Received 10 March 2024, accepted 31 March 2024, date of publication 3 April 2024, date of current version 10 April 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3384761

Advanced Fault Detection, Classification, and

Localization in Transmission Lines: A
Comparative Study of ANFIS, Neural
Networks, and Hybrid Methods
SHAZIA KANWAL , (Graduate Student Member, IEEE), AND
SOMCHAT JIRIWIBHAKORN , (Member, IEEE)
School of Engineering, King Mongkut’s Institute of Technology Ladkrabang (KMITL), Bangkok 10520, Thailand
Corresponding author: Somchat Jiriwibhakorn ([email protected])
This work was supported by the King Mongkut’s Institute of Technology Ladkrabang (KMITL), Bangkok, Thailand, under Grant
KDS2021/021.

ABSTRACT Electric systems are getting more complex with time, and primitive protection methods such
as traveling wave and impedance-based methods face limitations and shortcomings. This paper incorporates
and presents the applications of an adaptive neuro-fuzzy inference system and compares it with a back
propagation neural network, self-organizing map, and hybrid method of discrete wavelet with adaptive
neuro-fuzzy inference system for fault detections, classification, and localization in transmission lines. These
methods, in comparison with primitive methods, could be capable of detecting, identifying, and predicting
the location of the faults more accurately. The IEEE 9-bus system is utilized to obtain data from one end of
the transmission line to develop an ANFIS-based model. This system is simulated in MATLAB/Simulink
for different fault cases at various locations. The three-phase voltage and current at one end of IEEE 9-bus
number seven are taken for training. Three ANFIS models are developed for fault detection, classification,
and localization and compared with other models. For verification of the models, mean square error, mean
absolute error, and regression analysis have been computed and compared for all the models. All four
techniques have performed well for fault classification, detection, and location. However, the percentage
error for the ANFIS-based fault model is less compared to backpropagation, self-organizing map, and
discrete wavelet transform with ANFIS. Therefore, the proposed ANFIS models can be implemented for
deploying in real-time-based protection systems.

INDEX TERMS ANFIS, SOM, DWT fault detection, classification and location, IEEE 9-bus system,
transmission line.

I. INTRODUCTION in transmission lines is an undesirable event that may happen

The three stages of the power system are generation, trans- due to lightning, a short circuit between transmission lines,
mission, and distribution. The electrical system is built to an accident, any other unforeseen incident, or even human
provide stable and continuous electricity [2]. Some parts of error [1].
electrical systems are exposed to the environment, so these The short circuit faults in transmission lines are symmetri-
systems are more susceptible to faults. There are two main cal and asymmetrical. Asymmetrical faults are single-line-to-
types of faults in transmission lines: short circuit faults or ground, double-line, and double-line-to-ground faults. These
shunt faults and series faults or open conductor faults. A fault faults are commonly known as phase A to ground (A-G),
phase B to ground (B-G), phase C to ground (C-G), phase A
The associate editor coordinating the review of this manuscript and to phase B (A-B), phase B to phase C (B-C), phase C to
approving it for publication was Ruisheng Diao . phase A (C-A), phase AB to ground (AB-G), phase BC to
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 49017
 S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

ground (BC-G), phase CA to ground (CA-G) faults. These a signal processing tool, was established in the
faults cause instability in the power supply and severe damage nineties [4], [8].
to the system and humans [2]. Locating faults quickly and The discrete variant of the wavelet transform (WT), known
accurately is vital to reduce the outage time and save the as the discrete wavelet transform (DWT), is widely used to
system from major damage. Intelligent fault detection, iden- analyze discrete or sampled signals. The demand for DWT
tification, and location will increase the electrical system’s techniques in digital relaying systems has been rising recently
reliability, restore the power system, and reduce the power in an era of digital communication and analysis [10].
system’s outage time [1], [3]. For the safety of transmission lines, different AI-based
Protecting a transmission line uses relays at both ends fault detection, classification, and Location techniques, such
that continuously monitor voltages and currents and respond as ANN, Fuzzy logic, and hybrid algorithms using integrated
to a fault. The most frequently utilized transmission line wavelet transform ANN, SVM, and fuzzy logic, have been
protection method is distance protection, with a line safety used for the past few years [9], [11], [14].
of around 85% [4]. Artificial neural networks, or ANNs, have traditionally
There are a few limitations in using phasor-based methods been used successfully in various fault analysis studies.
in the protection system. First, delay associated with relay ANN’s ability to learn independently is its most valuable
operation results in uncertainty in voltage protection. Second, feature.
to ensure quick tripping from both ends, it is necessary to The author in [11] suggested an ANN-based method in
employ a communication link between the relays at the two which a feed-forward neural network is used to detect and
ends of the line [5]. Relay communication links provide a classify transmission line faults.
cyber threat when relays are connected to substation local ANN is frequently combined with fuzzy logic inference to
area networks (LANs). create an Adaptive Neuro-Fuzzy Inference System (ANFIS)
Researchers have presented several methods for transmis- topology, which many researchers have utilized to provide
sion line fault detection and location methods; the broader adequate security measures [6], [12], [13].
categories are mentioned here: Conventional methods, Signal ANFIS is similar to a neural network, and the function is
processing-based methods, and Artificial intelligence-based the same as a fuzzy inference system. ANFIS is used for the
methods. location and classification of faults in a transmission line.
The conventional methods that are often used for the An adaptive network is a multilayer network where every
detection, classification, and localization of the fault in a node operates a particular function of the applied data set.
transmission line are the traveling wave method and the The process of the node varies from node to node.
impedance measurement-based method [6]. Single-end or The author in [1] presented a neuro-fuzzy technique to
two-end impedance techniques are suggested based on a investigate fault location estimates in power systems. The
variation of current and voltage signals gathered from a trans- development of a 264 km, 132 kV, 50 Hz transmission line
mission line terminal. Due to the high fault impedance, load model and the simulation of various faults using MATLAB
on the line, source characteristics, and shunt capacitance, Simulink. To train the ANFIS, wavelet-processed data from
the fault location error of the impedance-based method is both ends of the line are used, including the detailed features
high [7]. of the signal.
To calculate the fault’s distance, the traveling wave-based In contrast to other machine learning methods, the effec-
method compares the forward and backward propagation tiveness of the Adaptive Neuro-Fuzzy Inference System
of waves in a transmission line. This approach is more (ANFIS) as an estimation model is discussed in [2]. In [13],
accurate in finding faults in high resistance faults. How- to localize the fault site precisely, the method feeds the
ever, the challenges are the high sampling frequency, high data into ANFIS and computes the line impedance using the
cost, and computational complexity that make practical use approximation coefficients of these signals.
difficult [8]. The use of an ANFIS and ANN for the classifica-
Signal processing methods such as fast Fourier transform tion and localization of faults in a lengthy transmis-
(FFT), wavelet transform (WT), multiresolution analysis sion line is presented in [8]. Artificial intelligence-based
(MRA), and discrete wavelet transform are some of the machine learning techniques do better in specific tasks
famous methods. WT is entirely accurate and can identify the than other approaches. Using current and voltage data
fault characteristics using a faulty waveform’s decomposed from the source end, ANFIS and ANN are used to pre-
frequency components [9]. For the calculation of coefficients cisely identify fault types and pinpoint the transmission line
to be proper and precise, the FFT needs the signal to be issue. Global System For Mobile Communication (GSM)
stationary in the broad sense. However, most signals in power and Global Positioning System (GPS) based ANN model
systems are flexible and fluctuate over time concerning their has been developed in [15]. The challenge with such a
properties [4]. communication-based protection system is that it could be
To address several issues with Fourier transform more reliable, and there are always issues with cybersecurity
analysis techniques, multiresolution analysis (MRA), threats.

49018 VOLUME 12, 2024

S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

Considering the mentioned restrictions, a new relaying

technique is required. This technique helps the system operate
without delay, just with currents and voltages, and without a
communication link.
Therefore, this paper presents an Artificial Intelligence
(AI) based technique that uses only voltage and current for
fault detection, classification, and Location and is tested using
only post-fault three-phase current and voltage.
The detection, classification, and localization of fault in a
transmission line in an IEEE 9-bus system using ANFIS are
studied in this paper. The post-fault three-phase voltage and
current data from one end bus of an IEEE 9-bus system are FIGURE 1. IEEE 9-bus system [16].
used to detect and determine the fault’s class and Location.
All Asymmetrical marks are considered while varying the
Location of the faults. where each node performs a specific function according to the
The main contributions in this paper are as follows: applied data set. The process of each node is different from
1) The ANFIS-based technique is proposed for fault detec- that of another node. This allows ANFIS to differentiate each
tion, classification, and Location, using one-end data for the input feature and detect the output.
IEEE 9- bus system. Some advantages, such as better generalization ability,
2) New technique called Self Organizing Map (SOM) for learning from experience, and better decision-making ability,
fault detection, classification, and location prediction is used make ANFIS better for complex problems. Therefore, ANFIS
with IEEE 9-bus. is helpful for fault detection, location, and classification in
3) Various asymmetrical faults under various conditions the transmission line. This approach has higher accuracy
and for different Locations have been studied. compared to other networks.
4) For the precision of the proposed ANFIS-based model, As mentioned above, a multilayer feedforward network
root mean square error, regression analysis, and mean abso- called an adaptive network, shown in Fig.3, allows each node
lute error are computed in MATLAB. to perform a unique function on incoming signals and a
5) A comparison of the ANFIS-based model with a set of parameters specific to that node. The node functions’
backpropagation neural network, self-organizing map, and formulae may differ from one node to the next, and the selec-
discrete wavelet transform-based ANFIS is presented. tion of each node function depends on the overall input-output
To present the mentioned goals, the paper is organized function that the adaptive network must perform. It should
as follows. A brief overview of the system, the techniques be noted that connections in an adaptive network do not
ANFIS, ANN SOM, and DWT-ANFIS, and the detailed carry weights; instead, they simply represent the signal flow
methodology are given in Section II. Simulation results for direction between nodes [19].
the proposed model for different fault conditions are given in As illustrated in Fig.3, the ANFIS construction includes
Section III. And finally concluded in Section IV. a Fuzzy Interface System (FIS) whose membership function
of parameters was altered using various methods, such as
II. SYSTEM AND TECHNIQUES the backpropagation algorithm or the least squared approach.
A. SYSTEM STUDIED Compared to fuzzy inference systems, which are not offered
by all fuzzy system possibilities, ANFIS is significantly more
The IEEE 9-bus system is used as a study model in this
sophisticated. ANFIS exclusively supports the Sugeno-type
paper [16]. The single line diagram is given in Fig.1. All the
data for the IEEE 9-bus system is taken from [17] and [18]
and is displayed in Fig. 2. The transmission line under study
is between bus number 7 and bus number 8. The IEEE 9-bus
system consists of three generators, three transformers, nine
buses, six transmission lines, and three loads. The intercon-
nection of all these utilities is shown in Fig. 2.
This system is simulated in MATLAB/Simulink version
2019a to implement the proposed ANFIS-based fault detec-
tion and classification technique.

B. PROPOSED TECHNIQUE
1) ADAPTIVE NEURO-FUZZY INFERENCE SYSTEM
ANFIS is a hybrid learning algorithm that best uses ANN
and fuzzy logic. An adaptive network is a multilayer network FIGURE 2. IEEE 9-bus system impedance diagram [17] [18].

VOLUME 12, 2024 49019

 S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

Layer 5:
In this layer, all the inputs from the previous layer are
summed together to get the predicted output. The total sum
output is given by below equation:
X
O51 = yi
Xi

= wi fi = w1 p1 x1 ) + w1 (q1 x2 ) + (w1 r 1 + w2
i

p2 x2 ) + w2 (q2 x2 ) + (w2 r 2 (5)
The succeeding parameters can be solved using a least
FIGURE 3. ANFIS architecture with input features [19].
squares technique in this final layer. The final equation can
be written as:
and has a number of features that are compatible with the 
p1

ANFIS system.  q1 
The mathematical structure of an Adaptive Neuro-fuzzy 
 r1 

inference system with one prediction (output) y and two y = [w1 x1 w1 x2 w1 w2 x2 w2 ] 
 p2  = XW
 (6)
inputs x1 (current) and x2 (voltage) is defined below [6]. 
q 

2
The rule base has two Takagi and Sugeno-type fuzzy if- r2
then rules.
Rule I: If x1 is A1 and x2 is B1 , then f1 = pl x1 + q1 x2 + rl . The flow diagram for ANFIS techniques is shown in
Rule 2: If x1 is A2 and x2 is B2 , then f2 = p2 x1 + Fig.4. The steps followed for the ANFIS model are as
q2 x2 + r2 . follows [28], [29]:
The five layers of the ANFIS structure are as follows: Step 1. Select the area in the power system for study; in
Layer 1: this study, the area between bus seven and bus eight has been
Each node i in this layer has a node function. An adaptive selected as shown in Fig. 1.
node is present in this labeling. The fuzzy membership grade Step 2. Generate fault between bus seven and bus 8,
of the inputs is the layer’s executed output and is represented and capture the voltage and current values from one
as follows: end.
Step 3. Apply all fault conditions, AG, BG, CG, AB, BC,
O1i = µAi (x) (1) CA, ABG, BCG, CAG and no-fault
where A is the linguistic label connected to this node. isO1i Step 4. Change the fault position and collect the faulty
the membership function of µAi (x). Every MF (membership signal’s voltage and current (rms)values.
function) is changed according to the layer’s parameter. Step 5. Normalization of the data
Layer 2: Step 6. Select the structure of ANFIS
The nodes are fixed, nodes with the symbol π, which Step 7. Choose input and target data
denotes that they perform function as a multiplier. Each node Step 8. Train ANFIS with the dataset
in this layer multiplies the input signals to determine the Step 9. Test the models
firing strength of each rule before sending the result out. The Step 10. Compute and record the Mean square error and
equation is written as follows: coefficient of determination(R2 ) for the ANFIS-based detec-
tor, classifier, and location to check the performance of the
O2i = wi = πj=1
m
µAi (x) (2)
ANFIS model.
Layer 3:
The nodes in this layer are fixed nodes as well. The nodes
2) BACK PROPAGATION NEURAL NETWORK (BPNN)
marked with a N, as shown in Fig.3, indicate that the firing
The ability of ANNs to work with complex systems makes
strength has been normalized from the preceding layer. The
ANNS a good network for fault detection, classification, and
equation for this layer is represented as follows:
wi location [25].
O3i = w = (3) Some of the properties that make ANNs better to work for
w1 + w2 complex and nonlinear tasks are as follows:
Layer 4:
a) Every electrical system fault causes a change in the
Each node in this layer is an adaptive node, and the output
power system, and a neural network (NN) can reorga-
parameters are modified in this layer. The output, which com-
nize according to these changes
bined the first-order polynomial with the normalized firing
b) ANN can make decisions and learn through
strength, was processed for each node in the layer. As a result,
experience
this layer’s outputs are provided as follows:
c) Due to their numerical strength, they can carry out
O4i = yi = Wi fi = Wi (pi x1 + qi x2 + ri ) , i = 1, 2, .. (4) several tasks at once.

49020 VOLUME 12, 2024

S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

FIGURE 4. Flow chart for ANFIS-based fault model.

The ANN offers a lot of benefits, but it also has certain the models.
drawbacks. The choice of network type, the number of hidden v
u n
layers, the number of neurons, and the settings of the learning uX
RMSE = t ypredicted − yactual

2
(7)
method are some crucial elements [4]. There are several limi- i=1
tations, such as post-fault values of line currents and voltages
for fault identification and classification. The line current This way, it generates the error for fault detection, classifi-
and voltage of the fault transmission lines are considerably cation, and location against the fault generated in the IEEE
different before and after the incident. Therefore, determining 9-bus system for each case of fault. The final output of BPNN
the type of faults from pre- and post-fault value patterns is generated for sample j can be calculated as below:
necessary for the fault classification procedure. The BPNN n
X
basic structure and the network used in this paper are given in yj = ωk,j f (hk ) + βj (8)
Fig. 5a and 5b, respectively. Correct weight tuning in BPNN j=1
reduces the error rate [30]. Weights are selected at random. The ωk,j in equation 8 is the weight of hidden node k at the
The consequences are updated after each iteration or epoch, jth epoch. βj is the bias value of the output node at the jth
and the procedure is repeated. The following are some of age. f (hk ) is the output value of hidden node k. The fitness
BPNNs’ most valuable benefits: of values can be calculated using the Error in the equation
Backpropagation may function without having thorough below.
network expertise since it is simple, quick, and straightfor-
ward to implement [7], [32]. Fitness (x) = minimum(RMSE) (9)
As shown in Fig. 5c the simple flow diagram of
ANN-based fault detection. It just requires tuning a small 3) SELF ORGANIZING MAP (SOM) NEURAL NETWORK
number of input parameters, such as the number of hid- SOM is an extensively used neural model in the competitive
den layers and neurons in each hidden layer. Levenberg- learning network category. It uses unsupervised learning.
Marquardt backpropagation is used to modify the weights and Therefore, there is no need for human assistance during train-
biases of the training process. Additionally, the mean absolute ing or extensive prior knowledge of the input and output data.
error (MAE), root mean square error (RMSE), and mean SOM is used for tasks like clustering input data and locating
square error (MSE) can be used to calculate learning errors. the faults.
The following equation 7 is used to compute RMSE, The self-organizing map (SOM), which has strong visu-
which has been used in this paper for the performance of alization capabilities, is particularly well suited for data

VOLUME 12, 2024 49021

 S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

FIGURE 5. (Continued.) c. Flow chart for ANN-based fault model.

categorization tasks [26]. It produces a collection of pro-

totype vectors to represent the data set. It projects the
prototypes from the high-dimensional input space onto a
low-dimensional grid while maintaining their structure. This
ordered grid may be a practical visualization surface to dis-
play various SOM properties, such as the cluster structure.
The structure is shown in Fig. 6a [21].
The voltage and current data set is first clustered using
the Kohonen-mean clustering method. The k-means cluster-
ing technique divides a given data set into a predetermined
number (k) of clusters.
The centroid’s k number is initially picked. A data point at
the cluster’s center is called the centroid.
For this work, a two-dimensional SOM is used. A 2D
lattice of ‘nodes’, each completely linked to the input layer,
makes up the network. A Kohonen network of 8∗ 8 nodes con-
nected to the input layer and representing a two-dimensional
vector can be seen in Fig. 6a. Each node in the lattice has
a distinct topological location and a vector of weights that
has the same dimension as the input vectors. Afterward,
a matching weight vector W with n dimensions is mapped
with clustered input, as shown in Fig. 6a.
Unlike many other types of networks, a SOM does not need
a specific target output. Instead, the area of the lattice where
the node weights match the input vector is specifically tuned
to resemble the data for the class the input vector belongs to
more closely. The SOM finally settles into a map of stable
zones after starting with a distribution of random weights
and going through several iterations. Training of SOM takes
several steps [27], [31], which are mentioned below.
Step 1. Weight initialization
Step 2. Best matching unit (BMU) calculation using
Euclidean distance method.
The Euclidean distance can be calculated using the
following:
r
Xi=n
FIGURE 5. a. BPNN basic structure [7] b. ANN structure for Fault
detection [20].
Distance = (Xi − Wi )2 (10)
i=0

49022 VOLUME 12, 2024

S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

FIGURE 6. a. SOM structure with clustered input b. Flow chart for SOM-based fault model.

In equation (10), Wi is the input vector, and Wi is the weight such as grid size and number of input features for fault detec-
vector of nodes. tion, classification, and location, are defined using the Matlab
Step 3. Calculating the size of BMU tool first, and changes are made in the SOM code according
 n to the learning rate. The size of the grid as we test the grid
σ (n) = σ0 exp − (11) for larger size; the higher the size, the better. We tested for as
T
In equation 11, σ (n) is the width of the lettuce the at time high as a 60∗60 grid. After defining the features, we train the
zero, n is the epoch, and T is the time constant SOM model for detection, classification, and Location. After
Step 4. Weights adjustment can be done using equation (12) training, the training data and testing points to the SOM grid
and (13) using the SOM function in Matlab [20].
Then, the fault type, fault detected value, and fault location
Wnew = Wold +L (X−Wold ) (12) within each mapped neuron were retrieved for training data.
W (n + 1) = W (n) + L (n) (X (n) − W (n)) (13) The flow diagram of the whole process is given in Fig. 6b. The
In equations (12) and (13), L is the learning rate, Wnew new weights generated by the SOM model for each input is given
weights, Wold is old weights, and X is the input vector. in Fig. 8a and Fig. 8b shows the distance created between all
The learning rate is calculated at each epoch using the neighboring weights.
equation (14). 4) DISCRETE WAVELET TRANSFORM AND ANFIS
 n
L (n) = L0 exp − (14) Wavelet transform, using information from the frequency
T domain and time domain, can detect and classify differ-
In this paper, we preprocessed the data: six features of three- ent kinds of faults. Wavelet transform is sensitive to signal
phase voltage and current. Meanwhile, the SOM parameters, abnormalities [22]. In this algorithm, the DWT is employed

VOLUME 12, 2024 49023

 S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

to extract the detailed features of the current signal using the

mother wavelet ‘‘db1’’. After several calculations using sev-
eral wavelet families, it was determined that the Daubechies
family’s ‘‘db1’’ wavelet can accurately identify fault in the
current signal [23]. The original current signals are divided
into several frequency levels using wavelet algorithms. The
mother wavelet is scaled and shifted throughout the signal to
achieve this. DWT is used because the data being analyzed is
digital [24].
The steps used to apply DWT are as follows.
Step 1. Record the current signal at bus number 7 in the
IEEE 9-bus system.
Step 2. Apply DWT using the below equation [6], [33].
Z ∞  
1 ∗ t − q2p
D= √ x(t)ϕ dt (15)
2p −∞ 2p
where p and q are the wavelet’s scale and the wavelet func-
tion’s positions. x(t) represents the current signal recorded at
bus 7. ϕ ∗ is the complex conjugate and the mother wavelet.
Step 3. Decompose the current signal into high-frequency
components and low-frequency components.
Step 4. Recording detailed components and approximate
components to use it with ANFIS. It can be represented
mathematically as.
X
a (t) = x (t) .hq (2t − q) (16)
q
X FIGURE 7. Flow chart for DWT-ANFIS-based fault model.
d (t) = x (t) .gq (2t − q) (17)
q
values of three-phase current and three-phase voltage, are
where a(t) is the approximate coefficient of the current signal collected and preprocessed for the next step. 80% of the total
extracted using the low pass filter (hq ), and d(t) is the detail data is used as training data, and 20% is used as testing
coefficient of the current signal extracted using the high pass data. For the development of ANFIS-based fault detection,
filter gq . classification, and Location models, a Fuzzy C-mean (FCM)
Step 5. Use an Approximate and detailed coefficient and Clustering algorithm is used. The total clusters used are 50,
apply ANFIS by following the steps in ANFIS. After the and the total membership for each cluster is 50, too. The
training and testing, record the results to check the model’s Gaussian input membership function is further used as a
performance. As shown in the flow diagram in Fig. 7. membership function for inputs.
After generating a trained ANFIS network for fault detec-
C. TRAINING AND TESTING tion, classification, and Location, the trained network is saved
1) TRAINING TESTING DATA and used with the simulated model to get the results. The
For developing neural network-based models, arranging input generated ANFIS Network is presented in Fig.10.
and targets for a particular task is necessary. The two basic An ANN-based fault model for detection, classification,
processes in a Neural network are training and testing with and location, a MATLAB toolbox, is used. The generated
the input and target data to make predictions for upcoming MATLAB code is further used. To adjust the parameters to
events, in our case, predictions about fault detection, fault make the model better. 80% of the data is used for training,
classification, and fault location. Training and testing input and 20% is used for testing and validation. Two hidden layers
and out-sample data is mentioned in Table 1. are used, and the hidden neurons in each layer are 18 and 15,
For ANFIS, ANN, SOM, and DWT-ANFIS-based mod- respectively, with sigmoid and tan sigmoid membership func-
els, data is collected from bus number 7 in the IEEE tions.
9-bus system after generating eight different types of faults The workflow for the developed ANFIS model is given
and for no-fault conditions. The eight fault cases are A-G, in Fig. 4.
B-G, C-G, AB, BC, CA, AB-G, BC-G, CA-G faults. All For SOM, the train data is the root mean squared value
these faults are generated at four locations at the transmis- of three-phase voltage and current. The dimension defined
sion line between bus number 7 and bus number 8 in the for the SOM for detection is 8∗8. The clusters generated
IEEE 9-bus model simulated in MATLAB/Simulink [20]. are mapped with the target to evaluate the performance. The
Training and testing data, in the form of root mean square clustered mapping is shown in Fig. 6a.

49024 VOLUME 12, 2024

S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

FIGURE 8. a). Weights for each input, b) Neighbor distance between weights.

TABLE 1. Sample data set for the ANFIS-based fault detection, classification, and for location.

2) FAULT DETECTION model of DWT-ANFIS, the detail coefficient of the faulty

The ANFIS-based fault detection model detects the fault current signal is used in addition to the raw voltage and
while analyzing the three-phase voltage and current signal. current signal values for training and testing of the DWT-
The RMS value of three-phase voltage and post-fault current ANFIS model.
are inputs for the fault detection model. The output of the fault
detection model is two binary states: 1 or 0. The one indicates 3) FAULT CLASSIFICATION
a fault, and 0 shows a no-fault condition. The fault detection The ANFIS-based fault classification model is designed and
modal structure in MATLAB for ANN and ANFIS is given trained for the same input as the fault detection model. The
in Fig. 5b and Fig.10. output for fault classification is changed; for fault classifica-
Fault detection takes six inputs (three-phase voltage and tion, we used one output indicating the fault class. Table 1
three-phase current) and one output to display whether there shows that the fault class output is eight numerical values for
is a fault or no fault in the system. eight different fault classes and 0 for no-fault conditions.
The same input and output datasets are used for ANN and The ANFIS-based fault classification structure is given
SOM for training and testing, and trained models are used in Fig. 7. The total input for the fault classification model
with the system for fault detection. For the fault detection is 6, and the output is 1. When the trained model is tested

VOLUME 12, 2024 49025

 S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

TABLE 2. Performance of ANFIS -based fault detection, classification and location model.

with the system post-fault rms values of three-phase voltage generated. According to the results shown in Table 1, the
and current, the model gives the output based on the fault system has the respective fault.

49026 VOLUME 12, 2024

S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

TABLE 3. Performance of ANN based fault detection, classification and location model.

The same inputs and outputs are employed by ANN and For the fault classification model of DWT-ANFIS, the detail
SOM-based classification models for training and testing. coefficient of the faulty current signal is used in addition to

VOLUME 12, 2024 49027

 S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

TABLE 4. Performance of SOM -based fault detection, classification and location model.

49028 VOLUME 12, 2024

S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

TABLE 5. Performance of DWT-ANFIS-based fault detection, classification, and location model.

the raw voltage and current signal values for training and 4) FAULT LOCATION
testing of the DWT-ANFIS model. The performance of all ANFIS fault location trained model is developed for six
models is simulated and recorded in Tables 6 and 7 for all the inputs and one output for four different locations. The Loca-
models. tion is in kilometers (km), and the eight various faults are

VOLUME 12, 2024 49029

 S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

TABLE 6. Percentage error comparison of Ann, Anfis, Wt-Anfis, and som models.

FIGURE 9. The performance of ANFIS, ANN, SOM, and DWT-ANFIS models for testing: a. Fault detection, b: fault classification, c: fault location.

generated at 100km, 75km, 50km, and 25km. The model The first six columns in the Table are six inputs, and the
is trained with the generated data for 40 different fault last three columns are output data.
conditions.
The trained model is then tested with different conditions III. RESULT AND COMPARISON OF MODELS
and three-phase current and voltage values; the results are A total number of 40 different fault conditions have been
given in the next section. studied. Ten fault types are generated in the IEEE 9-bus sys-
The same inputs and outputs as in ANFIS are used with tem at four locations. Nine asymmetrical faults and no-fault
ANN and SOM models for training and testing to find the states to analyze the performance of the proposed technique.
Location of the fault generated in the system. For the Fault These faults were caused on transmission lines at four differ-
detection model of DWT-ANFIS, the detail coefficient of the ent locations at different times.
faulty current signal is used in addition to the raw voltage The ANFIS model is trained and tested for six inputs and
and current signal values for training and testing of the DWT- one output for fault detection. The results from the training
ANFIS model. The results for all the models are displayed in and testing of ANFIS are given in Table 2. Predicted fault
the results section in Tables 6 and 7. detection, classification, and Location have been recorded,

49030 VOLUME 12, 2024

S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

TABLE 7. Training and testing performance of ANN, ANFIS, DWT-ANFIS and SOM based models.

and percentage error has been computed for each case. The Fault detection in the SOM-based model is computed in
average percentage error for ANFIS-based fault detection, Table 4. The SOM clustered accurately for fault detection
type, and Location is 0.008%. -0.005% and 0.547%, respec- for all the faults and did not miss any fault with zero per-
tively. Which is less than zero. Compared to other models, centage error. In the case of SOM-based fault classification,
the based model can detect all faults accurately, as shown in the percentage error for all classifications is zero. However,
Table 2, it detected all faults correctly. For Fault classifica- the percentage error is more than one percent for AB-G fault
tion, the ANFIS model classified all fault classes with less classification at 100 km, as shown in Table. 4. Fault location is
than zero percent error. However, for the AB-G fault, the per- predicted well for most fault locations with zero percent error;
centage error is -1.224 at 100 km. For fault location, ANFIS however, it cannot detect the Location for two fault conditions
performed well with an average percent error of 0.547%. The at 100km and once for 50km, as shown in Table 4. The weight
root mean square error and mean absolute error are calculated plane representation of the SOM classes displaying each of
and recorded in Table 7. the six input features is shown in Fig. 8 (a), providing a visual
ANN-based models are given in Table 3. The average depiction of the weights that link each input to a single neuron
percentage error is recorded in Table 6. The average percent- in the 6∗8 hexagonal grid. Heavy weights are indicated by
age error for ANN-based Fault detection, classification, and darker hues than lighter ones. Six inputs are highly correlated
Location is zero, 0.268, and 0.348 percent. The root mean when they have comparable weight planes (color gradients
square error. It can detect the fault correctly with zero percent can be inverted or identical).
error and classify the error with less than zero percent error It was also shown that the weight ratios of voltage and
except for phase C to A fault at 25km and phase A to phase B current are opposite to one another when all four variables
fault at 50 km. are weighted equally. Consequently, the Euclidean interval

VOLUME 12, 2024 49031

 S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

achieved precise location predictions, as depicted in Fig. 9c,

outperforming the other models in this regard.
The performance of models in the form of RMSE given in
equation, mean absolute error, and R2 is recorded in Table 7.
 
Predicted value − Actual value
%Error = ∗ 100 (18)
Actual value
Xn |Predicted value − Actual value|
MAE = (19)
i=1 n
The output graph between predicted values and actual fault
is given in Fig. 9a for ANFIS, ANN, SOM, and DWT-ANFIS
models. Similarly, Fig. 9b shows the predicted value and
actual classification for fault for all four models. Further,
Fig. 9c represents the exact fault location and predicted Loca-
FIGURE 10. ANFIS network for fault models [20], [28].
tion for model ANFIS, ANN, SOM, and DWT-ANFIS.

between each neuron’s class and its surroundings is shown IV. CONCLUSION
in Fig. 8(b). In this study, we compared the effectiveness of several meth-
The input space’s highly related regions are indicated by ods, including Artificial Neural Networks (ANN), Adaptive
the vivid connections. On the other hand, groups representing Neuro-Fuzzy Inference Systems (ANFIS), Self-Organizing
areas of the function space that have few or no members Maps (SOM), and a hybrid approach that combines Discrete
separating them are shown by the dark links. Wavelet Transform (DWT) with ANFIS, for fault detection,
Large swaths of the input space are divided by extended classification, and localization in transmission lines. Our
boundaries with dark connections, indicating that the groups study set out to investigate the effectiveness of these tech-
on either side of the boundary reflect substantially different niques in managing the complex tasks (faults) that are a part
neighborhoods. of power transmission networks.
The DWT-ANFIS-based fault model predicted results are Our findings show that the average percentage error for
computed and recorded in Table 5. The percentage error for each of the approaches under investigation differs signif-
fault detection for all the fault conditions is zero, as shown icantly. The average percentage error for the ANN-based
in Table 5. Performance of DWT-ANFIS based model needs models was 0.039, which is a good result. But with a
more improvement as the average percentage error for fault noticeably lower average percentage error of 0.0015, the
classification is -0.364% as shown in Table 6. For fault loca- ANFIS-based models beat their ANN counterparts, demon-
tion, the DWT-ANFIS-based model detected fault location strating the greater adaptability and robustness of the ANFIS
with a percentage error less than zero for most cases. Still, for technique in fault identification.
some, the percentage error is 1 percent, as shown in Table 5. The SOM-based models showed an average percentage
The average error for this model is recorded in Table 6. error of -0.197, which can be explained by the special proper-
The average percentage error for all the models is com- ties of SOMs and their effective mapping of high-dimensional
puted and recorded in Table 6. data to low-dimensional space. This negative error number
The performance of various models including ANN, shows how well the SOM-based models handled complicated
ANFIS, SOM, and DWT-ANFIS are illustrated in transmission line fault situations by continuously producing
Fig. 9 a, b, and c, representing fault detection, classifica- findings that were closer to ground truth values.
tion, and location respectively. Fig. 9a showcases the fault Additionally, with an average percentage error of -0.103,
detection capabilities of these models across forty distinct the hybrid strategy that combined DWT with ANFIS demon-
scenarios, each assessed with a testing dataset comprising strated a significant improvement. This enhancement high-
400 instances. The simulation was conducted for fault (1) lights how well DWT and ANFIS operate together, using
and no-fault (0) conditions, with all models demonstrating DWT’s signal processing capabilities to improve the ANFIS
accurate detection capabilities for both scenarios. framework’s fault detection, classification, and location
For fault classification, depicted in Fig. 9b, the models’ accuracy.
performance is evaluated across ten fault classes (AG, BG, Overall, our comparative analysis demonstrates how well
CG, AB, BC, CA, ABG, BCG, CAG) as well as the no-fault ANFIS and hybrid approaches work to address the problems
case. While overall performance was satisfactory, certain related to fault localization, classification, and detection in
fault classes exhibited lower accuracy, particularly with ANN transmission lines. These results highlight the need of utiliz-
and DWT-ANFIS models. ing cutting-edge computational methods and hybrid strategies
In the context of fault location prediction, the models were to improve the dependability and effectiveness of electricity
tested across four distinct locations. ANFIS and SOM notably transmission networks.

49032 VOLUME 12, 2024

S. Kanwal, S. Jiriwibhakorn: Advanced Fault Detection, Classification, and Localization

REFERENCES [22] I. M. Karmacharya and R. Gokaraju, ‘‘Fault location in ungrounded photo-

[1] D. Lu, Y. Liu, Q. Liao, B. Wang, W. Huang, and X. Xi, ‘‘Time-domain voltaic system using wavelets and ANN,’’ IEEE Trans. Power Del., vol. 33,
transmission line fault location method with full consideration of dis- no. 2, pp. 549–559, Apr. 2018, doi: 10.1109/TPWRD.2017.2721903.
tributed parameters and line asymmetry,’’ IEEE Trans. Power Del., vol. 35, [23] M. Bhatnagar, A. Yadav, and A. Swetapadma, ‘‘A resilient protection
no. 6, pp. 2651–2662, Dec. 2020, doi: 10.1109/TPWRD.2020.2974294. scheme for common shunt fault and high impedance fault in distri-
[2] A. Yadav and A. Swetapadma, ‘‘A single ended directional fault section bution lines using wavelet transform,’’ IEEE Syst. J., vol. 16, no. 4,
identifier and fault locator for double circuit transmission lines using pp. 5281–5292, Dec. 2022, doi: 10.1109/JSYST.2022.3172982.
combined wavelet and ANN approach,’’ Int. J. Electr. Power Energy Syst., [24] C. Hwan Kim and R. Aggarwal, ‘‘Wavelet transforms in power systems.
vol. 69, pp. 27–33, Jul. 2015, doi: 10.1016/j.ijepes.2014.12.079. Part 1: General introduction to the wavelet transforms,’’ Power Eng. J.,
[3] Z. Jiao and R. Wu, ‘‘A new method to improve fault location accu- vol. 14, no. 2, pp. 81–87, Apr. 2000, doi: 10.1049/pe:20000210.
racy in transmission line based on fuzzy multi-sensor data fusion,’’ [25] F. Martin and J. A. Aguado, ‘‘Wavelet-based ANN approach for trans-
IEEE Trans. Smart Grid, vol. 10, no. 4, pp. 4211–4220, Jul. 2019, doi: mission line protection,’’ IEEE Trans. Power Del., vol. 18, no. 4,
10.1109/TSG.2018.2853678. pp. 1572–1574, Oct. 2003, doi: 10.1109/TPWRD.2003.817523.
[4] A. Abdullah, ‘‘Ultrafast transmission line fault detection using a DWT- [26] N. Lightowler and H. Nareid, ‘‘Artificial neural network based control
based ANN,’’ IEEE Trans. Ind. Appl., vol. 54, no. 2, pp. 1182–1193, systems,’’ SAE transactions, vol. 112, pp. 43–539, Jan. 2003.
Mar. 2018, doi: 10.1109/TIA.2017.2774202. [27] C.-C. Hsu, ‘‘Generalizing self-organizing map for categorical data,’’
[5] M. T. Hoq, ‘‘Distance protection of transmission lines with high levels IEEE Trans. Neural Netw., vol. 17, no. 2, pp. 294–304, Mar. 2006, doi:
of series compensation: A study on frequency and time domain commu- 10.1109/TNN.2005.863415.
nication independent distance protection for series compensated lines,’’ [28] S. Kanwal and S. Jiriwibhakorn, ‘‘Artificial intelligence based faults
Ph.D. dissertation, Dept. Elect. Eng., KTH Roy. Inst. Technol., Stockholm, identification, classification, and localization techniques in transmission
Europe, 2023. lines—A review,’’ IEEE Latin Amer. Trans., vol. 21, no. 12, pp. 1291–1305,
Dec. 2023, doi: 10.1109/TLA.2023.10305233.
[6] M. J. B. Reddy and D. K. Mohanta, ‘‘Performance evaluation of an
[29] S. Jiriwibhakorn, ‘‘Critical clearing time prediction for power transmission
adaptive-network-based fuzzy inference system approach for location of
using an adaptive neuro-fuzzy inference system,’’ IEEE Access, vol. 11,
faults on transmission lines using Monte Carlo simulation,’’ IEEE Trans.
pp. 142100–142110, 2023, doi: 10.1109/ACCESS.2023.3341968.
Fuzzy Syst., vol. 16, no. 4, pp. 909–919, Aug. 2008.
[30] N. A. M. Leh, F. M. Zain, Z. Muhammad, S. A. Hamid, and
[7] D. P. Mishra and P. Ray, ‘‘Fault detection, location and classification of a
A. D. Rosli, ‘‘Fault detection method using ANN for power transmis-
transmission line,’’ Neural Comput. Appl., vol. 30, no. 5, pp. 1377–1424,
sion line,’’ in Proc. 10th IEEE Int. Conf. Control Syst., Comput. Eng.
Sep. 2018, doi: 10.1007/s00521-017-3295-y.
(ICCSCE), Penang, Malaysia, Aug. 2020, pp. 79–84, doi: 10.1109/ICC-
[8] E. E. Ngu and K. Ramar, ‘‘A combined impedance and traveling wave
SCE50387.2020.9204921.
based fault location method for multi-terminal transmission lines,’’ Int.
[31] J. Tian, H. A. Michael, and P. Michael, ‘‘Anomaly detection using self-
J. Electr. Power Energy Syst., vol. 33, no. 10, pp. 1767–1775, Dec. 2011,
organizing maps-based k-nearest neighbor algorithm,’’ in Proc. PHM Soc.
doi: 10.1016/j.ijepes.2011.08.020.
Eur. Conf., vol. 2, no. 1, 2014, pp. 1–9.
[9] A. Mukherjee, P. K. Kundu, and A. Das, ‘‘Transmission line faults in power
[32] Z. I. Sobhuza and B. A. Thango, ‘‘Study of fault detection on a 230kV
system and the different algorithms for identification, classification and
transmission line using artificial neural network (ANN),’’ in Proc. 31st
localization: A brief review of methods,’’ J. Inst. Engineers India: Ser. B,
Southern Afr. Universities Power Eng. Conf. (SAUPEC), Jan. 2023,
vol. 102, pp. 1–23, Jan. 2021, doi: 10.1007/s40031-020-00530-0.
pp. 1–6.
[10] S. Devi, N. K. Swarnkar, S. R. Ola, and O. P. Mahela, ‘‘Detec- [33] V. Veerasamy, N. I. Abdul Wahab, R. Ramachandran, M. Thirumeni,
tion of transmission line faults using discrete wavelet transform,’’ in C. Subramanian, M. L. Othman, and H. Hizam, ‘‘High-impedance fault
Proc. Conf. Adv. Signal Process. (CASP), Jun. 2016, pp. 133–138, doi: detection in medium-voltage distribution network using computational
10.1109/CASP.2016.7746152. intelligence-based classifiers,’’ Neural Comput. Appl., vol. 31, no. 12,
[11] M. Jamil, S. K. Sharma, and R. Singh, ‘‘Fault detection and classification pp. 9127–9143, Dec. 2019.
in electrical power transmission system using artificial neural network,’’
SpringerPlus, vol. 4, no. 1, pp. 1–13, Dec. 2015, doi: 10.1186/s40064-015-
1080-x.
[12] M. N. M. Salleh, N. Talpur, and K. Hussain, ‘‘Adaptive neuro-fuzzy SHAZIA KANWAL (Graduate Student Member,
inference system: Overview, strengths, limitations, and solutions,’’ in Proc. IEEE) received the B.S. degree in electronics from
Data Mining Big Data, 2nd Int. Conf., vol. 2, Jul. 2017, pp. 527–535. the Sir Syed University of Engineering and Tech-
[13] M. Paul and S. Debnath, ‘‘ANFIS based single line to ground fault location nology, Karachi, Pakistan, and the M.Sc. degree in
estimation for transmission lines,’’ in Proc. Michael Faraday IET Int. electronics from the King Mongkut’s University
Summit (MFIIS), vol. 2020, Oct. 2020, pp. 69–74, doi: of Technology North Bangkok, Thailand. She is
10.1049/icp.2021.1077. currently pursuing the D.Eng. degree in electrical
[14] Y. Zhang, Q. Zhang, W. Song, Y. Yu, and X. Li, ‘‘Transmission line engineering with the School of Engineering, King
fault location for double phase-to-Earth fault on non-direct-ground neutral Mongkut’s Institute of Technology Ladkrabang
system,’’ IEEE Trans. Power Del., vol. 15, no. 2, pp. 520–524, Apr. 2000, (KMITL).
doi: 10.1109/61.852978.
Her research interest includes artificial intelligence-based fault analysis in
[15] O. E. Obi, ‘‘An extended ANN-based high speed accurate transmission
transmission lines.
line fault location for double phase to Earth fault on non-direct-ground,’’
Int. J. Eng. Sci. Technol., vol. 1, no. 1, pp. 31–47, Aug. 2019, doi:
10.29121/ijoest.v1.i1.2017.04.
[16] I. Abdulrahman, ‘‘MATLAB-based programs for power system dynamic SOMCHAT JIRIWIBHAKORN (Member, IEEE)
analysis,’’ IEEE Open Access J. Power Energy, vol. 7, pp. 59–69,
received the B.Sc. and M.Sc. degrees in electrical
Nov. 2020, doi: 10.1109/OAJPE.2019.2954205.
engineering from the King Mongkut’s Institute
[17] P. M. Anderson and A. A. Fouad, Power System Control and Stability,
2nd ed. Hoboken, NJ, USA: Wiley, 2008. of Technology Ladkrabang (KMITL), Bangkok,
[18] P. Sauer, M. A. Pai, and J. H. Chow, Power System Dynamics and Stability. Thailand, in 1994 and 1997, respectively, and the
Hoboken, NJ, USA: Wiley, 2017. Ph.D. degree in electrical engineering from the
[19] J.-S. R. Jang, ‘‘ANFIS: Adaptive-network-based fuzzy inference system,’’ Imperial College London, U.K., in 2000.
IEEE Trans. Syst., Man, Cybern., vol. 23, no. 3, pp. 665–685, Jun. 1993, He has been an Associate Professor with the
doi: 10.1109/21.256541. Department of Electrical Engineering, School of
[20] MATLAB, document R2019b, Version: 9.7.0, MathWorks, Inc, 2022. Engineering, KMITL, since 2006. His research
[21] J. Vesanto and E. Alhoniemi, ‘‘Clustering of the self-organizing map,’’ interests include power system stability, optimization, planning and forecast-
IEEE Trans. Neural Netw., vol. 11, no. 3, pp. 586–600, May 2000, doi: ing, applications of neural networks, and ANFIS in power engineering.
10.1109/72.846731.

VOLUME 12, 2024 49033