Ain Shams Engineering Journal 15 (2024) 102427

Contents lists available at ScienceDirect

Ain Shams Engineering Journal

journal homepage: www.sciencedirect.com

Deep learning techniques for transmission line fault classification – A

comparative study
Priyanka Khirwadkar Shukla, K. Deepa *
Department of Electrical and Electronics Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Bengaluru 560035, India

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

Keywords: Despite advancements in technology, power system faults leading to electric power interruption remain a sig­
Accuracy nificant issue. Efficient restoration of the power system relies on the swift classification and clearance of faults.
Artificial neural network Among the various types of faults in transmission lines, open circuit and short circuit faults are commonly
Deep learning techniques
encountered. This study specifically focuses on the analysis of five types of short circuit faults: line-to-line, line-
Fault classification and long short-time
memory
to-ground, double line-to-ground, triple line, and triple line-to-ground faults. Faults can cause both power failure
and power loss in transmission lines. Once a fault occurs, it is crucial to restore electricity supply promptly to
prevent further losses. Therefore, the development of a system capable of accurately and swiftly detecting and
removing faults is essential. Traditionally, categorizing transmission line faults required sophisticated mathe­
matical modeling, intricate signal processing techniques, and expert knowledge to interpret the output signals. In
this paper, an alternative approach is proposed, utilizing deep learning techniques for transmission line fault
classification. Specifically, the paper employs techniques such as artificial neural network (ANN), long short-term
memory (LSTM), with and without window regression (WR). By implementing these deep learning techniques,
automatic feature extraction and signal processing are achieved, streamlining the fault classification process. The
experimental results obtained from this study demonstrate an accuracy of 42.98% for ANN, 99.98% for LSTM,
and 99.99% for LSTM-WR. These outcomes underscore the effectiveness of the deep learning techniques
employed in accurately classifying transmission line faults, with LSTM and LSTM-WR outperforming ANN in
terms of accuracy.

1. Introduction hampered when DGs are included into distribution networks [3,4]. The
usage of DERs, such as photovoltaic (PV) systems and wind, is expanding
Modern enterprises and the progress of our civilization depend on a globally, however due to the advent of nonlinear power electronics
steady supply of power. Power system problems may come from factors devices and loads, they introduce a number of power quality issues [5].
like as losses, corona discharge, and other issues that affect power Similar to this, the use of bidirectional electric vehicle charging stations
transmission and distribution. These flaws prevent customers from (EVCS) [8] is transforming the distribution networks due to the rising
receiving consistent energy supplies, which causes maintenance prob­ popularity of electric cars (EV) [6,7]. Therefore, as more and more DERs
lems, user unhappiness, and financial losses for enterprises. It is essential combine, it is getting harder to maintain the distribution network’s
to address these issues right once since they might result in protracted secure and dependable operation [9].
disruptions that impede corporate growth and interfere with routine To protect the power system network, a failure must be accurately
duties that depend on a reliable power supply. Every day, there is an located, detected, and categorized as soon as it occurs. As a result, there
increase in the need for electrical power supply. Distributed generators will be as little network outage as possible when the problematic phase
(DGs) are consequently being included into distribution networks to is unplugged [10]. Different fault detection, classification, and locali­
meet local demands [1,2]. Different fault current levels exist depending zation algorithms have emerged in the power system during the past few
on the kind of distributed energy resource (DER) connected to the sys­ years as a result of numerous research programs [11–13]. Manual
tem. The capacity of ordinary relays to recognize and classify faults is feature extraction is often required by conventional fault classification

Peer review under responsibility of Ain Shams University.

* Corresponding author.
E-mail addresses: [email protected] (P.K. Shukla), [email protected] (K. Deepa).

https://doi.org/10.1016/j.asej.2023.102427
Received 24 April 2023; Received in revised form 23 June 2023; Accepted 6 August 2023
Available online 9 August 2023
2090-4479/© 2023 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Engineering, Ain Shams University. This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
P.K. Shukla and K. Deepa Ain Shams Engineering Journal 15 (2024) 102427

methods based on mathematical modelling and signal processing; may have, considerable memory resources may be required for storing
however, this can be time-consuming and may not adequately extract all data during both training and inference. The size of the training dataset
the relevant information from the data. These strategies’ substantial may also have an effect on how much storage is required. The amount of
dependence on expert knowledge and assumptions about the underlying memory and storage space available to support the DL models and
fault processes may make it difficult to apply them to complex and associated data may need to be taken into consideration in resource-
variable fault scenarios. Since, these methods usually rely their as­ constrained scenarios in real-time power systems [26].
sumptions on linear and stationary systems, they may also have issues DL algorithms’ training and testing processes may be made more
with nonlinear and non-stationary signals [14–16]. efficient to reduce the amount of time needed for fault categorization.
The detection, classification, and localization of defects in the power The following strategies can improve the efficacy of these processes:
system have the potential to be greatly aided by methods based on
machine learning (ML) and deep learning (DL) [17,18]. Since DL algo­ 1. Model optimization techniques:
rithms train from start to finish, no suitable feature extractor needs to be Model optimization techniques help speed up the training and
made. On the other hand, deep learning techniques like convolutional testing processes. These techniques include network design optimi­
neural networks (CNNs) and recurrent neural networks (RNNs) may zation, parameter initialization strategies, regularization techniques,
automatically learn features from the raw data, doing away with the and optimization algorithms. By carefully designing and optimizing
requirement for manual feature engineering. The flexibility and ability the deep learning model, it is possible to speed up convergence
of DL models to identify complex patterns and correlations in the defect during training and reduce inference time during testing [25].
data are increased by the ability of DL models to infer hierarchical 2. Data Augmentation and Preprocessing:
representations from the data. They are also better suited to handle Data augmentation techniques like rotation, scaling, and flipping
nonlinear and non-stationary data because of their potential to learn and can help increase the diversity of the training data without the
predict nonlinear dynamics. Additionally, DL algorithms have shown requirement for further sample gathering. This improvement could
positive results for defect classification tasks across a range of domains, lead to faster convergence and better generalization during training.
including image analysis, audio processing, and sensor data analysis Additionally, preprocessing methods like feature scaling and
[19]. These tactics have the potential to be more accurate and resilient normalization can improve the quality of the data and reduce noise
than traditional methods [14–16]. Consequently, DL approaches, in to improve the performance of both training and testing operations.
general, offer significant advantages over conventional fault classifica­ [26].
tion methods by automating feature extraction, capturing complex fault 3. Parallelization and distributed computing:
patterns, generalizing to new fault scenarios, requiring less data pre- It can significantly shorten training times by using many GPUs or
processing, and scaling successfully to huge datasets. distributed computing frameworks. By dividing the workload across
DL has been used to diagnose and categorize faults across a variety of various computing resources and employing techniques like data
industries and fault types, including industrial processes and parallelism and model parallelism, deep learning model training may
manufacturing, HVAC systems [20], mechanical systems for acoustic be sped up [27].
signal analysis, misalignment, imbalance, wear or tear, rolling bearings
using time–frequency image analysis of vibration signals [21], and For the real-time implementation of the DL algorithms, the following
HVAC systems [20]; rolling bearings using time–frequency image steps must be taken into consideration:
analysis of vibration signals [21]; in medical faults in medical imaging,
electrocardiograms and electroencephalograms; for electrical faults like 1. Gathering and preparing data:
voltage sag or swell, harmonics, power quality issues, transformer faults, To enable real-time fault identification, pertinent data from the
motor faults, generator faults, transmission and distribution system power system must be gathered, such as current and voltage mea­
faults, or in wind turbines [22] and photovoltaic system [23] under surements. After normalising the values and ensuring that the data is
different operation conditions. of good quality, the data is then ready to be entered into the DL
These are only a few examples; many more fault detection and model [28,29].
classification tasks across other disciplines have been successfully 2. Handling real-time data streaming:
completed using DL approaches. Due to their versatility and adapt­ The collected data must be disseminated in real-time in order for
ability, these models may be utilized for a number of fault analysis ap­ the DL model to detect problems. For this, it is required to set up an
plications as long as there is enough labelled data and the models are appropriate data streaming infrastructure that can control the high-
created and trained properly. frequency input and give it to the model for inference. Effective
In summary, the DL for unsupervised feature with reference learning processing methods are essential to enable quick issue diagnosis and
technique provides the following advantages: repair [28].
3. Integration and deployment of the model:
i. The enormous volumes of unlabeled data that online monitoring The third phase is deploying and integrating the trained model
tools acquire can be used to their fullest potential. into the operational power system. To host the model, handle
ii. The extraction approach is completely compatible with powerful streaming data, and perform inference, the necessary hardware and
ML models. software infrastructure must be set up. Strategies like containeriza­
iii. Automated feature learning with improved generalizability may tion and microservices architecture may be used for efficient
take the place of the time-consuming feature design procedure deployment and integration [29,30].
[24]. 4. Making Choices for Fault Finding and Restoration
The real-time data is analysed by the DL model, which also pro­
A lot of computational power is usually required by DL approaches duces outputs for fault detection. Based on these outputs, the power
when analyzing complex datasets or training large models. Training system may choose how to identify, find, and repair defects. In order
frequently benefits from high-performance graphics processing units to launch the appropriate power system activities, decision-making
(GPUs) or specialized hardware accelerators. Adequate computer re­ algorithms must be developed during the integration phase [28-30].
sources are also required in order to analyze the input data efficiently 5. System verification and modification
during the inference phase, where the trained model is used for real-time It is crucial to test the functioning of the integrated system and to
classification [25]. continually adapt it to changing circumstances and emerging failure
Due to the potentially large number of parameters that DL models patterns. Regular validation tests, feedback loops, and system

2
P.K. Shukla and K. Deepa Ain Shams Engineering Journal 15 (2024) 102427

Table 1 done for a 4-bus transmission line system. A thorough dataset is pro­
Types of Activation Function. duced by simulating the system and recording voltage and current data
Name Function Application for each phase. Furthermore, five different fault types are purposefully
generated and documented inside the dataset to guarantee the efficacy
Rectified Linear Units relu • Complex mapping
• Non-linear function of the fault categorization technique. Traditional fault classification
Hyperbolic tangent Tanh • Continuous function approaches based on mathematical modelling and signal processing
• Output in range of [− 1, 1] techniques may be used to properly classify these issues. The use of
• Linear and non-linear artificial neural networks (ANN) is the main emphasis of this study, and
Sigmoid or logistic Logistic Output in range of [0, 1]
a comparison of the effectiveness of Long Short-Term Memory (LSTM)
•
Identity • Differential and non-linear
Linear or identity • Output in range [− inf, inf] and LSTM with Window Regression (LSTM-WR) approaches is made.
• Faster optimization The innovative aspect of this study is the use of these algorithms for
offline fault classification tasks that will eventually be applied to real-
time power systems using the trained models. This work intends to
develop fault classification techniques in the domain of power systems
by using the advantages of DL algorithms, such as their capacity to learn
from raw data and manage nonlinear dynamics. The paper is divided as
follows – Section 2 describes the ANN and LSTM architectures, Section 3
describes the test system, Section 4 describes the dataset preprocessing,
Section 5 discusses the results in Section 4 and finally, section 5 dis­
cusses the conclusion.

2. Artificial neural network

ANN has improved the comprehension of functioning of the brain

neurons and has served as the foundation to develop artificial intelli­
gence [31]. NN is a collection of neurons that receive, process, and
Fig. 1. Artificial neural network. transport information. They employ an activation function (AF) that
changes it weighted input into its activation. Linear and nonlinear
monitoring assist the accuracy and dependability of the problem functions may be distinguished using AF [32]. These functions are
identification and restoration method [28]. chosen based on the type of issue being resolved by the network and are
tabulated in Table 1.
In this case, offline modelling using MATLAB/Simulink has been The NN for one neuron considers the weight of the input vector

Fig. 2. Long short-term memory.

3
P.K. Shukla and K. Deepa Ain Shams Engineering Journal 15 (2024) 102427

followed by summator and AF (Fig. 1). Weights for the Ith neuron as wI1, Table 2
w2I, … and wIn. They measure the robustness of the input vector I = [I1, Comparison of ANN, LSTM and LSTM-WR.
I2, … In] of the Ith neuron. Each input is multiplied by the neuron’s Technique Advantages Disadvantages
corresponding weight and then, summed as SI and given as input to AF.
LSTM [16] • Ability to capture long-term • Requires more computational
AF is a non-linear function Z(I) which is also known as transfer function. dependencies in sequential resources and training time.
The final output YI is obtained from AF. data. May overfit small datasets.
The challenge with ANN is selecting the appropriate weight values Handles variable-length Can be sensitive to
depending on the known inputs and outputs. Different numbers of hid­ input sequences. hyperparameter tuning.
Can learn temporal pat­
den layers, neurons in each layer, AF, and training techniques can be terns and dynamics
utilized depending on the complexity of the problem to be solved effectively.
[33,34]. LSTM-WR • Combines the strengths of • Window size selection can
[35] LSTM and window regression impact performance.
Considers both local and
3. Long short-term memory
global information.
Captures temporal features
The system structure, known as the Recurrent NN (RNN), is applied and their variations.
to analyze time series. Its unique feature is that the preceding output ANN [36] • Simpler architecture and • May struggle with capturing
information along with the current input information influence the easier to interpret. complex temporal
Faster training and dependencies.
present output. Although this network has certain benefits, gradient inference time. May require manual feature
disappearance or explosion is more likely to happen when the data, Suitable for smaller engineering.
needed by RNN, to learn is far from the present projected value. A datasets. Prone to overfitting if not
variation of the RNN is LSTM. Through a device known as a gate unit, it properly regularized.
is able to achieve the self-circulating weight change, successfully
resolving the gradient disappearance and gradient explosion issues. As a 4. LSTM with window regression
result, it is more suited for handling time series data. LSTM is used for
dynamically spatial forecasting problems. It learns through layers which LSTM requires the following inputs
helps to understand the pattern for better performance. It holds the in­
formation until required and then discards it when it is not required any 1. Batch size
more for prediction. There are three types of LSTM– 2. Number of steps
3. Dimensions of the input layer
1. Frontward pass
2. Rearward pass The number of steps indicate the number of LSTM cells. Consider a
3. Bidirectional random dataset with two attributes and six steps. The dataset should be
restructured such that the step is calculated on the basis of former and
The first two types are unidirectional and can process information present step attributes. For this restructuring, sliding window, also
either in forward side or backward side. The third type of LSTM is known as window regression (WR), is created based on the following
bidirectional and can process data in either direction. algorithm
The LSTM structure is shown below in Fig. 2. It has three gate con­
trollers – first gate for the input, second gate for the forget and third gate 1. Define a, b as list ()
for output. The memory cells store the state information and decides 2. Create loop for the range of the entire dataset
whether the data needs to be retained or deleted. 3. Calculate new sliding window index
Let the input data be – {A1, A2, … An}. Consider the ith sample with Xi 4. If the index > dataset size, then get sequence of the dataset for a and
as forget value, pi-1 as previous data output, o as offset term, w as weight only last element for b
and µ as sigmoid activation function, then the forget value is given as Eq. 5. Append the a and b lists
(1) 6. Make final arrays from a and b
7. Return a and b
Xi = μ(wi pi− 1 + wi Ai + oi ) (1)
The sigmoid function µ controls the data to be updated as given in Eq. The size of the batch may not even need to be supplied; it might be
(2) specified purely for the purpose of accelerating the training process.
After the implementation of WR, the LSTM model is trained and tested to
si = μ(wt pi− 1 + wt Ai + ot ) (2) check the accuracy of the proposed algorithm. A comparison of the three
The tanh function produces the participants which updates the state techniques implemented in this paper are presented in Table 2.
as given in Eq. (3) and (4)
5. Test system and data preprocessing
M‘i = tanh(wr pi− 1 + wr Ai + or ) (3)
A simulation model for categorizing and finding defects is the test
Mi = Xi *Mi− 1 + sa *M‘i (4)
system. Every electrical block has to have a parameter defined before the
where, sa is data to be updated, M‘i alternate data to be updated, Mi-1 is
old state and Mi is the updated state. The output is evaluated by the
output gate using the sigmoid function and given as Eq. (5)
Za = μ(wz pi− 1 + wz Ai + oz ) (5)
The present state output is given as Eq. (6)
pi = Za *tanh(Mi ) (6)

Fig. 3. Test system.

4
P.K. Shukla and K. Deepa Ain Shams Engineering Journal 15 (2024) 102427

Fig. 4. Dataset.

simulation model is run. The amplitudes of the current and voltage will 4. Phase information:
be recorded during the simulation. The current amplitude would be The model may learn fault-related phase patterns and distinguish
significantly increased when there is a grid failure or a fault occurs. But between various fault types by taking into account the phase angles
the voltage amplitude will decrease because of energy losses in the or phase variances between distinct phases.
transmission lines.
There are four three-phase voltage sources shown in the test system So, based on the test system, the dataset (Fig. 4) comprised of six
(Fig. 3). Each source has an 11 kV rating. Two sources are connected by input attribute – three voltages Va, Vb, and Vc, three currents Ia, Ib and Ic
transformers at the end of the transmission line. At the midway of the and four output feature – G, A, B and C representing the fault position in
transmission line, the fault is simulated. After simulating both normal the corresponding lines. The four output attributes had a Boolean output
and fault scenarios for the test system, the related data was logged and – 0 and 1, for normal and fault conditions respectively. These values
sent as input to the DL algorithms. determined the label of the fault, given as FaultType.
The specific objectives, the available data, and the peculiarities of The following methods can be used to categorize short circuit
the transmission line fault classification problem all influence the choice faults according to the fault current and voltage:
and combination of input characteristics. When selecting input features,
it is important to take into account the underlying physics of the system, 1. Magnitude analysis:
the fault types of interest, and the features’ ability to accurately capture Takes into account the fault voltage and current’s size.
fault-related patterns. The classification of transmission line defects re­ 2. Phase analysis:
lies crucially on feature engineering and domain expertise to provide Takes into account the fault voltage and current’s phase angle or
meaningful and instructive input characteristics for DL models. phase discrepancy.
3. Symmetrical analysis:
1. Current and voltage: Examines the fault current and voltage signals’ positive, negative,
The electrical signals recorded at various points along the trans­ and zero sequence components.
mission line are represented by these waveforms. 4. Transient analysis:
2. Frequency components: Takes into account the fault current and voltage’s transient
The signals may be converted into the frequency domain using the response.
Wavelet or Fourier transforms, enabling the model to detect fault- or 5. Fault duration analysis:
disturbance-related patterns at certain frequencies. Examines the fault duration, including the fault current and
3. Time-frequency fluctuations: voltage signals’ rise and decay times.
These can record harmonic components, transient occurrences, or
frequency changes associated with fault situations. However, because the dataset is labelled, the DL algorithms can learn

Fig. 5. Boxplot for outlier detection (a) Ia (b) Ib (c) Ic (d)Ia (e) Ib (f) Ic.

5
P.K. Shukla and K. Deepa Ain Shams Engineering Journal 15 (2024) 102427

Fig. 6. Faults occurring in each phase.

Table 3 6. Implementation and results

Fault description.
Fault Type Fault Type Interpretation Number of Fault Occurrences The raw dataset is preprocessed and labeled as per the requirements.
A dataset correlation is performed for the dataset before implementing
0000 No fault 2365
1011 LLG fault 1134 ANN, LSTM and LSTM-WR. A correlation is a statistical association be­
1111 LLLG fault 1133 tween two features or variables. Positive correlation implies that
1001 LG fault 1129 together the features change identically i.e., an increase in the first
0111 LLL fault 1096 feature results in an increase in the second feature. A negative correla­
0110 LL fault 1004
tion infers that the features move in the opposite direction i.e., an in­
crease in the first feature results in a decrease in the second feature.
Fig. 8 shows the correlation between the various features.
The following hyperparameters can be optimised:

1. Learning rate:
The learning rate controls the step size at each iteration
throughout the model training process. It determines how frequently
the model changes its internal parameters based on the estimated
gradients.
2. Number of hidden layers and units:
By modifying the network’s depth and breadth, it is possible to
optimise the architecture and strike the perfect balance between
model complexity and accuracy.
3. Activation function:
Relu, sigmoid, or tanh are examples of popular activation func­
tions that may be optimised to find the best combination for a certain
fault classification task.
4. Dropout rate:
This regularisation technique randomly eliminates a fraction of the
model’s units during training in order to prevent overfitting.
5. Batch size:
Fig. 7. Fault occurrences. Depending on the batch size, different samples are treated in
different training iterations. Batch size optimisation may affect the
without utilizing these methods. model’s generalisation performance, memory requirements, and
Label encoding is a critical preprocessing stage for altering the convergence speed.
object-type fields to numerical-type. The library package in Python 3.0 6. Regularisation methods:
called sklearn.preprocessing was used for label encoding. L1 or L2 regularisation, dropout, or batch normalisation are reg­
From the dataset exploration, it was found that the dataset was ularisation methods that can help prevent overfitting and improve
almost normally distributed and without outliers as shown in boxplots of the generalisation abilities of the model.
the voltages and currents (Fig. 5 (a–f)).
The dataset included six output labels namely – no fault, LG, LL, DLG, There is no established process for choosing hyperparameters; they
LLL and TLG faults. The fault occurrences on each line i.e., G, A, B and C are chosen at random. To find the best configuration for the particular
is shown in Fig. 6. It is observed that all the lines have approximately job and dataset at hand, it is necessary to conduct several trials with
similar number of fault occurrences. various hyperparameter settings and assess their effects on metrics like
The different faults are labelled as shown in Table 3. The number of accuracy, loss, or validation error.
fault occurrences is shown in Fig. 7. It is observed that no fault cases Relu and tanh are two different kinds of activation functions that
have a majority and rest of the fault cases have almost equal occurrence. were used in this study to create the artificial neural network (ANN) for
fault classification. The treatment of non-linear datasets is suitable for

6
P.K. Shukla and K. Deepa Ain Shams Engineering Journal 15 (2024) 102427

Fig. 8. Dataset feature correlation.

both activation functions. The uniform kernel was chosen since it was configuration changes, according to this.
discovered that the distribution of the dataset was approximately nor­ The number of units in each layer did, however, have a considerable
mally distributed. This decision shows that no specialised kernel was impact on the algorithm’s accuracy. The best training and testing ac­
necessary for the dataset to perform better. curacy, respectively, were obtained by increasing the number of units in
Different configurations were examined during the experimental the first layer while maintaining the number of units in the second layer
process to see how they affected the algorithm’s accuracy (Table 4). It’s fixed at either 10 or 100. On the other hand, regardless of the number of
interesting that the algorithm’s accuracy did not noticeably change units in the second layer, the lowest training accuracy (30.15%) and
when the setup was changed, such as by changing the kernel or acti­ testing accuracy (29.81%) were seen when the first layer contained just
vation functions. The method may be somewhat resilient to one unit.

7
P.K. Shukla and K. Deepa Ain Shams Engineering Journal 15 (2024) 102427

Table 4
ANN Hyperparameters.
Input layer Hidden layer 1 Hidden layer 2 Output layer Accuracy
(%)
Kernel AF Kernel AF No. of units Kernel AF No. of units Kernel AF

Uniform Relu Uniform Relu 1 Uniform Relu 10 Uniform Relu 29.81

10 10 42.33
100 10 42.33
100 100 29.81
Uniform tanh Uniform tanh 1 Uniform tanh 10 Uniform tanh 29.81
10 10 42.33
100 10 42.33
100 100 29.81

Table 5 problems can result from a number of issues with the transmission and
Accuracy comparison. distribution of electricity, including losses, corona discharge, and other
Algorithm Train Accuracy (%) Test Accuracy (%) factors. These flaws not only stop the steady flow of energy to customers,
but they also cause maintenance problems, user annoyance, and finan­
ANN 42.99 % 42.33 %
LSTM 99.98 % 100 % cial losses for enterprises. It is essential to fix these issues as soon as
LSTM-WR 99.99 % 99.98 % possible because they might result in protracted disruptions that impede
corporate growth and interfere with routine duties that depend on a
These results underline how crucial it is to choose the right number of reliable power supply.
units for each layer in order to maximize the effectiveness of the ANN for This study focuses on the offline implementation of fault classifica­
fault classification. It is clear that the accuracy of the method is signifi­ tion methods using MATLAB/Simulink in a 4-bus transmission line
cantly influenced by the number of units in the first layer, although the system. In order to provide a diversified dataset for fault classification
activation functions and the selection of the uniform kernel do not analysis, the modelling procedure entails the thorough recording of
significantly affect accuracy. This information offers insightful strategies voltage and current data for each phase while purposefully introducing
for accurately setting the ANN and producing fault classification findings. multiple simulated fault types. The originality of this research is in using
The classification accuracy comparison of the methods used in this LSTM and LSTM-WR approaches, in particular, to accurately classify
work is shown in Table 5. Lowest training and testing accuracies for the faults in power systems.
ANN are 42.99% and 42.33%, respectively. The finest training and This work’s offline nature enables a thorough investigation of
testing accuracies, respectively 99.98% and 100% and 99.99% and various problem scenarios and the enhancement of fault classification
99.98%, were demonstrated by LSTM and LSTM-WR. accuracy. Comparing ANN, LSTM, and LSTM-WR approaches can pro­
The Fig. 9 shows the training and testing plots for LSTM and LSTM- vide important insights into their effectiveness and applicability for real-
WR. The plots are over-lapping because of high level of accuracies. time fault classification in power systems. By removing the requirement
Table 6 contrasts the fault classification algorithms utilized in this for manual feature extraction, ANN models have advantages that lead to
study with the accuracy of several approaches [33,34]. The simulation more effective fault categorization. Particularly the LSTM and LSTM-WR
findings presented in this research demonstrate that training and testing designs may successfully learn patterns and correlations among the fault
accuracy is close to 100%. However, the classification accuracy of CDASE, data because they can capture long-term dependencies and temporal
DWT and SVM, and hybrid DWT and ANN approaches was 92.22 percent, dynamics. Furthermore, they are appropriate for capturing the complex
87.67 percent, and 91.60 percent, respectively [33]. However, a very basic behaviours prevalent in power system breakdowns due to their capacity
power system was taken into consideration for the study and noise signal
may be included to assess the consistency of the system [34]. DNN also Table 6
demonstrated 100% correctness. However, this research suggested the Comparison of accuracies.
LSTM and LSTM-WR for a more complicated power system and demon­ Algorithm Classification Accuracy (%)
strated accuracy with both algorithms of about 100%.
ANN 42.33
LSTM 100
7. Conclusion LSTM-WR 99.98
DNN [18] 100
CDSAE [17] 92.22
The proper operation of contemporary enterprises and the growth of
DWT + SVM [17] 87.67
our civilization depend on a steady supply of power electricity system DWT + ANN [17] 91.60

Fig. 9. Training and testing plots for LSTM and LSTM-WR.

8
P.K. Shukla and K. Deepa Ain Shams Engineering Journal 15 (2024) 102427

to handle nonlinear and non-stationary signals. [16] Zhang S, et al. Deep learning for sensor-based fault diagnosis in complex
mechanical systems. IEEE Trans Ind Inf 2019;15(5).
The impact of changing the activation function was examined in the
[17] Zhang J, He ZY, Lin S, Zhang YB, Qian QQ. An ANFIS-based fault classification
context of fault detection using ANN, and it was discovered that the approach in power distribution system. Int J Electr Power Energy Syst 2013;49:
accuracy was not considerably impacted. However, the accuracy of the 243–52.
method is affected by modifying the hidden layers’ hyperparameters. [18] Chen K, Huang C, He J. Fault detection, classification and location for transmission
lines and distribution systems: a review on the methods. High Volt 2016;1(1):
ANN training and testing accuracies were found at 42.98% and 42.33%, 25–33.
respectively. The test system and its associated dataset were used for [19] AU. of E, Shihab T, MRU of E, Islam T. Artificial neural network based fault
LSTM and LSTM-WR implementation. The highest training and testing classification and location for transmission lines. In: IEEE conf power electron
renew energy, CPERE 2019; 2019. p. 140–4.
accuracies for LSTM were 99.98% and 100%, while for LSTM-WR they [20] Zhang C, et al. Deep learning for fault detection and diagnosis in HVAC systems.
were 99.99% and 99.98%. However, it should be highlighted that these Energ Build 2019;202.
algorithms’ extensive testing and training procedures, which are un­ [21] Zhang Y, et al. Deep learning-based fault diagnosis for rolling bearings using time-
frequency image analysis of vibration signals. Sensors 2018;18(9).
predictable, are a negative. [22] Sun Z, et al. Deep learning-based fault diagnosis for wind turbines under different
In conclusion, this study focuses on the offline implementation of operation conditions. Energies 2018;11(10).
fault classification algorithms in a 4-bus transmission line system [23] Xu X, et al. Deep learning-based fault diagnosis for photovoltaic systems. IEEE
Trans Ind Electron 2019;67(5).
employing ANN, LSTM, and LSTM-WR. The goal is to maximise fault [24] Rosle N, et al. Fault detection and classification in three phase series compensated
classification accuracy by utilising the advantages of these methodolo­ transmission line using ANN. J Phys Conf Ser 2020;1432(1).
gies and investigating diverse problem scenarios. The results show that [25] Yuan Y, et al. Power system fault classification based on deep learning: a review.
IEEE Access 2019;7.
LSTM and LSTM-WR have the capacity to accurately classify faults, but
[26] Fu W, et al. A review on deep learning techniques for power system disturbance
their testing and training are time-consuming. monitoring and protection. Electr Pow Syst Res 2020;176.
[27] Krizhevsky A. Learning multiple layers of features from tiny images. Master’s
Declaration of Competing Interest thesis, University of Toronto; 2009.
[28] Wang N, et al. Deep learning based fault location for power distribution systems
with limited measurements. IEEE Trans Power Syst 2019;34(2).
The authors declare that they have no known competing financial [29] Lin Z, et al. Application of deep learning in real-time overhead line fault diagnosis.
interests or personal relationships that could have appeared to influence IET Gener Transm Distrib 2020;14(14).
[30] Chen X, et al. Fault detection in smart grids using deep learning and hierarchical
the work reported in this paper. mixture of experts. IEEE Trans Smart Grid 2020;11(6).
[31] Seema S, Mamatha KR, Thejaswini S. Intelligent fault identification system for
References transmission lines using artificial neural network. IOSR J Comput Eng (IOSR-JCE)
2014.
[32] Tatjana K, Goran Š. “Odlučivanje i optimizacija”, Repro Karić d.o.o., Tuzla, 2010
[1] Chaitanya BK, Yadav A. An intelligent fault detection and classification scheme for
(PDF) Power System Fault Detection, Classification and Location using Artificial
distribution lines integrated with distributed generators. Comput Electr Eng 2018;
Neural Networks. <https://www.researchgate.net/publication/317428704_
69:28–40.
Power_System_Fault_Detection_Classification_and_Location_using_Artificial_Neural_
[2] Manohar M, Koley E. SVM based protection scheme for microgrid. In: 2017 int conf
Networks> [accessed Jan 27 2023].
intelligent computing instrumentation and control technologies ICICICT 2017;
[33] Heo S, Lee JH, Heo S, Lee JH, Lee JH. ScienceDirect. 2405-8963 © 2018, IFAC
2018.
(International Fed Autom Control Hosting by Elsevier Ltd All rights Reserv Peer
[3] Tokel HA, Al HR, Alirezaei G, Mathar R. A new approach for machine learning-
Rev under Responsib Int Fed Autom Control, 2018, doi: 10.1016/j.
based fault detection and classification in power systems. In: 2018 IEEE power
ifacol.2018.09.380.
energy society of innovative smart grid technology conference ISGT 2018; 2018.
[34] Nasrin MAM, Omar AMS, Ramli SSM, Ahmad AR, Jamaludin NF, Osman MK. Deep
[4] Abdelgayed TS, Morsi WG, Sidhu TS. A new approach for fault classification in
learning approach for transmission line fault classification. IEEE Conf Control Syst
microgrids using optimal wavelet functions matching pursuit. IEEE Trans Smart
Comput Eng 2021.
Grid 2018;9(5):4838–46.
[35] Yu L, et al. Deep learning-based fault diagnosis for bearings using sensors. Sensors
[5] Ortega MJ, Hernández JC, García OG. Measurement and assessment of power
2017;17(3).
quality characteristics for photovoltaic systems: harmonics, flicker, unbalance, and
[36] Xu D, et al. Deep learning for fault diagnosis using motor current signals. IEEE
slow voltage variations. Electrical Power Syst Res 2013.
Trans Ind Electron 2017;64(6).
[6] Tejesh B, Udaya BG, Krishna BN, Deepa K, Manitha PV, Sailja V. A comparative
analysis of power generation using solar and fuel cell for charging EVs. In:
International virtual conference on power engineering computing and control:
developments in electric vehicles and energy sector for sustainable future; 2022. Dr. K. Deepa graduated from Alagappa Chettiar college of en­
[7] S. Bhargavi, U.B. Jyothsana, K. Deepa, J. Ramprabhakar, Smart V2G/G2V charging gineering and Technology, T.N, India in 1998. She obtained M.
for grid connected electric vehicles in Indian scenario, International Conference on Tech degree from Anna University, Guindy campus, T.N, India
Electronic Systems and Intelligent Computing, (2022), 40-45. in 2005. She received Doctoral degree from Jawaharlal Nehru
[8] Sanchez-Sutil F, Hernández JC, Tobajas C. Overview of electrical protection Technological University, Anantapur, A.P, India in 2016.
requirements for integration of a smart DC node with bidirectional electric vehicle Dr.K.Deepa, is Associate Professor in EEE Dept and Lead, E-
charging stations into existing AC and DC railway grids. Electrical Power Syst Res Mobility Thrust Area Group, Amrita School of Engineering,
2015;122:104–18. Amrita Vishwa Vidyapeetham, Bangalore. She mainly works in
[9] Ardakanian O, Yuan Y, Dobbe R, Von Meier A, Low S, Tomlin C. Event detection the field of power electronics with focus on renewable energy
and localization in distribution grids with phasor measurement units. IEEE Power source-based systems, micro -pico-nano-grids, and electric ve­
Energy Soc Gen Meeting 2018. hicles. 20+PG graduates and two PhD scholar have completed
[10] Kumar A, Vidya HA. Application of k-Nearest Neighbor (kNN) machine algorithm their degree under her guidance, and currently she is guiding 7
for transformer fault classification. Int J Adv Sci Technol 2020;29:8441–8. PhD scholars. She has presented and published several research
[11] Vishakh KH, Abhishek O, Airavati S, Sunil J, Pratheek PR, Nair MG. Fault detection papers in reputed journals and conferences. She had 60+ International Journals and
isolation restoration of an active distribution network. In: Innovations in power 80+International conference papers to her credits. She has authored 2 books titled
and advanced computing techniques; 2019. 8960245. “Control systems” and “Electrical Machines”, 5 book chapters and has filed one Indian
[12] Nookesh VM, Kumar VB, Saketh BS, SS, Ramprabhakar J. Comparative study of patent. She has 23+ years of academic and research experience. She is Life Member of
wind speed forecasting using machine learning. In: 2021 2nd international Indian Society For Technical Education (ISTE - India), Associate Member of The Institution
conference on smart electronics and communication (ICOSEC), India; 2021. p. of Electronics and Telecommunication Engineers (IETE-India) and Senior Member of The
1121–6. Institution of Electrical and Electronics Engineers (IEEE –USA). She is currently the chair of
[13] Vanitha PPPV, RR. Wind speed forecasting using long short term memory IEEE-PELS Bangalore chapter and also Secretary for 2021 and 2022 IEEE PELS Bangalore
networks. In: 2019 2nd international conference on intelligent computing, chapter. She has received the outstanding Exofficio - 2021 award for her incredible vol­
instrumentation and control technologies (ICICICT), India; 2019. p. 1310–4. unteering in IEEE PELS Bangalore section. She is the advisor for the IEEE-PES & IAS stu­
[14] Xu D, et al. Deep learning for fault diagnosis using convolutional neural networks. dent branch joint chapter and counsellor for IEEE Amrita School of Engineering student
Sensors 2018;18(8). branch. She was the advisor for IEEE WIE in Amrita School of Engineering, Bengaluru from
[15] Liang H, et al. Deep recurrent neural networks for machinery fault diagnosis. Shock 2015 to 2020.
Vib 2015;2015.