IEEE - 61001

Fault Classification & Detection in IEEE 34 Bus

System Using Convolutional Neural Network
2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT) | 979-8-3503-7024-9/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICCCNT61001.2024.10725909

[Link] Kumar [Link] Jeyanthy [Link] K

Assistant Professor/ Dept of EEE Professor/Dept of EEE Professsor/Dept of EEE
Sir M Visvesvaraya Institute of Kalasalingam Academy of Research Sir M Visvesvaraya Institute of
Technology,Bengaluru-562157 and Education, Krishnankoil,India Technology,Bengaluru-562157
rajeshkumar_eee@[Link] [Link]@[Link] drmaheshk_eee@[Link]

Akila.P Thrisha V S Swetha B M

Assistant Professor/Dept of ETE Assistant Professor/Dept of CSE Assistant Professor/Dept of ETE
Sir M Visvesvaraya Institute of Sir M Visvesvaraya Institute of Sir M Visvesvaraya Institute of
Technology,Bengaluru-562157 Technology,Bengaluru-562157 Technology,Bengaluru-562157
akila_te@[Link] thrisha_cs@[Link] swethabm_ete@[Link]

Abstract- This paper presents a novel approach by means of improving fault detection and classification accuracy in
Convolutional Neural Networks (CNNs) for fault classification distribution schemes, which is crucial for maintaining system
and detection in the IEEE 34 bus system, which serves as a reliability and stability. This may have limitations in terms of
representative model for distribution networks. The IEEE 34 scalability and applicability to different types of distribution
bus system is known for its complexity, featuring multiple buses,
systems with varying levels of distributed generation
branches, and loads, making it an ideal testbed for evaluating
advanced fault management techniques. The proposed CNN- integration. [2] conducted an examination on fault location
based methodology leverages the ability of CNNs to and detection procedures in power distribution schemes. The
automatically learn hierarchical features from system data, review highlighted various techniques and methods used for
such as voltage and current measurements, without the need for fault location and detection in distribution systems with
manual feature extraction. The key contributions of this paper distributed generation, emphasizing the importance of
include the development and implementation of a CNN accurate fault detection for system protection and reliability.
architecture tailored for fault classification and detection tasks The review may have limitations in terms of the
in the IEEE 34 bus system. The CNN model is trained using comprehensiveness of the reviewed techniques and the
labeled data sets containing information about different fault
applicability of the findings to specific distribution system
scenarios, including 1-phase, 2-phase, and 3-phase faults. The
training process involves optimizing the network parameters to configurations.[3] discussed fault detection and protection
improve accuracy and robustness in fault identification. schemes for distributed generation integrated into
Simulation results demonstrate the effectiveness of the CNN- distribution networks. The study identified challenges and
based approach in accurately classifying and detecting faults in provided suggestions for improving fault detection and
the IEEE 34 bus system. The CNN model shows high accuracy protection schemes in distribution systems with distributed
rates in distinguishing between different fault types and generation, emphasizing the need for robust and reliable
provides rapid response times for fault detection, thereby protection mechanisms. Deep learning approaches for fault
enhancing the overall reliability and resilience of the recognition and category in the electrical secondary
distribution network. The implications of this research extend
distribution network were compared by [4]. The study
to the broader context of intelligent fault management systems
in power engineering. By harnessing the capabilities of CNNs, compared different methods and evaluated the accuracy of
power system operators can benefit from more efficient and recurrent neural networks for fault detection and
automated fault management processes, leading to improved classification, highlighting the potential of deep learning
system performance, reduced downtime, and enhanced grid techniques in improving fault detection performance. The
reliability. comparison conducted by [4] may have limitations in terms
Keywords— Fault Classification, Fault Detection, IEEE 34 Bus of the generalizability of the results to diverse distribution
System, Convolutional Neural Networks (CNNs), Distribution network structures and fault circumstances. [5] proposed an
Networks active recognition fault identification and fault locality
I. INTRODUCTION expertise for LVDC distribution networks. The investigation
determined to develop a technology for fault identification
Fault classification and detection are critical aspects and locality in low-voltage direct current (LVDC)
of ensuring the reliability and stability of electrical power distribution networks, aiming to improve system reliability
schemes. The IEEE 34 bus system serves as a benchmark and efficiency. The technology may have limitations in terms
system widely used in power system studies and research. of the practical implementation challenges and the scalability
Leveraging advanced technologies like Convolutional Neural of the technology to larger LVDC distribution networks.
Networks (CNNs) for fault classification and detection in the Enhancing the accuracy of fault classification and
IEEE 34 bus system represents a significant step towards position in an intelligent distribution system using stacked
enhancing the efficiency and accuracy of fault management ensemble learning was focused by [6]. The main finding of
in power systems. the study was that the stacked ensemble learning approach
An efficient methodology for the detection and organization improved the accuracy of fault category and position in the
of SC faults in distribution techniques with distributed distribution system. The study did not consider the impact of
generation was proposed by [1]. The study focused on varying fault types or network configurations on the

15th ICCCNT IEEE Conference,

Authorized licensed use limited to: VIT University- Chennai [Link] 24-28, 2024,
Downloaded on May 22,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
IIT - Mandi, Kamand, India
 IEEE - 61001

performance of the stacked ensemble learning approach.[7] II. SYSTEM DESCRIPTION

investigated fault position in distribution arrangements with
A. Fault Classification
distributed production elements grounded on the impedance
environment. The study found that using the impedance Various types of faults commonly encountered in electrical
matrix can help improve fault position accuracy in systems:
distribution systems. It did not explore the impact of different Short-Circuit Faults:
fault scenarios or network topologies on the effectiveness of When two phases of a 3-phase system come into direct
using the impedance matrix for fault location. Fault detection, contact with one another, they bypass the load and may cause
category, and position in microgrids using superimposed a phase-to-phase short circuit, which can result in excessive
elements and Wigner distribution utility was studied by [8]. current flow and equipment damage.
The main finding of the research was that the proposed Open-Circuit Faults:
approach using superimposed elements and Wigner supply An open circuit fault happens once there is a breakdown or
utility can effectively identify and localize faults in interruption in one of the conductors of a circuit, which can
microgrids. It did not address the computational complexity result from physical damage or poor connections.
or real-time implementation challenges of the proposed fault Specifically, in a phase open circuit fault, one phase of a
identification and localization method.[9] investigated the multi-phase system becomes disconnected, causing an
identification of fault types in distribution systems utilizing a imbalance that can impact the performance of connected
trivial ANN with distributed production. The study found that loads.
the trivial ANN approach can accurately identify the variety Ground Faults:
of faults in distribution systems with distributed production. A single line-to-ground fault happens when one phase of the
The impact of varying fault conditions or network parameters system contacts the ground, typically due to insulation
on the performance of the shallow neural network for fault degradation or equipment malfunction. On the other hand, a
type identification was considered. An assessment of fault double line-to-ground fault involves two phases
position and categorization methods in distribution grids was simultaneously shorting to ground, often resulting from
conducted by [10]. The review highlighted various fault insulation breakdown or severe system disturbances.
position and category methods used in allocation grids and Transient Faults:
discussed their advantages and limitations. The review did Voltage sag/dip refers to a temporary decrease in voltage
not provide a comprehensive comparison of different fault levels, usually caused by sudden load changes or short-
locations and classification methods or evaluate their duration faults. Conversely, voltage swell/surge represents a
performance in different network scenarios. temporary increase in voltage levels, often triggered by
Convolutional Neural Networks (CNNs) have switching operations or lightning strikes.
revolutionized numerous topics, incorporating image Intermittent Faults:
detection, language and signal processing. Their capability to Understanding fault types such as intermittent open, where
inevitably study classified features from facts makes them circuits sporadically open and close due to factors like
well-suited for analyzing complex patterns and structures, thermal expansion or loose connections, and intermittent
making them increasingly popular for applications in power short, causing intermittent short circuits from insulation
systems. The degradation or mechanical stress, is crucial for effective fault
IEEE 34 bus system is a realistic representation of a analysis, diagnosis, and mitigation strategies in electrical
distribution network, featuring multiple buses, branches, and systems. This understanding ensures system reliability and
loads. It is used extensively in research and testing due to its safety.
complexity and relevance to real-world distribution systems.
Faults in the IEEE 34 bus system can arise from various
III. FAULT CLASSIFIER
sources, including short circuits, equipment malfunctions,
and external disturbances. By applying CNNs to fault A. Wavelet Transformation
classification and detection in the IEEE 34 bus system,
researchers and engineers can benefit from improved Analyzing digital signals across various domains helps
accuracy, faster response times, and reduced manual understand their key characteristics like periodicity,
intervention in fault management processes. CNNs can autocorrelation, and power spectrum, which are challenging
analyze system data, such as voltage and current in a single domain. Wavelets are effective in signal analysis
measurements, and automatically identify and classify and are widely used in computer vision, music, and digital
different faults, including 1-phase, 2-phase, and 3-phase processing. They are small oscillating wave functions with
faults. The utilization of CNNs in fault classification and mother wavelets generating similar translated or dilated
detection for the IEEE 34 bus system represents a cutting- waves, forming wavelet coefficients. FFT and wavelet
edge approach that aligns with the ongoing digitalization and transform are similar yet different in representing signals.
automation trends in the power industry. This approach not Wavelets in DWTs extract features for fault detection in
only improves the dependability and resilience of distribution microgrids. For a constant-time signal x(t), the wavelet alter
networks but also lays the foundation for future 𝑊𝑥 (𝑎, 𝑏) is defined as
advancements in intelligent fault management systems. 1 ∞ 𝑡−𝑏
𝑊𝑥 (𝑎, 𝑏) = ∫ 𝑥(𝑡) ( 𝑎 ) 𝑑𝑡 (1)
√|𝑎| −∞

Finding the inner product among x(t) and a scaled and

translated representation of a single function ψ(t) is the first

15th ICCCNT IEEE Conference,

Authorized licensed use limited to: VIT University- Chennai [Link] 24-28, 2024,
Downloaded on May 22,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
IIT - Mandi, Kamand, India
 IEEE - 61001

step in computing the wavelet transform. By changing the This is particularly useful for visual datasets with strong
scaling parameter a, the wavelet's bandwidth and center patterns in the data.
frequency may be changed. The transformation becomes a CNN architecture
convolution of x(t) with the time-reversed and scaled Multiple layers are incorporated in Convolutional Neural
wavelets when an is fixed. Additionally, the wavelet Networks (CNNs), including the input, Convolutional,
transform must meet the admissibility requirement for the Pooling, and fully linked layers. The input image undergoes
wavelet ψ(t) to guarantee an impeccable reconstruction of the filtering in the Convolutional layer to extract features,
signal while changing it into other representation domains. followed by downsampling in the Pooling layer to decrease
computational load, and the final prediction is made in the
∞ |𝜓(𝜔) 2 |
𝐶𝜓 = ∫−∞ 𝑑𝜔 < ∞ (2) fully connected layer. The network's learning of optimal
|𝜔|
filters is achieved through backpropagation and gradient
descent.
Where 𝜓(𝜔) is the transformation ability of the wavelet,
∞
𝜓(0) = ∫−∞ 𝜓(𝑡)𝑑𝑡 = 0 (3)

To address challenges related to plotting a wave in discrete

time fields, the introduction of the Discrete Wavelet
Transform (DWT) is considered. The signal is scaled and Fig.1 Representation of Convolution Neural Network
translated in discrete steps by the DWT, retaining ample
information for signal analysis and synthesis. Moreover, it
reduces computational time and is relatively straightforward 1. Convolutional Layer:
to implement. The feature extraction in a CNN is handled by the
convolutional layer, which employs filters (kernels) on the
1 ∞ 𝑡−𝑘𝑏0 𝑎0
𝑗 input image, generating feature maps. Mathematically, the
𝑑𝑗,𝑘 =
𝑗
∫−∞ 𝑥(𝑡)𝜓 ∗ ( 𝑗 ) 𝑑𝑡, (4) convolution operation is represented as:
√|𝑎0 | 𝑎0

𝑍[𝑖, 𝑗] = ∑𝑀−1 𝑁−1

𝑚=0 ∑𝑛=0 𝑋[𝑖 + 𝑚, 𝑗 + 𝑛]. 𝑊[𝑚, 𝑛] + 𝑏 (6)
The DWT evaluates the signal by decomposing it into
approximations and detail parameters, where j and k are
Where 𝑍[𝑖, 𝑗] is the output value at location (i, j) in the
integer representations of the level and location, respectively,
and each has a variable resolution across various frequency attribute map, 𝑋[𝑖 + 𝑚, 𝑗 + 𝑛]is the input value at location
bands. The time-domain signal is successively filtered using 𝑖 + 𝑚, 𝑗 + 𝑛 , 𝑊[𝑚, 𝑛] +is the weight of filter at location
low-pass and high-pass filters throughout this decomposition [𝑚, 𝑛], 𝑏 is the bias label, and M and N are the dimensions
process, usually with the aid of wavelet and scaling value.
algorithms. First, half of the samples in the original signal 2. Activation Function:
x(n) are removed by filtering it using a half-band low-pass To add non-linearity to the network after convolution, a
filter and a half-band high-pass filter h(n). This produces a function called activation is utilized element-by-element.
halved time resolution and a doubled frequency resolution. The sigmoid and ReLU (Rectified Linear Unit) are two
The detail coefficients are produced by the low-pass filter, common activation functions. The definition of the function
while the approximation coefficients are produced by the that activates ReLU is:
high-pass filter. At each stage of decomposition, the sub band
coding process doubles the frequency resolution and 𝑓(𝑥) = max⁡(0, 𝑥) (7)
decreases temporal resolution until there are only two
samples left in the signal. This process is repeated recursively 3. Pooling Layer:
with the detail coefficients. Attaining multiresolution The longitudinal dimensions of the characteristic maps are
breakdown at level M may be explained mathematically as minimized by the pooling layer, which preserves critical
follows: information. Max pooling is a widely applied procedure for
pooling.
1 𝑡 𝑡 𝐾−1
𝑌[𝑖, 𝑗] = 𝑚𝑎𝑥𝑚=0 𝐿−1
𝑚𝑎𝑥𝑛=0 ⁡𝑋[𝑖. 𝐾 + 𝑚, 𝑗. 𝐿 + 𝑛] (8)
𝑥(𝑡) = ∑𝑘 𝑎𝑀,𝑘
√ 2𝑀 𝑗 ∑𝑘 𝑑𝑗,𝑘 𝜓 (2𝑗 − 𝑘) (5)
𝜑 (2𝑀 − 𝑘) + ∑𝑀
Where 𝑌[𝑖, 𝑗] is the output of the pooling operation at
The approximation coefficients at level M, denoted as 𝑎𝑀,𝑘 , position (i, j), 𝑋[𝑖. 𝐾 + 𝑚, 𝑗. 𝐿 + 𝑛] is the input position, and
and the scaling function φ(t) are involved in the wavelet K and L are the pooling window dimensions.
transform. In each decomposition level, the DWT signal 4. Fully linked Layer:
highlights high-amplitude frequencies corresponding to the The flattened outcome from the previous levels is taken
most significant frequencies in the original signal's respective by the fully connected layer, which then executes
region. classification or regression tasks. The calculation of the
output in the fully linked layer is:
B. Basics of Convolution Neural Network 𝑍 = 𝑊. 𝑋 + 𝑏 (9)
In a CNN, elements such as the weight matrix W, input
A modified version of ANN, the CNN is mostly used to vector X, bias vector b, and output vector Z are used to learn
extract features from matrix datasets with grid-like structure. hierarchical features from input data. These components

15th ICCCNT IEEE Conference,

Authorized licensed use limited to: VIT University- Chennai [Link] 24-28, 2024,
Downloaded on May 22,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
IIT - Mandi, Kamand, India
 IEEE - 61001

work together to make activities possible in power systems, efficient fault management and maintenance of electrical
such as defect detection & identification. The effectiveness power systems.
of CNNs for fault classification relies heavily on the size,
diversity, and preparation of data sets. Large data sets provide IV. SIMULATION RESULTS & DISCUSSIONS
numerous examples, helping the model learn diverse patterns The fault analysis study was simulated in
and reducing overfitting. Diversity ensures the model can MATLAB/Simulink using the IEEE-34 bus structure to
handle various fault scenarios accurately. Proper preparation evaluate the proposed technique in three distinct scenarios.
includes meticulous labeling, data normalization, noise These scenarios consider variations in fault type, location,
removal, and class balancing, ensuring high-quality input for and the presence of DG units. In the first scenario, only DG
effective learning and robust performance. units are modeled. The second scenario includes a wind
generator, while the third scenario incorporates both a PV and
wind generator for electricity control. During the simulation,
C. CNN for Fault Detection & Classification
four metrics were used: QNN (true-positive instances of
faultless normal cases), QNF (false-negative cases of normal
To use Convolutional Neural Networks (CNNs) for fault instances identified as faults), QFN (false-positive cases of
recognition and category in the IEEE 34 bus system, the faults identified as normal), and QFF (true-negative cases of
procedure typically involves several steps: faults correctly identified).
1. Data Collection and Preprocessing:
Collect data from the IEEE 34 bus system, including voltage Table.1 scenario's number of simulations and fault type
and current measurements during normal operation and fault
scenarios. Preprocess the data to remove noise, normalize the Type of Low/High Total No of Simulations
values, and prepare it for input into the CNN model. Fault Impedanc Scenari Scenari Scenari
2. Dataset Preparation: e Fault o1 o2 o3
Divide the pre-processed data into training, validation, and Rg, Yg, Bg Low 830 830 595
testing [Link] the data according to fault types, such
High 1580 1630 1580
as single-phase, double-phase, three-phase faults, etc.
RY, RYg Low 540 540 580
3. CNN Design:
High 520 520 600
Create a CNN architecture that works well for jobs including
categorization and fault identification. Typically, the RB,RBg Low 540 540 580
architecture consists of fully linked layers for classification, High 520 520 520
pooling layers for reducing dimensionality, and YB,YBg Low 540 540 580
convolutional layers for feature extraction. High 520 520 520
4. Training of Model: RYB,RY Low 540 540 540
Utilizing the labeled training dataset, train the CNN [Link] Bg High 520 520 520
minimize the loss function during training, use optimization Total 6650 6700 6615
methods like Adam optimizer or stochastic gradient descent
.To avoid overfitting, keep an eye on the model's performance The table provides a breakdown of the number of simulations
using the validation dataset. conducted for different fault scenarios and fault types in a
5. Model Evaluation: power system. The fault types are categorized based on their
Utilizing the testing dataset, evaluate the trained CNN impedance levels, either low or high impedance faults. Each
model's performance in defect detection and classification. scenario, labeled as Scenario 1, Scenario 2, and Scenario 3,
Assess the model's efficacy using measures like precision, involves varying conditions or configurations in the system.
recall, and the F1-score The fault types include faults between ground (Rg, Yg, Bg),
6. Fault Detection and Classification: between phases (RY, RYg, RB, RBg, YB, YBg), and
Input voltage and current data from the IEEE 34 bus system combinations of phases (RYB, RYBg). For each fault type
into the trained CNN model. The CNN model will and impedance level, the table lists the number of replications
automatically extract features from the input data and classify executed under each scenario. The total number of
the fault type based on learned patterns. replications for each scenario is also provided, summing up
7. Validation and Optimization: the simulations across all fault types and impedance levels.
Validate the CNN model's performance in real-world This data is essential for understanding the distribution and
scenarios by testing it with new fault data from the IEEE 34 frequency of fault simulations across different scenarios and
bus system. Fine-tune the model parameters and architecture fault types, which is crucial for analyzing fault behavior and
if necessary to improve performance and robustness. system response in power systems.
8. Deployment and Integration: Table.2 CNN Performance in Scenario 2 & 3
Once validated, deploy the trained CNN model for Errors
Accuracy
Scenario CNN Test Single Two Three
real-time fault finding and categorization in the IEEE 34 bus (%)
phase Phase Phase
system. Integrate the CNN-based fault detection system with 1- 2342 46 0 40 96.32
existing control and monitoring systems for seamless phase
operation and enhanced system reliability. By following this 2
2- 3054 0 122 4 95.87
procedure, CNNs can effectively detect and classify faults in phase
3- 1050 0 4 8 98.85
the IEEE 34 bus system, providing valuable insights for phase

15th ICCCNT IEEE Conference,

Authorized licensed use limited to: VIT University- Chennai [Link] 24-28, 2024,
Downloaded on May 22,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
IIT - Mandi, Kamand, India
 IEEE - 61001

1- 2050 22 12 21 97.31 Test Actual Measured Actual Measured Error

phase No Type Type of Location Location (%)
2- 3100 0 125 32 94.93 of Fault of Fault of Fault
3
phase
Fault
3- 1125 2 30 0 97.15
phase
1 LG 150 145.51 2.99
2 LG 240 248.82 3.68
LG
3 LG 295 289.55 1.85
Table 2 presents the working metrics of a CNN model in fault
4 LG 800 796.25 0.47
categorization for Scenario 2 and Scenario 3 in a power 5 LLG 60 63.05 5.08
system. The scenarios represent different operational 6 LLG 300 298.56 0.48
conditions or configurations. The CNN's performance is LLG
7 LLG 720 715.32 0.65
evaluated based on test data, categorized by fault types such 8 LLG 905 9.3.26 0.19
as 1-phase, 2-phase, and 3-phase faults. For each scenario and 9 LLLG 218 216.85 0.53
fault type, the table provides the number of test cases, the 10 LLLG 608 6.5.86 0.35
LLLG
number of errors (misclassifications) made by the CNN, and 11 LLLG 965 960.38 0.48
the accuracy achieved in percentage. In Scenario 2, the CNN 12 LLLG 900 896.20 0.42
achieved high accuracies ranging from 95.87% to 98.85%
across different fault types, with single-phase faults having Table 3 presents fault location and error calculations for
the lowest error rate. In Scenario 3, the CNN's accuracy various test cases in a power system. Each test case is
ranged from 94.93% to 97.31%, again with single-phase numbered, and the table includes columns for the actual type
faults exhibiting the highest accuracy. These results of fault, measured type of fault, actual location of the fault,
demonstrate the CNN's effectiveness in accurately measured position of the fault, and the error percentage. The
classifying faults in the power system under varying actual sort of fault refers to the known fault type in the
scenarios, highlighting its potential for enhancing fault system, such as LG (Line-to-Ground), LLG (Line-to-Line-to-
detection and classification capabilities in real-world Ground), or LLLG (Triple Line-to-Ground). The measured
applications. type of fault indicates the fault type identified or measured by
the system under test. The actual location of the fault
represents the true location of the fault in the power system,
while the measured position of the fault indicates where the
system detected the fault. The error percentage column
calculates the percentage difference between the actual and
measured fault locations, providing insight into the accuracy
of fault location detection. Overall, the table demonstrates the
system's ability to accurately detect and locate faults, with
most error percentages ranging from less than 1% to around
5%, highlighting the effectiveness of fault location
algorithms or techniques used in the power system.

Table.4 Fault classification accuracy using CNN

Fault Accuracy Classification
Type Voltage Current Voltage & Current
(%) (%) (%)
Without With
Fig.2 Training & test data for accuracy using CNN ME ME
R-g 100 84 100 90
Y-g 100 100 100 100
B-g 78 100 98 100
RY-g 100 100 100 100
YB-g 100 64 100 90
RB-g 80 100 100 100
R-Y 100 100 100 100
Y-B 100 86 100 85
B-R 100 100 100 100
RYB-g 100 100 100 100
No fault 100 100 100 100
Table 4 presents the fault classification accuracy using a
Convolutional Neural Network (CNN) for different fault
types based on voltage, current, or both voltage and current
Fig.3 Training & test data for Loss using CNN measurements. The table includes columns for accuracy
classification percentages for voltage, current, and combined
Table.3 Fault Location & error calculation voltage and current measurements, with and without
mechanical effects (ME). Each row corresponds to a specific
fault type, such as R-g (Line-to-Ground fault in the Red

15th ICCCNT IEEE Conference,

Authorized licensed use limited to: VIT University- Chennai [Link] 24-28, 2024,
Downloaded on May 22,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
IIT - Mandi, Kamand, India
 IEEE - 61001

phase), Y-g (Line-to-Ground fault in the Yellow phase), B-g architecture for fault categorization and finding tasks in the
(Line-to-Ground fault in the Blue phase), and various IEEE 34 bus system showcase the potential of advanced
combinations like RY-g, YB-g, RB-g, R-Y, Y-B, B-R, and machine learning techniques in addressing complex
RYB-g. Additionally, the table includes a row for no-fault challenges in power engineering. Through the training
conditions. The accuracy percentages indicate CNN's ability process using labeled data sets containing various fault
to correctly classify faults based on the input data, with higher scenarios, including 1-phase, 2-phase, and 3-phase faults, the
percentages indicating more accurate classification. The table CNN model is optimized to improve accuracy and robustness
also compares the classification accuracy with and without in fault identification. Simulation results validate the
considering mechanical effects, showing the impact of ME effectiveness of the CNN-based approach in accurately
on fault classification performance. Overall, the table classifying and detecting faults, highlighting its practical
demonstrates CNN's effectiveness in accurately classifying applicability in real-world scenarios. The ability of CNNs to
different fault types based on voltage, current, or combined distinguish between different fault types contributes to more
measurements, highlighting its potential for fault detection efficient fault management processes, leading to reduced
and classification in power systems. The accuracy & loss downtime and improved grid reliability. The broader
representation of test & train data were shown in Fig.2 & 3 implications of this research extend to the field of intelligent
respectively fault management systems in power engineering. By
Table.5 Comparative analysis of different methods with harnessing the capabilities of CNNs, power system operators
CNN can streamline fault management processes, enhance system
Sl. Method used Input Signal Accuracy performance, and ultimately contribute to a more reliable and
no Considered (%) resilient electrical grid.
1 WPT +SVM Current 98.2
2 Fuzzy Voltage & 99.1 REFERENCES
Current [1] Santos, A. D. S., Faria, L. T., Lopes, M. L. M., Lotufo, A. D. P., &
Minussi, C. R. (2022). Efficient methodology for detection and
3 Proposed CNN Voltage & 99.6 classification of short-circuit faults in distribution systems with
Current distributed generation. Sensors, 22(23), 9418.
[2] Gururajapathy, S. S., Mokhlis, H., & Illias, H. A. (2017). Fault location
The Table.5 presents a comparison of three methods for and detection techniques in power distribution systems with distributed
generation: A review. Renewable and sustainable energy reviews, 74,
signal processing and classification: WPT + SVM, Fuzzy 949-958.
logic, and the proposed CNN. WPT + SVM uses only the [3] Nsaif, Y. M., Lipu, M. H., Ayob, A., Yusof, Y., & Hussain, A. (2021).
current signal and achieves 98.2% accuracy. It is effective for Fault detection and protection schemes for distributed generation
feature extraction and classification, particularly with non- integrated to distribution network: Challenges and suggestions. IEEE
stationary signals, but it relies on a single input type and Access, 9, 142693-142717.
requires complex parameter optimization for SVM. Fuzzy [4] Mnyanghwalo, D., Kundaeli, H., Kalinga, E., & Hamisi, N. (2020).
Deep learning approaches for fault detection and classifications in the
logic, using both voltage and current signals, achieves a electrical secondary distribution network: Methods comparison and
higher accuracy of 99.1%. It handles uncertainty and recurrent neural network accuracy comparison. Cogent
imprecision well but requires expert knowledge to define Engineering, 7(1), 1857500.
membership functions and rules. The proposed CNN, also [5] Zhang, C., Wang, H., Wang, Z., & Li, Y. (2023). Active detection fault
diagnosis and fault location technology for LVDC distribution
using both voltage and current signals, achieves the highest networks. International Journal of Electrical Power & Energy
accuracy of 99.6%. CNNs excel in automatic feature Systems, 148, 108921.
extraction and learning complex patterns without manual [6] Ghaemi, A., Safari, A., Afsharirad, H., & Shayeghi, H. (2022).
feature engineering. However, they require substantial data Accuracy enhance of fault classification and location in a smart
and computational resources for training. Despite this, their distribution network based on stacked ensemble learning. Electric
Power Systems Research, 205, 107766.
superior accuracy and versatility make them a promising
[7] Hosseinimoghadam, S. M. S., Dashtdar, M., & Dashtdar, M. (2021).
choice for signal processing and classification tasks. Fault location in distribution networks with the presence of distributed
generation units based on the impedance matrix. Journal of The
V. CONCLUSION Institution of Engineers (India): Series B, 102, 227-236.
[8] Waqar, H., Bukhari, S. B. A., Wadood, A., Albalawi, H., & Mehmood,
K. K. (2024). Fault identification, classification, and localization in
In conclusion, the application of Convolutional Neural microgrids using superimposed components and Wigner distribution
Networks (CNNs) for fault classification and detection in the function. Frontiers in Energy Research, 12, 1379475.
IEEE 34 bus system represents a significant advancement in [9] Awasthi, S., Singh, G., & Ahamad, N. (2023). Identification of type of
power system management. The study demonstrates the a fault in distribution system using shallow neural network with
effectiveness of CNNs in automatically learning hierarchical distributed generation. Energy Eng, 120(4), 811-829.
features from system data, specifically voltage and current [10] Stefanidou-Voziki, P., Sapountzoglou, N., Raison, B., & Dominguez-
Garcia, J. L. (2022). A review of fault location and classification
measurements, without requiring manual feature extraction. methods in distribution grids. Electric Power Systems Research, 209,
By leveraging this capability, the CNN-based methodology 108031.
achieves high accuracy rates in fault classification and rapid
response times for fault detection, thereby enhancing the
reliability and resilience of distribution networks. The
development and implementation of a tailored CNN

15th ICCCNT IEEE Conference,

Authorized licensed use limited to: VIT University- Chennai [Link] 24-28, 2024,
Downloaded on May 22,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
IIT - Mandi, Kamand, India