophys. 87,
19–27.
Firth, S., Klomas, K.J., Rees, S.J., 2010. A simple model of PV system performance and its
use in fault detection. Sol. Energy 84, 624–635.
Ge, Z., Chen, J., 2016. Plant-wide industrial process monitoring: a distributed modeling
framework. IEEE Trans Ind. Inf. 12, 310–321.
Harrou, F., Suna, Y., Taghezouit, B., Saidi, A., Hamlati, M.-E., 2018. Reliable fault de-
tection and diagnosis of photovoltaic systems based on statistical monitoring ap-
proaches. Renew. Energy. https://doi.org/10.1016/j.renene.2017.09.048.
Hasan, R., Mekhilef, S., 2017. Highly efficient flyback microinverter for grid-connected
rooftop PV system. Sol. Energy 146, 511–522.
Hassani, S., Saidur, R., Mekhilef, S., Taylor, R.A., 2016. Environmental and exergy benefit
of nano fluid-based hybrid PV/T systems. Energy Convers. Manage. 123, 431–444.
Hossain, M., Mekhilef, M., Danesh, M., Olatomiwa, L., Shamshirband, S., 2017.
Application of extreme learning machine for short term output power forecasting of
three grid-connected PV systems. J. Clean. Prod. 167, 395–405.
Kim, il. S., 2016. On-line fault detection algorithm of a photovoltaic system using wavelet
transform. Sol. Energy 126, 137–145.
Ku, W., Storer, R.H., Georgakis, C., 1995. Disturbance detection and isolation by dynamic
principal component analysis. Chemom. Intell. Lab. Syst. 30, 179–196.
Liu, X., Li, K., McAfee, M., Irwin, G.W., 2011. Improved nonlinear PCA for process
monitoring using support vector data description. J. Process Control 21, 1306–1317.
Lou, Z., Shen, D., Wang, Y., 2017. Two-step principal component analysis for dynamic
769
Solar Energy 174 (2018) 762–769
processes monitoring. Can. J. Chem. Eng. https://doi.org/10.1002/cjce.22855.
Madeti, S.R., Singh, S.N., 2017. Monitoring system for photovoltaic plants: a review.
Renew. Sustain. Energy Rev. 67, 1180–1207.
Naso, D., Scalera, A., Aurisicchio, G., Turchiano, B., 2006. Removing spike noise from
railway geometry measures with a fuzzy filter. IEEE Trans. Syst. Man Cybernetics 36,
485–494.
Ramahaleomiarantsoa, J., Heraud, N., Sambatra, E.J.R., Razafimahenina, J.M., 2012. On
the sensitivity of principal components analysis applied in wound rotor induction
machines faults detection and localization. Int. J. Energy Sci. 2, 262–271.
Rato, T.J., Reis, M.S., 2013. Defining the structure of DPCA models and its impact on
process monitoring and prediction activities. Chemom. Intell. Lab. Syst. 125, 74–86.
Rato, T.J., Reis, M., Schmitt, E., Hubert, M., De Ketelaere, Bart, 2016. A systematic
comparison of PCA-based statistical process monitoring methods for high-dimen-
sional, time-dependent processes. Process Syst. Eng., AIChE 62, 1478–1493.
Russell, E.L., Chiang, L.H., Braatz, R.D., 2000. Fault detection in industrial processes
using canonical variate analysis and dynamic principal component analysis.
Chemom. Intell. Lab. Syst. 51, 81–93.
Shang, L., Liu, J., Zhang, Y., 2017. Efficient recursive kernel canonical variate analysis for
monitoring nonlinear time-varying processes. Can. J. Chem. Eng. https://doi.org/10.
1002/cjce.22897.
Sheriff, M.Z., Mansouri, M., Karim, M.N., Nounou, H., Nounou, M., 2017. Fault detection
using multiscale PCA-based moving window GLRT. J. Process Control 54, 47–64.
Silvestre, S., López, L.M., Kichou, S., Pacheco, F.S., Pumar, M.D., 2016. Remote super-
vision and fault detection on OPC monitored PV systems. Sol. Energy 137, 424–433.
Spataru, S., Sera, D., Kerekes, T., Teodorescu, R., 2015. Diagnostic method for photo-
voltaic systems based on light I-V measurements. Sol. Energy 119, 29–44.
Wold, S., Trygg, J., Berglund, A., Antti, H., 2001. Some recent developments in PLS
modeling. Chemom. Intell. Lab. Syst. 58, 131–150.
Zhong, N., Deng, X., 2016. Multimode non-Gaussian process monitoring based on local
entropy independent component analysis. http://doi.org/10.1002/cjce.22651.
Zhu, J., Geand, Z., Song, Z., 2017. Distributed parallel PCA for modeling an monitoring of
large-scale plant-wide processes with big data. IEEE Trans. Ind. Inf. 13, 1877–1885.