Computers and Electrical Engineering 120 (2024) 109763

Contents lists available at ScienceDirect

Computers and Electrical Engineering

journal homepage: www.elsevier.com/locate/compeleceng

An intelligent ANFIS-based fault detection, classification, and

location model for VSC-HVDC systems based on hybrid PCA-DWT
signal processing technique
Ehsan Akbari a, Milad Samady Shadlu b,*
a
Department of Electrical Engineering, Mazandaran University of Science and Technology, Babol, Iran
b
Young Researchers and Elite Club, Bojnourd Branch, Islamic Azad University, Bojnourd, Iran

A R T I C L E I N F O A B S T R A C T

Keywords: Designing an accurate model for fault detection, classification, and location is vital from the
VSC-HVDC system protection viewpoint. The adaptive neuro-fuzzy inference system (ANFIS) is a commonly used
Adaptive neuro-fuzzy inference system (ANFIS) learning-based fault location model that performs independently from the propagating wave
Harris hawks optimization (HHO) algorithm
characteristics, whose performance can be improved by optimizing the membership functions
Cuckoo search (CS) algorithm
Principal component analysis (PCA)
associated with the inputs. Accordingly, two metaheuristic algorithms with the quick-search
Discrete wavelet transform (DWT) capability of the population space, i.e., Harris Hawks optimization (HHO) and cuckoo search
(CS) optimization algorithms, are used in this paper to optimize the ANFIS, and their perfor­
mances are compared with the traditional ANFIS training algorithms. Moreover, principal
component analysis (PCA) is employed to detect the fault, and the discrete wavelet transform
(DWT) strategy is exploited to acquire the ANFIS training and testing dataset according to the
statistics T2 obtained by PCA. Three statistical indices, i.e., mean value, standard deviation, and
norm entropy, are computed corresponding to the extracted wavelet coefficients from the current
signal and applied to train and test the ANFIS. Optimized ANFIS conducts fault classification and
location tasks, and the accuracy of the proposed model is compared with the commonly used
traveling wave (TW) –based models and recently proposed fault location methods in the litera­
ture. Three fault types on the DC-link with three fault resistances are examined to confirm the
fault classification superiority of the proposed model and its fault type/impedance-independent
estimation capability. An accuracy rate of 99.995% is obtained for the fault locating task,
while fault detecting and classifying are accomplished with an accuracy of 100%. Simulations and
numerical studies are performed in MATLAB software.

1. Introduction

High voltage direct current (HVDC) transmission networks are vital sectors in the power transmission system, which have many
complications from the point of view of protection and control [1]. The optimal protection performance in these systems depends on
accurate and timely identification of the occurrence of the fault, the type of fault, and the location of the fault [2]. For this purpose,
intelligent methods have been introduced to detect, classify, and locate faults in HVDC networks [3]. The fundamental principles of
these methods are based on the measurement of the voltage signal [4] or current [5] of the transmission line so that the fault is

* Corresponding author at: Islamic Azad University, Bojnourd Branch, Bojnourd, North Khorasan Province, 1651153311, Iran.
E-mail addresses: e.akbari@ustmb.ac.ir (E. Akbari), m.samady@bargh-project.ir (M. Samady Shadlu).

https://doi.org/10.1016/j.compeleceng.2024.109763
Received 9 March 2024; Received in revised form 27 July 2024; Accepted 1 October 2024
Available online 10 October 2024
0045-7906/© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

identified based on the distinctive features that this signal exhibits during different fault conditions [5]. The fault signal can be
analyzed both in the frequency and time domains; hence, fault detection methods are divided into two general categories: frequency
domain [6,7], and time domain [8]. From the point of view of the data acquisition method, fault detection techniques in the power
network are divided into three classifications: model-based, signal-based, and knowledge-based approaches [9]. In model-based
methods, an equivalent model (EM) of the transmission system is constructed, and the parameter changes of the EM are evaluated
under faulty conditions to detect and locate faults [10,11]. The main disadvantage of these methods is the high dependence of the fault
detector on the system model and their complicated modeling procedure for complex and large power systems. In addition, the de­
pendency of these methods on the system model reduces their accuracy.
Signal-based methods perform fault detection and location by processing the measured signal in the time or frequency domain [12].
The conventional traveling wave method (TW) is a fault detection and location method based on high-frequency measurements of the
propagated signal (wave) from the fault point and determining its arrival time [13]. The main drawback of the TW-based methods is
the dependency of the wave propagation speed on frequency disturbances [14]. Nevertheless, TW-based methods have been used in
many studies associated with fault diagnosis and localization [15-18]. On the other hand, the accuracy of TW-based methods is
intensely affected by the signal sampling rate. Accordingly, high-resolution digital signal processing methods must be used to detect
the arrival time of wavefront [19]. This shortage increases the computational burden of the TW-based methods because the charac­
teristics of the wavefront must also be extracted in addition to the traveling wave processing. Signal processing methods are divided
into three categories: time domain, frequency domain, and time-frequency domain [20]. The first includes tools that perform signal
processing in the time domain, such as the Hilbert-Hang transform [21-24], the Intrinsic Time Decomposition (ITD) method [25],
sliding time window [26] and time-domain convolution power ratio [27] method. The main disadvantage of time-domain signal
processing methods is the high sensitivity of the processing result to noise and external disturbances. The second category includes
techniques that analyze the frequency spectrum of the fault signal [28], such as discrete Fourier transform [29] and Lomb-Scargle
Periodogram (LSP) [30]. Due to the non-stationary nature of DC faults in the HVDC system, frequency domain methods do not
perform satisfactorily. To overcome this shortcoming, time-frequency domain methods have been introduced for processing the fault
signal, and two wildly used methods in this category are short-term Fourier transforms (STFT) [31] and S transform (ST) [32]. The
major shortcoming of S transform-based methods is the limitation of time-frequency resolution; thus, improved versions of these
methods have been proposed in the literature to provide higher resolution [33]. Gabor transform (GT) is another time-frequency
domain method that has been used for fault detection/classification [34] and fault location [35] in the transmission system.
Despite the high speed of GT in extracting the arrival times of the first peak wave, the accuracy of this method decreases significantly
with the increase of fault inception time (FIT) [35]. One of the widely used signal processing techniques in the time-frequency domain
is the wavelet transform (WT), which has been widely exploited to process the fault signal in HVDC networks [36-38]. The WT-based
signal processing is implemented in both continuous and discrete modes. Continuous Wavelet Transform (CWT) is an effective method
for processing signals with variable frequency content over time; thus, it is a practical tool for analyzing non-stationary signals [39].
However, CWT is principally utilized for offline analysis, and it is not practical for online fault detection and power grid protection
purposes [40]. In contrast, discrete wavelet transform (DWT) has been widely used in DC fault detection applications [41-45].
As already stated, signal-based methods suffer from a long computational process and are difficult to implement for complex
networks. For this purpose, knowledge-based methods have been proposed for detecting, classifying, and locating faults in power
transmission systems, the most common of which are methods based on neural networks [46-48], machine learning algorithms [49-51]
and knowledge graph platform [52,53]. An effective and efficient method for detecting, classifying, tracking, and locating faults in the
power system is the adaptive neuro-fuzzy inference system (ANFIS), which has been rarely used to detect and locate faults in HVDC
systems [54-56], despite that the superior performance of this method has already been proven in other applications such as fault
detection in photovoltaic systems [57,58]. However, traditional ANFIS training and testing methods such as LS-BP, LSE, and GD
exhibit poor search performance, which leads to significant attenuation of the estimation accuracy [59,60]. In addition, the tradi­
tionally trained ANFIS has two limitations: (i) optimal membership functions must be determined, and (ii) only influential training
data must be applied as input; otherwise, the estimation error increases [61]. To deal with the first limitation, heuristic and
meta-heuristic algorithms have been proposed for ANFIS optimization and determination of the optimal parameters for the premise
and consequence layers [62], such as the Aquila optimization algorithm [63], particle swarm optimization [64], gray wolf optimi­
zation (GWO) and the particle swarm optimization (PSO) [65], genetic algorithm [66], cuckoo search (CS) algorithm [67], and
non-dominated sorting genetic algorithm type 2 (NSGA-II) [68]. In addition, feature extraction methods have been introduced to deal
with the second limitation, the most widely used of which is DWT [69]. In [68], the DWT technique with db-2 mother wavelet is used
to extract influential features from the fault signal with the assumption of 6 wavelet levels, and statistical indices are calculated for
each DWT level separately. This assumption increases the dimensions of the extracted feature matrix because as the number of wavelet
levels increases, the wavelet coefficients practically tend to zero and do not provide valuable information about the fault. Furthermore,
detecting the faults directly using the wavelet coefficients is impractical. For this purpose, an additional ANFIS layer is needed, which
is trained based on statistical indicators calculated in normal and faulty conditions.
In order to address the above shortcomings, the principal component analysis-based discrete wavelet transform (PCA-DWT)
method is used in this paper to generate a dataset for ANFIS training and testing. The PCA calculates the statistics T2 from the current
signal and accordingly performs fault detection in the first phase. In this way, the need for a separate ANFIS layer for fault detection is
eliminated. Next, statistics T2 obtained by the PCA are applied to the DWT instead of the original signal since this statistic includes
detailed information about the fault occurrence time, fault type, fault distance, and faulty pole. A sequence of detail wavelet co­
efficients is obtained for each fault scenario, and the corresponding statistical indices are calculated. Next, these indices are used as
input for ANFIS training and testing. This process will reduce the dimensions of the feature matrix and improve the effectiveness and

2
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig. 1. The proposed ANFIS-based fault detection, classification, and location model.

Fig. 2. A three-input ANFIS structure.

Fig. 3. The design procedure of the ANFIS according to the membership functions.

3
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Algorithm 1
Pseudo-code of Harris Hawk Optimization algorithm.
Inputs: The population size, N, and maximum number of iterations, Itermax.
Outputs: The position of the prey and its fitness value.
1. Determining the random population Xi (i = 1, 2,…, N).
2. Repeat the following steps until the stop condition is met:
Define Xrabbit as the prey position.
For each hawk (Xi), repeat the following steps:
Initial energy updating based on the following equation:
E0 = 2rand(.) - 1
Update E according to (1)
If |E| ≥1, then the following steps are repeated:
Update position vector according to (2)
If |E|<1, then the exploitation phase is performed
3. End

Algorithm 2
Pseudo-code of Cuckoo Search Optimization algorithm.
Inputs: The population size and maximum number of iterations
Outputs: The global best solution (nest) and its fitness value.
1. Determine the objective function f(X), X=(x1,…,xd)T
2. Production of the initial population corresponding to n host nests, Xi (i=1,...,n)
3. Repeat the following steps until t < Maximum Iteration:
Randomly select a cuckoo based on the Levy flight criterion.
Calculate the fitness function Fi for the selected cuckoo.
Randomly choose nest α among n nests and calculate Fj.
If Fi > Fj,
Replace j with a new solution.
Calculate the probability Pα for nest j.
Choose nests with the lowest Pα as the best solution.
Determine the optimal solution among the set of solutions during the discarding phase.
4. Apply the mutation phase to the new generation.
5. Select the global best solution according to the fitness function value.
7. End

Table 1
Statistical measures calculated for ANFIS training [79].
Feature Mathematical formulation

Mean 1 ∑N
μi = cij
N j=1
Standard Deviation ( )1
2
1 ∑N ( )2
σi = cij − μi
N j=1

∑N
Norm Entropy NEi = cP P ≥1
j=1 ij

speed of ANFIS training. Another advantage of the PCA method is that it can be used online and offline for fault detection, and it
provides distinct statistics for each fault condition, which means that this method can be used under noisy and transient fault con­
ditions [70]. For a better comparison, the time-domain signal processing method based on the Hilbert-Hung transform (HHT) is also
implemented to extract the features, and its performance is compared with the proposed PCA-DWT technique. Since the accuracy of
the fault locator depends on the features extracted from the fault signal for ANFIS training, the performance of feature extraction
patterns can be easily compared and evaluated. In addition, an effective meta-heuristic algorithm with fantastic searching capability
and excellent performance in escaping the local optimality trap called the Harris Hawk Optimization (HHO) algorithm is used for
ANFIS optimization. In addition, the Cuckoo search algorithm (CSA) is applied for ANFIS training and testing, and its performance is
compared with the HHO algorithm in terms of convergence speed and minimization of the objective function. Numerical analysis is
performed for four scenarios (two ANFIS optimization scenarios and two feature extraction strategies), and the results are compared.
Besides, a comparative study with other fault detection, classification, and location methods presented in the literature is conducted.
The proposed model based on optimized ANFIS can successfully classify faults in addition to accurate fault distance estimation. For
fault classification, classified data corresponding to different fault scenarios with unique identification codes (ICs) are applied for
ANFIS training and testing.
Unlike the TW-based methods, the proposed strategy in this paper does not require any time-domain information about the fault, e.
g., arrival time corresponding to the wavefronts, wave velocity, etc. In addition, any disturbances can affect the accuracy of the TW-
based methods, due to the dependency of the algorithm on the signal processing technique. On the other hand, the majority of the fault

4
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig 4. Hilbert-Huang transform-based feature extraction flowchart.

classification strategies proposed in the literature are based on conventional schemes, e.g., decision tree (DT) and support vector
machine (SVM) [16,18,22], which are threshold-dependent approaches. This implies that the thresholds must be changed by varying
the fault characteristics (e.g., type, impedance, location, etc.). This feature limits the applicability of the fault classifier particularly for
online protections. However, no thresholds are integrated into the proposed ANFIS-based classifier in this paper, and solely the
extracted features from the single-end measured currents are exploited to train and test the classifier.
The innovations and advantages of this study can be stated as follows:

• The proposed hybrid feature extraction model (PCA-DWT) can detect the faults in the first stage without any additional fault
detection layer in the ANFIS structure.
• The dimension of the extracted features matrix is reduced by calculating the wavelet coefficients from the PCA statistics instead of
the original signal.
• Only one endpoint signal measurement is required for executing the proposed model; hence, the number of measuring devices (e.g.,
current transformers) will be reduced.
• A comparison study is performed between two signal processing techniques belonging to the time domain (HHT) and time-
frequency domain (PCA-DWT) techniques.
• A knowledge-based, wave-velocity independent method with training capability and high accuracy for detecting, classifying, and
locating faults is proposed, which can be used for all fault scenarios during one training step.

The paper is organized as follows. In Section 2, the proposed model is introduced. This section describes the ANFIS optimization
algorithms and the feature extraction techniques. In Section 3, the studied VSC-HVDC system is described, and various faults under
three scenarios (PPTG, NPTG, and PTN) corresponding to the three fault resistances are simulated to generate the initial signals. Then,
the ANFIS input dataset is generated using the feature extraction techniques. The results of the simulations and comparative studies are
also presented and evaluated in Section 3. Section 4 is devoted to the conclusions of the study.

2. The proposed model

The general structure of the proposed fault detection, classification, and location model is shown in Fig. 1. The proposed model

5
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig. 5. The DC-link current waveforms measured on the rectifier side for faults that occurred at 10 km from the measurement point for (a) PPTG, (b)
NPTG, and (c) PTN fault scenarios.

consists of three main blocks. The first block shows the VSC-HVDC system under study, the second is the fault detector and the feature
extractor, and the third one is an ANFIS-based fault classifier and locator. In the following subsections, the ANFIS structure and
optimization algorithms are described.

2.1. Adaptive neuro-fuzzy inference system (ANFIS)

The schematic illustration of a three-input ANFIS structure is shown in Fig. 2. Two design parameters, Antecedent (premise) and
Conclusion (consequence), are used to create a logical relation between the inputs and outputs of the model. These two parameters are
determined according to the membership functions (MFs); hence, they can be appropriately selected by optimizing the MFs [71]. The
design procedure of the ANFIS according to the membership functions is illustrated in Fig. 3. It is noteworthy that the Takagi-Sugeno
(T-S) fuzzy inference system (FIS) is used in this study. The ANFIS layers and the mathematical representation of each layer’s output
are described in [72]. The Takagi-Sugeno fuzzy inference system exhibits a more stable behavior than the Mamdani FIS since its output
MFs are crisp constant values. Furthermore, the Takagi-Sugeno FIS is more accurate and faster than the Mamdani FIS [73].

2.2. Optimization algorithms

The Harris Hawk optimization (HHO) and Cuckoo search (CS) are nature-inspired metaheuristic algorithms that have exhibited
excellent performance in solving optimization problems. Both of these algorithms are applied to optimize the ANFIS in this study.

a) Harris Hawk Optimization Algorithm

HHO is inspired by the behavior of the Harris Hawk in nature to search for food [74]. In this algorithm, several hawks simul­
taneously attack a prey, usually a rabbit, to confuse it. This step of the algorithm is called the exploration phase. Based on the
behavior of the prey, the hawks change the strategy of attacking the prey to eventually overcome it. This step is also called the
exploitation phase. According to the prey energy defined in (1), this algorithm switches between the exploration and exploitation
phases.

6
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig. 6. The T2 statistics corresponding to the fault currents depicted in Fig. 4, (a) PPTG, (b) NPTG, and (c) PTN.

( )
itre
E = 2E0 1 − (1)
itremax

where, E indicates the prey energy in each iteration, itermax denotes the stopping criteria of the algorithm, E0 represents the initial
prey energy that changes between [-1,1], and iter symbolizes a counter for each iteration of the algorithm. |E|≥1 implies that the
hawks must look elsewhere for prey (exploration phase). In other words, according to (2), the hawk’s position is updated.
Conversely, if |E|<1, the algorithm enters the exploitation phase. In (2), if q indicates the probability of an individual hawk being in
position t relative to the prey, then the next position of the hawk will depend on the positions of the other members and the prey.
⎧
⎨
X(t + 1) = xrand (t) − r1 |xrand (t) − 2r2 X(t)| if q ≥ q∗ (2)
⎩
xprey (t) − xm (t) − r3 (LB + r4 (UB − LB)) if q < q∗

where, X(t+1) denotes the vector containing the position of the hawks in the next iteration, Xprey(t) indicates the position of the
prey, X(t) represents the current position of the hawks, r1, r2, r3, and r4 are random numbers between (0,1) which are updated in
each iteration, and LB and UB signify the lower and upper bounds of the variables, respectively. q indicates a random number in the
range of [0,1], and q* denotes its threshold value (which equals 0.5 in most cases). Xrand(t) is a randomly selected hawk from the
current population vector, and Xm is the average position of the hawks in the current population. The pseudo-code of the HHO
algorithm is presented in Algorithm 1 [74]. First of all, an initial set of solutions (initial population of hawks) is generated. In the
second step, the values of E0 and E are updated for each hawk, and the exploration or exploitation phases are performed according
to the value of E. This process continues as long as the stopping condition is met. Next, the position of the prey in the last iteration is
introduced as the optimal solution to the optimization problem.
b) Cuckoo Search Optimization Algorithm
The use of the CS algorithm in the field of optimization was proposed by Yang and Deb in 2009 [75]. This algorithm is based on
the brood-parasitic behavior observed in some cuckoo species. This type of bird performs a Levy random flight to find the desti­
nation nest for laying eggs, which is the exploration stage in the search space. Levy flight is expressed mathematically using (3):

7
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig. 7. The wavelet coefficients corresponding to the statistics T2 depicted in Fig. 5, (a) PPTG, (b) NPTG, and (c) PTN.

Levy(λ) = t− λ , 1 < λ ≤ 3 (3)

where, λ denotes the mathematical expectation of cuckoo or nest selection at time step t.

The following three rules are valid for the cuckoo search algorithm:

• Each cuckoo lays one egg at time t and places it in a randomly selected nest.
• The best nest containing the highest quality eggs is selected and introduced as the optimal solution.
• The number of host nests is fixed. Also, the probability of finding cuckoo eggs by the host bird is represented by Pαϵ(0,1).

The CS algorithm provides two sets of solutions during two stages, i.e., Levy flight and discarding stages [76]. By choosing the
suitable nest, the promising solution in each nest is selected based on the conditions defined in (4). Within nest s, only one solution
(egg) among two promising solutions, Xs and Ys, is kept and the other is discarded.
{
Ys , if FF(Ys ) > FF(Xs )
Zs = (4)
Xs , otherwise

where, FF(.) denotes the fitness function corresponding to the promising solution. Next, the mutation operator is applied to the solution
Zs to improve its quality as follows:
{
Zs + rand(Zr1 − Zr2 ), if rands > P
Us = (5)
Xs , otherwise

where, rands denotes a random value in the range of (0,1), P indicates the mutation probability, and Zr represents the randomly
selected solutions from the current population. The discarding phase is repeated to retain the promising solution between Zs and Us,
and the solution with the highest fitness value is picked up. Based on these three rules, the pseudo-code of the Cuckoo search algorithm
is presented in Algorithm 2 [76].

8
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig. 8. The extracted features using PCA-DWT, (a) Rf=1Ω, (b) Rf=10Ω, and (c) Rf=100Ω.

Table 2
Parameters of PCA.
Parameter Symbol Value

Number of principal components (PCs) a 2

Number of sampled data n 1500
Level of significance α 0.1
Critical value of the F-distribution F(a,n-a,α) 2.30259
Control limit of T2 T2α ~5

2.3. Feature extraction methods

a) Principal Component Analysis based Discrete Wavelet Transform (PCA-DWT)

In this paper, a PCA-DWT method is proposed to extract features from the current signals. Furthermore, PCA can detect the fault
occurrence and reduce the dimension of the fault signals matrix. In addition, PCA can detect any fault under unlimited conditions,
e.g., severe noisy, simultaneous, and transient fault conditions [70]. T2 is a commonly used index to detect any deviations from
normal conditions. Suppose Xn×m is the input matrix including sampled data, where m denotes the number of variables and n

9
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig. 9. The empirical mode decomposition results for the PPTG fault scenario with Rf=100Ω at a distance of 10 km.

Fig. 10. (a) Original signal, (b) real and imaginary parts of the original signal, (c) the Fourier spectrum corresponding to the real part, and (d) the
Fourier spectrum corresponding to the imaginary part, for the PPTG fault scenario with Rf=100Ω at a distance of 10 km.

indicates the number of observations. T2 can be determined using the eigenvectors and eigenvalues of the covariance matrix of
Xn×m, as follows:

T2 = xi Aλ− 1 AT xTi (6)

where, A denotes a matrix consisting of eigenvectors and λ represents a vector comprising the eigenvalues of the covariance matrix
of Xn×m, respectively. xi indicates a vector that includes the variables sampled during the ith observation. Any deviations from

10
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig. 11. The extracted features using Hilbert-Huang transform, (a) Rf=1Ω, (b) Rf=10Ω, and (c) Rf=100Ω.

Fig. 12. (a) Initial and (b) optimal membership functions obtained using HHO and CS algorithms.

11
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig. 13. The convergence characteristics of the optimization algorithms, (a) Cuckoo Search algorithm, (b) Harris Hawk optimization algorithm.

Table 3
Identification codes corresponding to the fault cases.
Fault Case Label Identification Code (IC)

PPTG Rf=1Ω PPTGR1 1

Rf=10Ω PPTGR10 2
Rf=100Ω PPTGR100 3
NPTG Rf=1Ω NPTGR1 4
Rf=10Ω NPTGR10 5
Rf=100Ω NPTGR100 6
PTN Rf=1Ω PTNR1 7
Rf=10Ω PTNR10 8
Rf=100Ω PTNR100 9

normal conditions can be detected by comparing the T2 with a threshold value named control limit, which is determined according
to the F-distribution as follows [77,78]:
a(n − 1)
Tα2 = Fa,n− a,α (7)
n− a

where, a denotes the number of principal components (PCs), n represents the number of sampled data, α indicates the level of
significance, and F(a,n-a,α) symbolizes the critical value of the F-distribution. In the DWT method, the input signal, xi, decomposes
into detail and approximate components with different resolution levels called wavelet coefficients and represented by ci. The
Daubechies (db) mother wavelet is used in this paper. Daubechies mother wavelets have been widely utilized in fault detection
applications due to their high resolution and signal traceability in patterns where the initial signal varies severely [69]. Due to the
high sampling resolution of the fault current signal, the DB mother wavelet is used in this study.
First, the current signal is measured at the rectifier side and T2 statistics are calculated by applying PCA to the sampled signals.
Comparing the T2 by the pre-defined threshold specifies the fault occurrence. Next, the obtained T2 statistics are input to the DWT
instead of the original current signals, and the three DWT features listed in Table 1 are calculated using the wavelet coefficients. In
this study, 24 fault signals (corresponding to 24 distances) are recorded for three fault resistances (Rf=1Ω, 10Ω, 100Ω) under PPTG,
NPTG, and PTN fault scenarios. Since we have three fault cases with three fault resistances, nine 24 × 3 matrices can be obtained,
resulting in 648 features, where 453 quantities are used for training the ANFIS and the rest are used for testing. To better display the
extracted features, all amounts are normalized. In Table 1, cij indicates the wavelet coefficients corresponding to the ith fault case
and jth fault resistance.

b) Hilbert-Huang Transform
To verify the superiority of the PCA-DWT-based extracted features, the time domain-based method used in [21], called
Hilbert-Huang (HH) transform, is also applied to extract the features from the original signal, and its effectiveness is compared with
the proposed PCA-DWT method. The comparison criterion is the accuracy rate of classifying and locating the faults. Fig. 4 shows the
HH transform process to extract the features from the current signal, where the empirical mode decomposition (EMD) is applied to

12
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig. 14. (a) training and (b) testing results of the ANFIS-based HHO-PCA-DWT fault classifier.

Table 4
Statistical analysis of the ANFIS-based fault classifiers.
ANFIS-based Classifier MSE RMSE R Accuracy%

Training Testing Training Testing Training Testing Training Testing

CS-HHT 0.25689 1.10965 0.50685 1.05340 0.98050 0.91360 96.14 95.06

HHO-HHT 0.04946 0.44855 0.22240 0.66974 0.99632 0.96420 96.28 95.83
CS-PCA-DWT 0.03042 0.05670 0.17442 0.23812 0.99774 0.99561 97.53 97.37
HHO-PCA-DWT 0.00022 0.00039 0.01509 0.01999 0.99998 0.99298 98.67 98.23

the original signal to obtain the intrinsic mode functions (IMFs). Next, instantaneous amplitude and frequency corresponding to
each IMF are derived by employing the Hilbert transform. Furthermore, each IMF can be decomposed into real and imaginary parts.
Then, by averaging the real (RE) and imaginary (IM) graphs over the IMF numbers (N), one characteristic curve for both RE and IM
can be derived. Next, three statistical indicators of Table 1 are calculated according to each pair of RE and IM. Since 216 fault
signals are evaluated considering “3 fault types × 3 fault resistances × 24 fault distances”, and a pair of RE-IM is derived for each
signal, 1296 independent quantities are calculated utilizing the HH transform. In the EMD process, the stopping criterion is met
when the residue (r(t) in (8)) becomes a monotonic function, i.e., no more intrinsic mode function can be obtained from the original
signal.
r(t) = X(t) − IMFi (t) (8)

where, X(t) denotes the original signal, and IMFi(t) indicates ith intrinsic mode function extracted by EMD. The HH transform tech­
nique is beyond the scope of this paper; hence, interested readers are referred to [22] for further theoretical details.

13
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig 15. Block diagram of the proposed fault detecting, classifying, and locating strategy.

3. Simulation results

In this section, a 250 km VSC-HVDC system (Vnominal=230 kV Snominal=2000 MVA) is implemented in the MATLAB/Simulink
R2018b and three fault cases, i.e., PPTG, NPTG, and PTN, are simulated with three fault resistances, i.e., Rf=1Ω, 10Ω, 100Ω at 24
distances, i.e., d=10 km, 20 km,…, 240 km. The simulation and sampling times are 1.5 sec and 10–3 sec, respectively, so it can be
obtained 1500 sampling points. It is worth mentioning that the faulty mode considered in this paper is an outage caused by the
permanent fault on the DC transmission line [80]. For example, the positive and negative DC-link currents measured on the rectifier
side for nine fault cases that occurred at a distance of 10 km from the measurement point at t=7.5 s are plotted in Fig. 5.

3.1. Feature extraction task

In order to evaluate the performance of the proposed PCA-DWT-based feature extraction technique, the fault cases shown in Fig. 5
are applied to PCA-DWT and conventional HH transform to extract the features used for ANFIS training. At first, the T2 statistic is
calculated for each fault case according to (6), as illustrated in Fig. 6. Next, the DWT is employed to generate the wavelet coefficients
for each fault case and calculate the statistical indices. Fig. 7 shows the wavelet coefficients, and the obtained feature matrices cor­
responding to each fault case are depicted in Fig. 8. The PCA parameters are given in Table 2.
Furthermore, the Hilbert-Huang transform technique is applied to extract the features from the current signals shown in Fig. 5. For
instance, the intrinsic mode functions (IMFs) and residue corresponding to the PPTG fault scenario with Rf=100Ω at a distance of 10
km are plotted in Fig. 9. Then, by calculating the RE and IM quantities, as shown in Fig. 10, a Fourier spectrum can be produced for
each graph. The same procedure is executed for each current signal in Fig. 5 for 24 distances, and the ultimately extracted feature
matrices are drawn graphically in Fig. 11. In this figure, subscripts RE and IM indicate the statistics extracted from the real and
imaginary quantities, respectively.

3.2. ANFIS optimization task

In this step, two optimization algorithms, i.e., HHO and CS, are used to obtain the optimal membership functions of ANFIS. In this
paper, Gaussian MFs, as defined in (9), are adopted to model ANFIS inputs. Indeed, the optimization algorithms are applied to generate

14
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig 16. (a) training and (b) testing results of the CS-PCA-DWT fault locator.

the optimal center (c) and width (b) for each membership function.
{ ( )2 }
( ) − xj − cij
μij xj = exp (9)
b2ij

where,cij and bij represent the center and width parameters of Gaussian MFs corresponding to the ith (i=1,2,…,m) fuzzy rule and jth
(j=1,2,...,n) fuzzy variable, in which m and n denote the number of rules and input variables, respectively. The optimization procedure
to achieve the optimal MFs is described in detail in [81].
In the optimization problem, the objective function is the mean squared error (MSE), which can be calculated as follows,
m
1 ∑
MSE = y i )2
(yi − ̂ (10)
m i=1

where, yi and ̂
y i represent the actual and estimated values of the output of rule i, respectively, which is formulated as follows:

yi = x T π i (11)
[ ]
where, x = [x1x2⋅⋅⋅xn] indicates the input variables vector, and πi = p1i p2i ⋯ pni (i=1,2,…,m) denotes the conclusion vector of ith
fuzzy rule. The actual output of the ANFIS is the pre-defined “Target” dataset. Meanwhile, the estimated output is calculated according
to the procedure illustrated in Fig. 3 starting with initial MFs.
It is worth mentioning that the MSE exhibits faster convergence in optimization problems where the error is minor. Since the input
data are normalized values, the main limitation of the MSE in handling the outliers would be eliminated [82]. Eventually, the opti­
mized MFs using two optimization algorithms are depicted in Fig. 12. In this process, the optimal MF design parameters are obtained
by minimizing the objective function defined in (10) by employing two algorithms. The convergence characteristics of the optimi­
zation algorithms during five runs are depicted in Fig. 13. The performance of the HHO algorithm in solving engineering optimization
problems is promising due to the use of the exploitation and exploration phases. However, it is worth pointing out that the HHO

15
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig 17. (a) training and (b) testing results of the HHO-PCA-DWT fault locator.

algorithm has a critical shortcoming, i.e., it falls into stagnation before reaching the global optimum. However, this challenge can be
addressed by choosing the appropriate threshold value (q*) to update the position of the hawk relative to the prey. It is worth pointing
out that this shortcoming does not exist in the CS algorithm; hence, another way to improve the performance of the HHO algorithm is
to combine the CS algorithm to improve the exploitation and exploration processes [83]. On the other hand, this combination increases
the execution time and complexity of the algorithm. To deal with this problem, we improve the execution speed of the HHO algorithm
in this article by accelerating the hawk position updating phase. This task is accomplished by defining the variable q as a Gaussian
function and bounding the Gaussian RMS width (parameter c).

3.3. Fault classification task

In this subsection, the fault classification procedure is explained, and the results are analyzed. In this paper, nine fault cases are
studied (3 fault types associated with 3 fault resistances). Hence, 9 identification codes (ICs) are required to train and test the ANFIS, as
given in Table 3. As described in Section 3.2, the optimized Gaussian MFs are derived corresponding to each identification code and
used to train and test the ANFIS. The training and testing results for the proposed HHO-PCA-DWT model are shown in Fig. 14. The same
procedure is executed for three other ANFIS-based classifiers, i.e., HHO-HHT, CS-PCA-DWT, and CS-HHT, and the results are compared
in Table 4. Conventional statistical measures are employed to compare the performance of the ANFIS-based classifiers during the
training and testing stages. These statistics include Mean Squared Error (MSE), Root Mean Square Error (RMSE), Regression (R), and
Accuracy (%), which are defined in (12)–(15).
Ns
1 ∑
MSE = (Ti − Oi )2 (12)
Ns i=1

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√ Ns
√1 ∑
RMSE = √ (Ti − Oi )2 (13)
Ns i=1

16
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig 18. (a) training and (b) testing results of the CS-HHT fault locator.

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑Ns ̅
2
i=1 (Ti − Oi )
R= 1− ∑Ns 2
(14)
i=1 (Ti )

NCE
Accuracy% = × 100 (15)
Ns
In the above equations,Ti and Oi are the target and output sampled data, respectively. In addition, NCE and Ns represent the number
of correctly estimated data and the total number of sampled data, respectively (Ns = 648).

3.4. Fault location task

In order to locate the faults, the ANFIS model must be trained and tested using the extracted features. In the proposed model, the
extracted features are applied to the ANFIS, which was previously optimized using HHO and CS algorithms. The fault detection task
has already been completed by comparing the statistic T2 with the pre-defined threshold value, Tα2. Hence, contrarily to the previous
ANFIS-based fault detection models where an independent layer must be trained and tested to distinguish the faulty and normal
conditions, the proposed ANFIS-based HHO-PCA-DWT model includes only two layers for classifying and locating the faults. Fig. 15
illustrates schematically this explanation.
Next, the features acquired from the HHT technique are employed to train and test the optimized ANFIS model. The training and
testing results of the four ANFIS-based models are presented in Figs. 16–19. A comparison of the statistical measures obtained during
training and testing processes corresponding to the fault location task is accomplished in Table 5. This table demonstrates that the
HHO-PCA-DWT strategy performs more accurately in the fault location phase. Furthermore, the MSE and RMSE indices are reduced
significantly by using the HHO-PCA-DWT method. This is due to the profit of the privileges of the HHO optimization algorithm and the
PCA-DWT feature extraction technique. The results confirm the prominent capability of the HHO algorithm in escaping the local
optimum trap when compared with the CS algorithm. Hence, the optimal membership functions designed by exploiting the HHO are
preferable to those of the CS. On the other hand, the features extracted by PCA-DWT provide more distinct characteristics for fault
scenarios in different locations due to the higher resolution. Therefore, training and testing the ANFIS utilizing these characteristics

17
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig 19. (a) training and (b) testing results of the HHO-HHT fault locator.

Table 5
Statistical analysis of the ANFIS-based fault locators.
ANFIS-based Locator MSE RMSE R Accuracy%

Training Testing Training Testing Training Testing Training Testing

CS-HHT 0.04774 0.04939 0.21851 0.22226 0.83312 0.83971 89.63 83.54

HHO-HHT 0.01464 0.01543 0.12102 0.12423 0.92437 0.92394 94.59 88.68
CS-PCA-DWT 0.00913 0.00829 0.09560 0.09107 0.94591 0.94381 98.01 96.39
HHO-PCA-DWT 0.00008 0.00049 0.00898 0.02223 0.99954 0.99677 98.45 97.42

improves the fault location scheme.

In the next step, 40 independent fault cases are simulated to conduct a more detailed investigation of the proposed model, and the
performance of the four ANFIS-based fault classifiers/locators is evaluated. It is worth noting that the fault detection task is fulfilled,
and only the classification and location capabilities of the models are examined. Table 6 presents the obtained results, and the sta­
tistical measures corresponding to the estimation models are compared graphically in Fig. 20. Moreover, the scatter plots corre­
sponding to each model are shown in Fig. 21. In this figure, the blue spheres demonstrate the actual fault class/location, and the red
spheres exhibit the estimation results attained by the ANFIS models. As shown, the proposed HHO-PCA-DWT model can more
accurately classify and locate the faults than the other models. The fault estimation accuracy rates of the proposed HHO-PCA-DWT
model are 100% and 99.995% for classifying and locating tasks, respectively, while the CS-PCA-DWT model also yields acceptable
results (100% and 99.980% accuracy rates for fault classification and location tasks, respectively).
However, the ANFIS models based on the HHT technique exhibit poor accuracies due to limitations of the EMD technique,
particularly the mode mixing effect [84]. This phenomenon causes the decomposed IMFs to fluctuate, which this fluctuation disturbs
the original signal by severely aliasing the instantaneous time/frequency distributions. On the contrary, integrating the conventional
PCA into the wavelet domain not only reduces the dimensionality of the dataset, but also increases the usefulness of the extracted
features by removing the correlated features and revealing the hidden non-stationary data [85,86]. In addition, the HHT-based model
needs an additional ANFIS layer, which has already been trained and tested by using the features extracted from the normal and

18
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Table 6
Evaluation of the ANFIS-based fault classification and location models under various fault scenarios.
Actual Fault Distance (km) Fault Case Fault Classification Results [IC] Fault Location Results [km]

ANFIS-HHO ANFIS-CS

PCA-DWT Features HHT Features PCA-DWT Features HHT Features

86.4075 PPTGR10 2 86.4028 86.4737 86.378 86.1145

109.011 PTNR100 9 109.0062 109.0048 109.101 109.0346
92.7336 PPTGR1 1 92.7399 92.4979 92.7037 92.8115
186.1331 NPTGR1 4 186.1236 186.0507 186.1027 186.8621
176.2251 NPTGR10 5 176.2267 176.1003 176.2482 176.3037
103.2667 PTNR100 9 103.3268 102.9572 103.2227 103.3446
166.5006 NPTGR100 6 166.5201 166.3943 166.4122 166.7448
226.8512 PPTGR10 2 226.8498 227.4734 226.7999 227.4157
188.2158 PPTGR100 3 188.1956 188.2476 188.2372 188.7118
169.3372 PTNR1 7 169.3542 169.3369 169.3637 169.5146
26.2402 PTNR10 8 26.2284 26.3028 26.2073 26.0568
93.5834 PTNR100 9 93.6379 93.3285 93.506 93.2643
141.8171 PPTGR100 3 141.829 141.8205 141.8952 141.7717
110.2512 PTNR10 8 110.2748 110.0977 110.2143 110.4269
12.0816 NPTGR10 5 12.0897 11.9897 12.0641 11.6181
54.885 PTNR1 7 54.8996 54.9702 54.8869 54.7531
200.2054 PPTGR1 1 200.2215 200.4017 200.0979 200.7124
3.7547 PPTGR10 2 3.7547 3.7514 3.8428 3.8867
207.2906 NPTGR100 6 207.2909 208.1624 207.1852 207.3081
18.7366 NPTGR10 5 18.753 18.7677 18.755 18.4619
160.5702 PPTGR10 2 160.5356 160.8157 160.6014 160.8931
120.0507 PTNR1 7 120.068 119.8568 119.9987 119.9837
52.3185 PPTGR100 3 52.3162 52.4785 52.3144 52.2086
137.1878 PTNR100 9 137.1838 137.2626 137.2697 137.461
29.3254 PPTGR10 2 29.321 29.2063 29.3476 28.7458
161.0799 NPTGR1 4 161.0743 160.9093 161.1043 161.1178
143.9005 PTNR10 8 143.893 143.7096 143.9728 143.8408
13.4343 PPTGR100 3 13.433 13.2979 13.4323 13.5098
13.5223 PPTGR1 1 13.5261 13.4338 13.5822 13.4929
36.6002 NPTGR100 6 36.6066 36.7155 36.5867 36.3543
4.7091 NPTGR1 4 4.6951 4.4605 4.8296 4.0723
104.4421 PTNR100 9 104.4229 104.6475 104.4481 104.5173
199.7332 PPTGR10 2 199.7299 199.8005 199.7176 199.9898
148.1736 PTNR1 7 148.1793 148.1434 148.1699 147.9871
124.8311 PPTGR100 3 124.7885 124.9709 124.6826 125.3007
207.3284 NPTGR10 5 207.3475 207.2604 207.25 207.8052
23.4475 PPTGR10 2 23.4403 23.0653 23.4012 22.5737
217.9325 NPTGR1 4 217.9324 217.9092 217.9456 218.3772
25.924 PTNR1 7 25.9164 25.833 25.9793 25.4775
124.0792 PTNR10 8 124.0394 124.4263 124.1868 124.1134
MAE 0.0137 0.1609 0.0489 0.2678
MAPE 0.000254 0.0037 0.0018 0.0091
MSE 0.000391 0.0536 0.0038 0.1192
RMSE 0.0198 0.2315 0.0614 0.3452

abnormal signals to detect the faults. This raises the complexity of the model and also computation time. Conversely, the
PCA-DWT-based model not only profits from the privileges of both PCA and DWT techniques but also has the capability of fault
detection during the PCA execution step without the need for an additional ANFIS layer.

3.5. Comparative study

A comparison with the models presented in the literature is performed in this subsection to prove the supremacy of the proposed
fault detection, classification, and location model. A fault location model based on gap frequency spectrum analysis has been proposed
in [87]. Furthermore, a Fast Fourier transform (FFT) -based model combined with the PRONY algorithm has been proposed in [88] to
estimate the dominant natural frequency, where the dependency of the natural frequency of the transmission line to the fault distance
and traveling wave speed has been confirmed. A significant estimation error has been demonstrated in this reference since the fault
signal has been applied directly as the input to the estimator. Additionally, the dependency of the traveling TW-based models on wave
velocity is a challenging issue; hence, wave speed-independence TW-based strategies have been introduced [89].
On the other hand, the HH transform-based feature extraction method has been employed in [56], where the chaotic dynamic
weight PSO (CDWPSO) algorithm has been used to train the ANFIS-based estimator. Unfortunately, there is no information about the
feature extraction procedure in [56]; hence, only fault location accuracy will be compared in this subsection. A similar study was
conducted in [68] by the same authors, where the DWT method was employed to extract the statistical features directly from the

19
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig 20. Comparison of statistical measures related to the ANFIS-based models.

Fig 21. The scatter plots corresponding to the ANFIS-based fault classification/location models.

20
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Fig. 22. Comparison of the estimation accuracy rate of the models proposed in the literature and this work.

Table 7
Comparison of the models proposed in the literature from the estimation error point of view.
Real Fault Estimated Fault Distance [km] Estimation Error [km]
Distance
Proposed Proposed Proposed Proposed This Work Proposed Proposed Proposed Proposed This
(km)
Method in Method in Method in Method in (HHO- Method in Method in Method in Method in Work
[56] [87] [88] [68] PCA-DWT) [56] [87] [88] [68] (HHO-
PCA-
DWT)

25 24.862 24.97 – – 25.0036 -0.1380 -0.0300 – – 0.0036

25.608 – – – 25.6413 25.6094 – – – 0.0333 0.0014
30 29.919 – 29.67 – 29.9995 -0.0810 – -0.3300 – -0.0005
54.672 – – – 54.7999 54.6792 – – – 0.1279 0.0072
75 74.621 74.46 - – 75.0498 -0.3790 -0.5400 – – 0.0498
78.144 – – – 78.1795 78.1620 – – – 0.0355 0.0180
90 90.345 - 91.17 – 90.0186 0.3450 – 1.1700 – 0.0186
92.040 – – – 92.2849 92.0602 – – – 0.2449 0.0202
115.080 – – – 115.0256 115.0573 – – – -0.0544 -0.0227
115.872 – – – 115.6430 115.8664 – – – -0.2290 -0.0056
117.5 117.193 – 119.05 – 117.5115 -0.3070 – 1.5500 – 0.0115
119.544 – – – 119.7463 119.5430 – – – 0.2023 -0.0010
125 124.745 125.25 – – 125.0099 -0.2550 0.2500 – – 0.0099
126.648 – – – 126.8802 126.6678 – – – 0.2322 0.0198
134.544 – – – 134.3509 134.5465 – – – -0.1931 0.0025
137.928 – – – 138.0545 137.9538 – – – 0.1265 0.0258
140.640 – – – 140.8855 140.6873 – – – 0.2455 0.0473
147.024 – – – 146.8665 147.0564 – – – -0.1575 0.0324
159.936 – – – 160.1158 159.9786 – – – 0.1798 0.0426
167.208 – – – 167.0181 167.2428 – – – -0.1899 0.0348
175 175.195 174.27 – – 174.9495 0.1950 -0.7300 – – -0.0505
177.264 – – – 177.4663 177.2738 – – – 0.2023 0.0098
190 189.725 – 189.18 – 189.9472 -0.2750 – -0.8200 – -0.0528
192.312 – – – 192.3857 192.2770 – – – 0.0737 -0.0350
195.696 – – – 195.6493 195.6691 – – – -0.0467 -0.0269
202.848 – – – 202.8054 202.8143 – – – -0.0426 -0.0337
207.696 – – – 207.6122 207.6845 – – – -0.0838 -0.0115
210 209.814 – 211.45 – 209.9867 -0.1860 – 1.4500 – -0.0133
210.960 – – – 210.8860 210.9792 – – – -0.0740 0.0192
216.216 – – – 216.2493 216.2307 – – – 0.0333 0.0147
223.104 – – – 223.2841 223.2162 – – – 0.1801 0.1122
225 224.798 225.32 – – 225.0115 -0.2020 0.3200 – – 0.0115
232.536 – – – 232.4100 232.5442 – – – -0.1260 0.0082
237.600 – – – 237.7291 237.6028 – – – 0.1291 0.0028
Mean Absolut Error (MAE) 0.2363 0.3740 1.0640 0.1351 0.0229

21
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

Table 8
Comparison of the models proposed in the literature from the MSE value and fault detection/classification capability points of view.
Reference Measured Extracted Features Feature Extraction Optimization MSE Fault Fault
Signal Method Algorithm Detection Classification

[56] 0-mode (I0) Time difference between I0 and I1, Hilbert-Huang (HH) CDWPSO 0.0592 No Yes
and 1-mode characteristic frequency of I1, Transform
(I1) Current energy attenuation coefficients of
Modes I0 and I1, high-frequency energies
of I0 and I1
[68] DC-link Energy, Shannon entropy, log Discrete Wavelet NSGA-II 0.0197 No No
Current energy entropy, norm entropy, Transform (DWT)
root mean square, mean value,
standard deviation, Skewness,
kurtosis
[87] DC-link Gap-based frequency spectrum Fast Fourier – 0.1052 Yes No
Current Transform (FFT)
[88] DC-link Dominant natural frequency Fast Fourier PRONY 0.0996 No No
Current Transform (FFT) Algorithm
This Work DC-link Mean value, standard deviation, Hilbert-Huang (HH) Cuckoo Search 0.1192 No Yes
(CS- Current and norm entropy of the real and Transform (CS)
HHT) imaginary Fourier spectrums
This Work DC-link Mean value, standard deviation, Principal Component Cuckoo Search 0.0038 Yes Yes
(CS- Current and norm entropy of the wavelet Analysis- Discrete (CS)
PCA- coefficients Wavelet Transform
DWT) (PCA-DWT)
This Work DC-link Mean value, standard deviation, Hilbert-Huang (HH) Harris Hawk 0.0536 No Yes
(HHO- Current and norm entropy of the real and Transform Optimization
HHT) imaginary Fourier spectrums (HHO)
This Work DC-link Mean value, standard deviation, Principal Component Harris Hawk 0.00039 Yes Yes
(HHO- Current and norm entropy of the wavelet Analysis-Discrete Optimization
PCA- coefficients Wavelet Transform (HHO)
DWT) (PCA-DWT)

Table 9
Comparison of the ANFIS optimization algorithms.
ANFIS Optimization Algorithm The Converged MSE Value (Best Result) Maximum Iteration Until Convergence (Best Result)

CDWPSO [56] 0.0592 Not mentioned

LSþBP [56] 0.1839 Not mentioned
LSþBP [68] 0.2445 ~365
PSO [68] 0.0546 ~280
GA [68] 0.0378 ~215
CDWPSO [68] 0.0255 ~50
NSGA-II [68] 0.0197 ~30
PRONY [88] 0.0996 Not mentioned
This Work (CS) 0.003191 20
This Work (HHO) 0.002464 9

current signal, and the Non-dominant Sorting Genetic Algorithm Type 2 (NSGA-II) was utilized to optimize the ANFIS. Although the
results obtained in [68] are better than previous works from the MSE and coverage speed viewpoints, evaluating the computational
speed is absent in this reference. The main advantage of the NSGA-II optimization algorithm is that it offers an efficient approach to
address multi-objective optimization problems at the cost of the higher computational burden due to genetic operators [90]; however,
the ANFIS optimization is a single-objective problem. In addition, the NSGA-II algorithm suffers from the dominance-resistant problem
existing in the solutions [91]. Moreover, additional features must be extracted by applying the DWT technique to the current signal to
reach acceptable estimation accuracy. This is due to poor directionality and shift sensitivity of the DWT method [92]. This means that
shifting the input signal leads to unpredictable changes in wavelet coefficients. Varying the fault occurrence instant may generate
undesirable coefficients that fail to distinguish between abnormal and normal conditions to detect the fault. Consequently, more
statistical features, e.g., Skewness and Kurtosis, are required to improve the fault detection capability of the ANFIS-based model.
Conversely, integrating the PCA method into the DWT solves the shift-sensitivity problem by guaranteeing fault detection capa­
bility. Therefore, reduced features would be required to train and test the ANFIS. This property increases the computational speed of
the PCA-DWT model compared to the HHT- and DWT-based estimators and makes it more advantageous for long-distance HVDC
systems. In the following, Fig. 22 and Table 7 present the results of the comparative study with references [56,68,87], and [88]. Since
the data in [87] and [88] are limited to only five distances, the estimation error of the proposed method is also calculated at these five
distances. Fig. 22 shows that the accuracy rate of the proposed ANFIS-HHO-PCA-DWT model is significantly higher than other
methods. This feature is due to the perfect capability of the HHO algorithm to escape from the local optimal traps. In addition, the final
solution obtained by HHO is the global one. Since a Pareto front of possible solutions is created in the first step of the HHO algorithm,

22
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

thus the fitness criterion is checked for all solutions. Meanwhile, a comparative study from the MSE value and fault detection/­
classification capability points of view is performed in Table 8. This table reveals that the MSE is significantly reduced using the
proposed ANFIS-HHO-PCA-DWT model. In addition, a comparison of the convergence speeds and the converged MSE values of the
ANFIS optimization algorithms is performed in Table 9.

4. Conclusions

An adaptive neural-fuzzy inference system (ANFIS) estimator based on a hybrid PCA-DWT signal processing technique is proposed
in this paper to detect, classify, and locate the faults in a VSC-HVDC system. The PCA-DWT technique detects the fault occurrence and
extracts features from the DC-link current signals. The Harris Hawk optimization (HHO)and Cuckoo Search (CS) algorithms are
employed for ANFIS optimization. The PCA technique can detect the fault occurrence by comparing the T2 statistic with a pre-defined
control limit. Next, DWT is applied to extract features from the statistics obtained by PCA, and the features matrix is obtained. In the
simulation stage, 40 fault scenarios are independently simulated in a 250 km VSC-HVDC system. The proposed model can successfully
detect, classify, and locate faults. A comparative study proves the superiority of the proposed ANFIS-HHO-PCA-DWT model compared
to other models presented in previous studies. The numerical value of mean square error (MSE) obtained by the proposed model is
0.002464, which has been significantly reduced compared to other models. In addition, the accuracy rate of 99.995% obtained during
the fault location stage confirms the high accuracy of the proposed locator. Furthermore, the optimization results verify the superiority
of the HHO algorithm in terms of convergence speed. The HHO algorithm exhibits superior performance due to its ability to escape
from local optimality and powerful operators for searching the space of possible solutions. Fast response and accuracy of the fault
locators are vital in VSC-HVDC systems because the system operator can track and fix the fault before the system components are
damaged. Therefore, the proposed model has a meaningful advantage over other models.

Credit author statement

Ehsan Akbari: Conceptualization, Methodology, Data curation, Writing- Original draft preparation, Visualization. Milad Samady
Shadlu: Investigation, Supervision, Methodology, Software, Validation, Writing- Reviewing and Editing.

Funding statement

The authors did not receive support from any organization for the submitted work.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Data availability

No data was used for the research described in the article.

References

[1] Pragati A, Mishra M, Rout PK, Gadanayak DA, Hasan S, Prusty BR. A comprehensive survey of HVDC protection system: fault analysis, methodology, issues,
challenges, and future perspective. Energies (Basel) 2023;16(11):4413. https://doi.org/10.3390/en16114413.
[2] Muniappan M. A comprehensive review of DC fault protection methods in HVDC transmission systems. Prot Control Mod Power Syst 2021;6(1). https://doi.org/
10.1186/s41601-020-00173-9.
[3] Kanwal S, Jiriwibhakorn S. Artificial intelligence based faults identification, classification, and localization techniques in transmission lines-a review. Revista
IEEE América Latina 2023;21(12):1291–305. https://doi.org/10.1109/tla.2023.10305233.
[4] Silva AS, Santos RC, Torres JA, Coury DV. An accurate method for fault location in HVDC systems based on pattern recognition of DC voltage signals. Electr
Power Syst Res 2019;170:64–71. https://doi.org/10.1016/j.epsr.2019.01.013.
[5] Agarwal S, Swetapadma A, Panigrahi C, Dasgupta A. A method for fault section identification in High voltage direct current transmission lines using one End
measurements. Electr Power Syst Res 2019;172:140–51. https://doi.org/10.1016/j.epsr.2019.03.008.
[6] Zhang Y, Cong W. An improved single-ended frequency-domain-based fault detection scheme for MMC-HVDC transmission lines. Int J Electr Power Energy Syst
2021;125:106463. https://doi.org/10.1016/j.ijepes.2020.106463.
[7] Wang M, Beerten J, Van Hertem D. Frequency domain based DC fault analysis for bipolar HVDC grids. J Mod Power Syst Clean Energy 2017;5(4):548–59.
https://doi.org/10.1007/s40565-017-0307-y.
[8] Imani A, Moravej Z, Pazoki M. A novel time-domain method for fault detection and classification in VSC-HVDC transmission lines. Int J Electr Power Energy Syst
2022;140:108056. https://doi.org/10.1016/j.ijepes.2022.108056.
[9] Gao Z, Cecati C, Ding S. A survey of fault diagnosis and fault-tolerant techniques Part II: fault diagnosis with knowledge-based and hybrid/active approaches.
IEEE Trans Ind Electron 2015;1. https://doi.org/10.1109/tie.2015.2419013.
[10] Mardani R, Ehteshami MZ. Efficient model-based DC fault detection and location scheme for multi-terminal HVDC systems with voltage source converter. Electr
Eng 2021;104(3):1553–64. https://doi.org/10.1007/s00202-021-01412-4.
[11] Xu J, Lu Y, Zhao C, Liang J. A model-based DC fault location scheme for multi-terminal MMC-HVDC systems using a simplified transmission line representation.
IEEE Trans Power Deliv 2020;35(1):386–95. https://doi.org/10.1109/tpwrd.2019.2932989.

23
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

[12] Chang B, Cwikowski O, Barnes M, Shuttleworth R, Beddard A, Coventry P. Review of different fault detection methods and their impact on pre-emptive VSC-
HVDC dc protection performance. High Voltage 2017;2(4):211–9. https://doi.org/10.1049/hve.2017.0024.
[13] Wilches-Bernal F, Bidram A, Reno MJ, Hernandez-Alvidrez J, Barba P, Reimer B, Montoya R, Carr C, Lavrova O. A survey of traveling wave protection schemes
in electric power systems. IEEE Access 2021;9:72949–69. https://doi.org/10.1109/access.2021.3080234.
[14] Chen W, Wang D, Cheng D, Liu X. Travelling wave fault location approach for MMC-HVDC transmission line based on frequency modification algorithm. Int J
Electr Power Energy Syst 2022;143:108507. https://doi.org/10.1016/j.ijepes.2022.108507.
[15] Tiwari RS, Gupta OH, Sood VK. Fault detection using backward propagating traveling waves for bipolar LCC-HVDC lines. Electr Power Compon Syst 2022;50
(9–10):516–29. https://doi.org/10.1080/15325008.2022.2136788.
[16] Muzzammel R. Traveling waves-based method for fault estimation in HVDC transmission system. Energies (Basel) 2019;12(19):3614. https://doi.org/10.3390/
en12193614.
[17] Zhang S, Zou G, Huang Q, Gao H. A traveling-wave-based fault location scheme for MMC-based multi-terminal DC grids. Energies (Basel) 2018;11(2):401.
https://doi.org/10.3390/en11020401.
[18] Liu T, Zhang Y, Wang S, Li X, Gooi HB, Ghias AM. Fault identification and fault location methods for VSC-HVDC transmission lines based on the traveling
waveform difference. Int J Electr Power Energy Syst 2023;147:108867. https://doi.org/10.1016/j.ijepes.2022.108867.
[19] Muzzammel R, Raza A, Hussain MR, Abbas G, Ahmed I, Qayyum M, Rasool MA, Khaleel MA. MT–HVDC systems fault classification and location methods based
on traveling and non-traveling waves—A comprehensive review. Appl Sci 2019;9(22):4760. https://doi.org/10.3390/app9224760.
[20] Imani A, Moravej Z, Pazoki M. A novel MODWT-based fault detection and classification scheme in VSC-HVDC transmission line. Electr Power Syst Res 2023;221:
109434. https://doi.org/10.1016/j.epsr.2023.109434.
[21] Lacerda VA, Monaro RM, Campos-Gaona D, Coury DV, Anaya-Lara O. Distance protection algorithm for multiterminal HVDC systems using the Hilbert–Huang
transform. IET Gener Transm Distrib 2020;14(15):3022–32. https://doi.org/10.1049/iet-gtd.2019.1551.
[22] Zhu X, Li S, Guo Y, Chen X, He C, Deng J. Novel wavefront detection and fault location method based on Hilbert-Huang transform for long HVDC transmission
lines. Electr Power Syst Res 2022;211:108213. https://doi.org/10.1016/j.epsr.2022.108213.
[23] Li D, Ukil A, Satpathi K, Yeap YM. Hilbert–Huang transform based transient analysis in voltage source converter interfaced direct current system. IEEE Trans Ind
Electron 2021;68(11):11014–25. https://doi.org/10.1109/tie.2020.3038056.
[24] Radwan M, Azad SP. A single-ended protection scheme for multi-terminal HVDC grids based on Hilbert-Huang transform. In: IEEE Power & energy society
general meeting; 2023. https://doi.org/10.1109/pesgm52003.2023.10252557.
[25] Mishra S, Gupta S, Yadav A. Intrinsic time decomposition based fault location scheme for unified power flow controller compensated transmission line. Int Trans
Electr Energy Syst 2020;30(11). https://doi.org/10.1002/2050-7038.12585.
[26] Wang X, Zhang D, Zhang X, Ni P. Fault diagnosis method for hybrid HVDC transmission system based on stacked sparse auto-encoder. In: 7th Asia conference on
power and electrical engineering (ACPEE); 2022. https://doi.org/10.1109/acpee53904.2022.9783919.
[27] Peng G, Chen L, Wu J, Jiang H, Wang Z, Li H. Convolution power ratio based on Single-Ended Protection Scheme for HVDC transmission lines. Electronics
(Basel) 2023;12(23):4883. https://doi.org/10.3390/electronics12234883.
[28] Li D, Cai Z, Qin B, Deng L. Signal frequency domain analysis and sensor fault diagnosis based on artificial intelligence. Comput Commun 2020;160:71–80.
https://doi.org/10.1016/j.comcom.2020.05.034.
[29] Agarwal S, Swetapadma A, Panigrahi C, Dasgupta A. Fault detection in direct current transmission lines using discrete fourier transform from single terminal
current signals. In: 1st International conference on electronics, materials engineering and nano-technology (IEMENTech); 2017. https://doi.org/10.1109/
iementech.2017.8076975.
[30] Lacerda VA, Monaro RM, Coury DV. Fault distance estimation in multiterminal HVDC systems using the Lomb-Scargle periodogram. Electr Power Syst Res 2021;
196:107251. https://doi.org/10.1016/j.epsr.2021.107251.
[31] Yeap YM, Ukil A. Fault detection in HVDC system using short time Fourier transform. In: IEEE Power and energy society general meeting; 2016. https://doi.org/
10.1109/pesgm.2016.7741323.
[32] Singh K, Ukil A. Fault detection in HVDC transmission line by S-transform technique. In: IEEE PES Asia-pacific power and energy engineering conference; 2019.
https://doi.org/10.1109/appeec45492.2019.8994571.
[33] Li D, Ukil A, Satpathi K, Yeap YM. Improved S Transform-Based fault detection method in voltage source converter interfaced DC system. IEEE Trans Ind
Electron 2021;68(6):5024–35. https://doi.org/10.1109/tie.2020.2988193.
[34] Prakash R, Koley E. A combined gabor transform-EKNN based protection scheme for AC-HVDC transmission line with DFIG wind turbine. In: 13th IEEE PES Asia
pacific power & energy engineering conference; 2021. https://doi.org/10.1109/appeec50844.2021.9687784.
[35] Moravej Z, Movahhedneya M, Pazoki M. Gabor transform-based fault location method for multi-terminal transmission lines. Measurement 2018;125:667–79.
https://doi.org/10.1016/j.measurement.2018.05.027.
[36] Yeap YM, Geddada N, Ukil A. Analysis and validation of wavelet transform based DC fault detection in HVDC system. Appl Soft Comput 2017;61:17–29. https://
doi.org/10.1016/j.asoc.2017.07.039.
[37] Huai Q, Liu K, Ding H, Qin L. Protection scheme for multiterminal HVDC system based on Wavelet transform modulus maxima. IEEJ Trans Electr Electron Eng
2020;15(8):1147–59. https://doi.org/10.1002/tee.23175.
[38] Psaras V, Vozikis D, Adam G, Burt G. Real time evaluation of Wavelet transform for fast and efficient HVDC grid non-unit protection. In: IEEE Power & energy
society general meeting; 2020. https://doi.org/10.1109/pesgm41954.2020.9281507.
[39] Ahmed N, Hashmani AA, Khokhar S, Tunio MA, Faheem M. Fault detection through discrete wavelet transform in overhead power transmission lines. Energy Sci
Eng 2023;11(11):4181–97. https://doi.org/10.1002/ese3.1573.
[40] Da Silva DM, Costa FB, Miranda V, Leite H. Wavelet-based analysis and detection of traveling waves due to DC faults in LCC HVDC systems. Int J Electr Power
Energy Syst 2019;104:291–300. https://doi.org/10.1016/j.ijepes.2018.07.011.
[41] Joshi A, Khathoon R, Devikrishna N, Peter PA, Vinod V. A computationally less expensive fault detection technique in VSC-HVDC system using wavelet
decomposition and support vector machine classifier. In: IEEE International conference on power electronics, smart grid, and renewable energy (PESGRE); 2022.
https://doi.org/10.1109/pesgre52268.2022.9715818.
[42] Jawad RS, Abid H. HVDC fault detection and classification with artificial neural network based on ACO-DWT method. Energies (Basel) 2023;16(3):1064.
https://doi.org/10.3390/en16031064.
[43] Fayazi M, Joorabian M, Saffarian A, Monadi M. A single-ended traveling wave based fault location method using DWT in hybrid parallel HVAC/HVDC overhead
transmission lines on the same tower. Electr Power Syst Res 2023;220:109302. https://doi.org/10.1016/j.epsr.2023.109302.
[44] Wang Z, Xu L. Fault detection of the power system based on the chaotic neural network and Wavelet transform. Complexity 2020;2020:1–15. https://doi.org/
10.1155/2020/8884786.
[45] Ahmed N, Ram N, Memon AP, Ahmed S. Comparative analysis of fault detection for HVDC transmission system using Wavelet transform based on standard
deviation. In: 3rd International conference on computing, mathematics and engineering technologies (iCoMET); 2020. https://doi.org/10.1109/
icomet48670.2020.9074065.
[46] Jawad RS, Abid H. Fault detection in HVDC system with gray wolf optimization algorithm based on artificial neural network. Energies (Basel) 2022;15(20):
7775. https://doi.org/10.3390/en15207775.
[47] Ding C, Wang Z, Ding Q, Yuan Z. Convolutional neural network based on fast Fourier transform and gramian angle field for fault identification of HVDC
transmission line. Sustain Energy Grids Netw 2022;32:100888. https://doi.org/10.1016/j.segan.2022.100888.
[48] Merlin VL, Santos RCD, Pavani APG, Vieira JCM. A frequency spectrum-based method for detecting and classifying faults in HVDC systems. Electr Power Syst
Res 2022;207:107828. https://doi.org/10.1016/j.epsr.2022.107828.
[49] Goni MF, Nahiduzzaman M, Anower M, Rahman M, Islam M, Ahsan M, Haider J, Shahjalal M. Fast and accurate fault detection and classification in transmission
lines using extreme learning machine. e-Prime Adv Electr Eng Electron Energy 2023;3:100107. https://doi.org/10.1016/j.prime.2023.100107.

24
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

[50] Johnson JM, Yadav A. Complete protection scheme for fault detection, classification and location estimation in HVDC transmission lines using support vector
machines. IET Sci Meas Technol 2017;11(3):279–87. https://doi.org/10.1049/iet-smt.2016.0244.
[51] Ghashghaei S, Akhbari M. Fault detection and classification of an HVDC transmission line using a heterogenous multi-machine learning algorithm. IET Gener
Transm Distrib 2021;15(16):2319–32. https://doi.org/10.1049/gtd2.12180.
[52] Chen Q, Wu J, Li Q, Gao X, Yu R, Guo J, Peng G, Yang B. Long short-term memory network-based HVDC systems fault diagnosis under knowledge graph.
Electronics (Basel) 2023;12(10):2242. https://doi.org/10.3390/electronics12102242.
[53] Chen Q, Li Q, Wu J, He J, Mao C, Li Z, Yang B. State monitoring and fault diagnosis of HVDC system via KNN algorithm with knowledge graph: a practical china
power grid case. Sustainability 2023;15(4):3717. https://doi.org/10.3390/su15043717.
[54] Mustari MR, Hashim MN, Osman MK, Ahmad AR, Ahmad F, Ibrahim MN. Fault location estimation on transmission lines using neuro-fuzzy system. Procedia
Comput Sci 2019;163:591–602. https://doi.org/10.1016/j.procs.2019.12.141.
[55] Kumar AN, Sanjay C, Chakravarthy M. Adaptive neuro fuzzy inference system-based fault location technique in double circuit transmission line against
simultaneous faults. Multiscale Multidiscip Model Exp Des 2019;3(2):143–50. https://doi.org/10.1007/s41939-019-00066-x.
[56] Rohani R, Koochaki A. A hybrid method based on optimized neuro-fuzzy system and effective features for fault location in VSC-HVDC systems. IEEE Access
2020;8:70861–9. https://doi.org/10.1109/access.2020.2986919.
[57] Abbas M, Zhang D. A smart fault detection approach for PV modules using adaptive neuro-fuzzy inference framework. Energy Reports 2021;7:2962–75. https://
doi.org/10.1016/j.egyr.2021.04.059.
[58] Bendary AF, Abdelaziz AY, Ismail MM, Mahmoud K, Lehtonen M, Darwish MMF. Proposed ANFIS based approach for fault tracking, detection, clearing and
rearrangement for photovoltaic system. Sensors 2021;21(7):2269. https://doi.org/10.3390/s21072269.
[59] Ghaghishpour A, Koochaki A. An intelligent method for online voltage stability margin assessment using optimized ANFIS and associated rules technique. ISA
Trans 2020;102:91–104. https://doi.org/10.1016/j.isatra.2020.02.028.
[60] Akbari E, Shadlu MS. Fault detection, classification, and location in VSC-HVDC systems using metaheuristic optimized ANFIS and Hilbert Huang transform
method. In: International conference on protection and automation in power system; 2024. https://doi.org/10.1109/ieeeconf62358.2024.10453213.
[61] Salleh MNM, Talpur N, Hussain K. Adaptive neuro-fuzzy inference system: overview, strengths, limitations, and solutions. Lecture notes in computer science.
2017. p. 527–35. https://doi.org/10.1007/978-3-319-61845-6_52.
[62] Liu X, Hussein SH, Ghazali KH, Tung TM, Yaseen ZM. Optimized adaptive neuro-fuzzy inference system using metaheuristic algorithms: application of shield
tunnelling ground surface settlement prediction. Complexity 2021;2021:1–15. https://doi.org/10.1155/2021/6666699.
[63] AlRassas AM, Al-Qaness MaA, Ewees AA, Ren S, Elaziz MA, Damaševičius R, Krilavičius T. Optimized ANFIS model using aquila optimizer for oil production
forecasting. Processes 2021;9(7):1194. https://doi.org/10.3390/pr9071194.
[64] Dirik M. Optimized Anfis model with hybrid metaheuristic algorithms for facial emotion recognition. Int J Fuzzy Syst 2022;25(2):485–96. https://doi.org/
10.1007/s40815-022-01402-z.
[65] El-Hasnony IM, Barakat SI, Mostafa RR. Optimized ANFIS model using hybrid metaheuristic algorithms for Parkinson’s disease prediction in IoT environment.
IEEE Access 2020;8:119252–70. https://doi.org/10.1109/access.2020.3005614.
[66] Habibi E, Salehi M, Yadegarfar G, Taheri A. Optimization of the ANFIS using a genetic algorithm for physical work rate classification. Int J Occup Saf Ergon
2018;26(3):436–43. https://doi.org/10.1080/10803548.2018.1435445.
[67] Anshori MY, Rahmalia D, Herlambang T, Karya DF. Optimizing adaptive neuro fuzzy inference system (ANFIS) parameters using Cuckoo search (Case study of
world crude oil price estimation). J Phys Conf Ser 2021;1836(1):012041. https://doi.org/10.1088/1742-6596/1836/1/012041.
[68] Rohani R, Koochaki A, Siahbalaee J. Fault location in VSC-HVDC systems based on NSGA-II and discrete wavelet transform. Int J Renew Energy Res 2022;12(3).
https://doi.org/10.20508/ijrer.v12i3.13050.g8519.
[69] Swetapadma A, Agarwal S, Ranjan A, Abdelaziz AY. A novel fault distance estimation method for voltage source converter-based HVDC transmission lines.
Electr Power Compon Syst 2021:1–17. https://doi.org/10.1080/15325008.2021.1908447.
[70] Shadlu MS. Open-Circuit fault detection and location in modular multilevel converters based on principal component analysis. In: 13th Power electronics, drive
systems, and technologies conference (PEDSTC); 2022. https://doi.org/10.1109/pedstc53976.2022.9767383.
[71] Karaboga D, Kaya E. Adaptive network based fuzzy inference system (ANFIS) training approaches: a comprehensive survey. Artif Intell Rev 2018;52(4):
2263–93. https://doi.org/10.1007/s10462-017-9610-2.
[72] Abbasi Z, Shafieirad M, Mehra AHA, Zamani I. Vaccination and isolation based control design of the COVID-19 pandemic based on adaptive neuro fuzzy
inference system optimized with the genetic algorithm. Evol Syst 2022;14(3):413–35. https://doi.org/10.1007/s12530-022-09459-9.
[73] Pop M, Pescaru D, Micea MV. Mamdani vs. Takagi–Sugeno fuzzy inference systems in the calibration of continuous-time car-following models. Sensors 2023;23
(21):8791. https://doi.org/10.3390/s23218791.
[74] Alabool HM, Alarabiat D, Abualigah L, Heidari AA. Harris hawks optimization: a comprehensive review of recent variants and applications. Neural Comput Appl
2021;33(15):8939–80. https://doi.org/10.1007/s00521-021-05720-5.
[75] Yang X, Deb NS. Cuckoo search via levy flights. In: World congress on nature & biologically inspired computing; 2009. https://doi.org/10.1109/
nabic.2009.5393690.
[76] Nguyen T, Vo D, Quynh NV, Van Dai L. Modified Cuckoo search algorithm: a novel method to minimize the fuel cost. Energies (Basel) 2018;11(6):1328. https://
doi.org/10.3390/en11061328.
[77] Shaikh SM, A I, H N, S M, Hussain A, A A. Data-driven based fault diagnosis using principal component analysis. Int J Adv Comput Sci Appl 2018;9(7). https://
doi.org/10.14569/ijacsa.2018.090725.
[78] Liu C, Bai J, Wu F. Fault diagnosis using dynamic principal component analysis and GA feature selection modeling for industrial processes. Processes 2022;10
(12):2570. https://doi.org/10.3390/pr10122570.
[79] Huang N, Lu G, Cai G, Xu D, Xu J, Li F, Zhang L. Feature selection of power quality disturbance signals with an entropy-importance-based random forest.
Entropy 2016;18(2):44. https://doi.org/10.3390/e18020044.
[80] Jin X, Lu S, Yu J. Adaptive NN-Based consensus for a class of nonlinear multiagent systems with actuator faults and faulty networks. IEEE Trans Neural Netw
Learn Syst 2022;33(8):3474–86. https://doi.org/10.1109/tnnls.2021.3053112.
[81] Ren Y, Liu F, Lv J, Meng A, Wen Y. T-S fuzzy systems optimization identification based on FCM and PSO. EURASIP J Adv Signal Process 2020;2020(1). https://
doi.org/10.1186/s13634-020-00706-2.
[82] Bottmer L, Croux C, Wilms I. Sparse regression for large data sets with outliers. Eur J Oper Res 2022;297(2):782–94. https://doi.org/10.1016/j.
ejor.2021.05.049.
[83] Kaur M, Kaur R, Singh N. Fusion of powerful features of Harris Hawks with Cuckoo Search algorithm for engineering applications. Results Control Optim 2023;
13:100308. https://doi.org/10.1016/j.rico.2023.100308.
[84] Wang Y, Hu K, Lo M. Uniform phase Empirical mode decomposition: an optimal hybridization of masking signal and ensemble approaches. IEEE Access 2018;6:
34819–33. https://doi.org/10.1109/access.2018.2847634.
[85] Lim Y, Kwon J, Oh H. Principal component analysis in the wavelet domain. Pattern Recognit 2021;119:108096. https://doi.org/10.1016/j.patcog.2021.108096.
[86] Nobre J, Neves RF. Combining principal component analysis, discrete wavelet transform and XGBoost to trade in the financial markets. Expert Syst Appl 2019;
125:181–94. https://doi.org/10.1016/j.eswa.2019.01.083.
[87] Yang Q, Blond SL, Cornelusse B, Vanderbemden P, Li J. A novel fault detection and fault location method for VSC-HVDC links based on gap frequency spectrum
analysis. Energy Procedia 2017;142:2243–9. https://doi.org/10.1016/j.egypro.2017.12.625.
[88] Song G, Chu X, Cai X, Gao S, Ran M. A fault-location method for VSC-HVDC transmission lines based on natural frequency of current. Int J Electr Power Energy
Syst 2014;63:347–52. https://doi.org/10.1016/j.ijepes.2014.05.069.
[89] Gao S, Xu Z, Song G, Shao M, Jiang Y. Fault location of hybrid three-terminal HVDC transmission line based on improved LMD. Electr Power Syst Res 2021;201:
107550. https://doi.org/10.1016/j.epsr.2021.107550.

25
E. Akbari and M. Samady Shadlu Computers and Electrical Engineering 120 (2024) 109763

[90] Verma S, Pant M, Snasel V. A comprehensive review on NSGA-II for multi-objective combinatorial optimization problems. IEEE Access 2021;9:57757–91.
https://doi.org/10.1109/access.2021.3070634.
[91] Pang LM, Ishibuchi H, Shang K. NSGA-II with simple modification works well on a wide variety of Many-Objective problems. IEEE Access 2020;8:190240–50.
https://doi.org/10.1109/access.2020.3032240.
[92] Fernandes F, Van Spaendonck R, Burrus C. A new framework for complex wavelet transforms. IEEE Trans Signal Process 2003;51(7):1825–37. https://doi.org/
10.1109/tsp.2003.812841.

26