FAULT CLASSIFICATION LOCATION AND DETECTION IN POWER

SYSTEM USING NEURAL NETWORK

This project report is submitted for partial fulfilment of requirement for the
degree of Bachelor of Technology in Electrical Engineering under Maulana
Abul Kalam Azad University of Technology, West Bengal.

Session: 2023-2024

Submitted by

NAME MAKOUT ROLL NO REGISTRATION NO

SOHAM DAS 14801620009 201480101610007

ARPAN KUMAR BHANDARI 14801620008 201480101610008

TRIDIB PRADHAN 14801620004 201480101610012

Under the Guidance of: Prof. Arnab Roy (Asst. Prof. Dept. of
EE)

1
 ACKNOWLEDGEMENT
With a deep sense of gratitude, I wish to express my
sincere thanks to my guide, Prof. Arnab Roy, Assistant
Professor in the department of Electrical Engineering
for giving me the opportunity to work under him on
this project. I also want to thank Dr. Avijit Saha, Head
of the Department, Electrical Engineering for providing
me a solid background for my research thereafter. He
has been a great source of inspiration to me and I
thank him from the bottom of my heart.
Finally, I would like to thank all those people whose
direct and indirect support helped me completing my
project in time. I would like to thank my department
for giving me the opportunity and platform to make my
effort a successful one.
Soham Das
Arpan Kumar Bhandari
Tridib Pradhan
(Department of Electrical Engineering)

2
 CONTENT

ABSTRACT 4

CHAPTER 1 INTRODUCTION 5

1.1 INTRODUCTION 6

CHAPTER 2 REVIEW OF LITARATURE 7

2.1 NEURAL NETWORK 8

CHAPTER 3 METHODOLOGY 9

3.1 METHODOLOGY 10

CHAPTER 4 WORKING OF THE 11

PROJECT AND CIRCUIT DIAGRAMS

4.1 MODELLING THE TRANSMISSION NETWORK 12

4.2 HOW TO TRAIN NEURAL NETWORK 14
4.3 CIRCUIT DIAGRAM 22

CHAPTER 5 RESULT AND DISCUSSION 24

25
5.1 RESULT AND DISCUSSION 28
5.2 RESULTS OF FAULT CLASSIFIER AND LOCATOR FOR
DIFFERENT FAULT TYPES AND FAULT LOCATIONS

CHAPTER 6 ADVANTAGES AND CONCLUSION 30

5.1 ADVANTAGES 31

5.2 CONCLUSION 33

CHAPTER 7 BIBLIOGRAPHY 34
BIBLIOGRAPHY 35

3
ABSTRACT
In our project, we're using neural networks to help
manage faults in power systems. Instead of relying on
complex methods or manual checks, our approach uses
a computer system that learns from data to quickly
figure out what's wrong and where it's happening in the
power grid.

We trained our system with information about different

types of faults, like short-circuits or equipment failures,
and the signals that show up when these faults occur.
With this training, the system can accurately recognize
and locate faults in real-time, helping prevent
disruptions and keep the power flowing smoothly.

Our experiments show that this approach works well

even in different situations, making power systems
more reliable and responsive to problems.

4
CHAPTER- 1

5
1.1 INTRODUCTION
 Fault classification, location and detection in power
systems are critical tasks for ensuring the stability
and reliability of electrical grids. Traditional
methods for fault detection often rely on complex
algorithms and specialized equipment. However,
with advancements in artificial intelligence,
particularly in neural networks, there’s a growing
interest in using these technologies for fault
analysis in power system.
 On of the key advantages of neural networks is their
ability to adopt and learn from new data, which can
improve the accuracy of fault detection systems
over time. Challenges remain in developing neural
network models that are robust enough to handle
the variability and complexity of real-world power
systems.
 We explore the application of neural networks for
fault classification, location and detection in power
systems. We aim to develop a reliable and efficient
system that can help improve the stability and
reliability of electricals grids, ultimately
contributing to the advancement of smart grid
technologies.
6
CHAPTER- 2

7
2.1 NEURAL NETWORK
Traditional methods of fault detection and location
often rely on complex algorithms and manual analysis.
These methods may not always be accurate or
efficient, especially in large and complex power
systems. There's a need for smarter, more automated
approaches to fault management.
So, here we can use neural networks as a smarter
option. Neural networks are computer systems
inspired by the human brain, capable of learning from
data and making intelligent decisions. In the context of
power systems, Neural networks are very effective for
managing faults in power systems because they can
learn from data patterns. By using advanced data
processing and training methods, they create a strong
system for detecting, classifying, and locating faults,
which makes power systems more reliable and efficient.

8
CHAPTER- 3

9
3.1 METHODOLOGIES
This section presents the methodologies of the research
such as the modelling the transmission network,
simulation of various fault conditions, training the
neural network and test the neural network.

10
CHAPTER- 4

11
4.1 MODELLING THE TRANSMISSION
NETWORK
A simple transmission line system which was modelled
using MATLAB Simulink. The model consists of
equivalent Thevenin source, voltage and current
measurements, transmission line, fault configuration
and three-phase load. The parameters used in the model
is listed in below. The transmission line is spur
connection type where the current flow is in one
direction from the source to the load.

Transmission line system

12
 Transmission line system parameters

The transmission line system in above was used to

generate the training data for the fault classifier and
locator. For every fault simulation, the sequence values
of three-phase voltages and currents were recorded.
Figure 3 shows the blocks inside the sequence values
subsystem. The first part is the filter used to remove the
high frequency components from the fundamental
frequency waveforms. Then, the filtered three-phase
voltage and current are fed into sequence analyser
blocks to get the positive, negative and zero sequence
values of the voltage and current. Next, the magnitudes
of the sequence values during fault occurrence are sent
to workspace of the MATLAB for recording purpose.

13
4.2 HOW TO TRAIN NEURAL
NETWORK
For training purpose, the fault conditions were varied
for different fault types, fault resistance values and
fault locations. The higher the number of training data,
the higher the accuracy of the fault classifier and
locator. Five different fault types have been simulated
which are AG, AB, ABG, ABC and ABCG which represent
the single line-to-ground (SLG) fault, line-to-line (LL)
fault, line-to-line-to-ground (LLG) fault, three-line (LLL)
fault and three-line-to-ground (LLLG) fault respectively.
For each SLG, LL, and LLG fault, only one type of fault
needs to be simulated because the other two types of
faults will have the same values for the same simulated
fault type. This is the advantage of using sequence
values as the input for the neural network compared
with using the phase values where the voltage and
current for all phases have to be recorded. Below table
shows the variation of fault types, fault resistance
values and fault locations for the simulated fault cases
for training data purpose. The output for the neural
network should be in value. Thus, the different fault
types have been assigned with different values as
shown below-
14
INPUT FAULT TYPES REPRESENTATION

FAULT Va Vb Vc Ia Ib Ic
TYPE/FAULT
NO FAULT 167600 167600 167600 766.1 766.1 766.1
AG(0) 97.67 178600 161800 1448 788.9 750.7
BG(0) 161800 97.67 178600 751.3 1408 788.5
CG(0) 178600 161800 98.24 788.4 750.9 1432
AG(50) 38330 172700 161900 1228 708.4 747.1
BG(50) 161600 37000 172800 746.7 1196 708.7
CG(50) 173200 162200 43250 708.8 746.9 1214
AG(100) 59820 171700 162100 1077 645.7 723.6
BG(100) 162000 59700 171700 722.5 1050 646.5
CG(100) 171700 162100 59850 646.8 723.2 1064
AB(0) 83712.6 83761.7 167476.7 1539.0 995.1 777.3
AB(50) 75882.1 96900.0 168524.0 1386.0 884.5 724.1
AB(100) 75629.6 108963.8 170369.8 1261.0 794.5 676.9
BC(0) 167476.7 83712.5 83761.7 777.3 1522.0 988.2
BC(50) 168523.9 90116.2 108881.8 724.1 1369.0 877.5
BC(100) 170369.8 87267.2 117421.3 676.9 1245.0 787.6
AC(0) 83761.7 167476.7 83712.6 1016.0 777.3 1559.0
AC(50) 109964.0 168523.9 92452.5 902.1 724.1 1403.0
AC(100) 119083.2 170369.8 89690.7 809.0 676.9 1275.0
ABC(0) 55.5 55.5 55.5 1449.0 1409.0 1434.0
ABC(50) 44708.9 39487.6 72253.5 1294.0 1259.0 1276.0
ABC(100) 57142.0 52140.3 76202.6 1168.0 1137.0 1154.0
ABG(0) 59.3 103.5 177186.2 1449.0 1408.0 695.8
ABG(50) 35067.6 41270.5 171308.7 1290.0 1202.0 673.6
ABG(100) 48321.3 60136.7 169264.8 1168.0 1055.0 641.2
BCG(0) 177199.4 58.8 103.2 696.5 1409.0 1433.0
BCG(50) 171009.0 57663.4 61377.0 673.6 1258.0 1219.0
BCG(100) 169159.3 64932.2 73869.2 640.8 1141.0 1067.0
CAG(0) 102.8 177199.4 59.2 1448.0 696.6 1434.0
CAG(50) 63825.5 170972.1 60013.1 1237.0 673.5 1280.0
CAG(100) 76724.7 169142.6 67669.3 1086.0 640.6 1160.0

15
 OUTPUT FAULT TYPES
REPRESENTATION
FAULT TYPE/FAULT A B C G
NO FAULT 0.0 0.0 0.0 0.0
AG(0) 1.0 0.0 0.0 1.0
BG(0) 0.0 1.0 0.0 1.0
CG(0) 0.0 0.0 1.0 1.0
AG(50) 1.0 0.0 0.0 1.0
BG(50) 0.0 1.0 0.0 1.0
CG(50) 0.0 0.0 1.0 1.0
AG(100) 1.0 0.0 0.0 1.0
BG(100) 0.0 1.0 0.0 1.0
CG(100) 0.0 0.0 1.0 1.0
AB(0) 1.0 1.0 0.0 0.0
AB(50) 1.0 1.0 0.0 0.0
AB(100) 1.0 1.0 0.0 0.0
BC(0) 0.0 1.0 1.0 0.0
BC(50) 0.0 1.0 1.0 0.0
BC(100) 0.0 1.0 1.0 0.0
AC(0) 1.0 0.0 1.0 0.0
AC(50) 1.0 0.0 1.0 0.0
AC(100) 1.0 0.0 1.0 0.0
ABC(0) 1.0 1.0 1.0 0.0
ABC(50) 1.0 1.0 1.0 0.0
ABC(100) 1.0 1.0 1.0 0.0
ABG(0) 1.0 1.0 0.0 1.0
ABG(50) 1.0 1.0 0.0 1.0
ABG(100) 1.0 1.0 0.0 1.0
BCG(0) 0.0 1.0 1.0 1.0
BCG(50) 0.0 1.0 1.0 1.0
BCG(100) 0.0 1.0 1.0 1.0
CAG(0) 1.0 0.0 1.0 1.0
CAG(50) 1.0 0.0 1.0 1.0
CAG(100) 1.0 0.0 1.0 1.0
16
The fault classifier and locator have been trained using
the neural fitting tool inside the MATLAB where the
zero sequence values of voltage and current were the four
input elements for the tool. The target for the tool was
fault location in km for fault locator and represented fault
type value for fault classifier. The total number of
training data is 30 samples which consist of training data
for different fault types, fault location values as shown
below.

17
DATA SHEET FOR DISTANCE
CALCULATION

FAULT DISTANCE V SEQUENCE I SEQUENCE

NO FAULT 0.00 0.00
AG 0.00 88190.00 555.50
BG 0.00 88190.00 534.40
CG 0.00 88190.00 548.40
ABG 0.00 83460.00 519.20
BCG 0.00 83470.00 527.40
ACG 0.00 83470.00 506.60
ABCG 0.00 0.00 0.05
AG 20.00 77500.00 488.80
BG 20.00 77500.00 469.40
CG 20.00 77500.00 481.70
ABG 20.00 72200.00 448.70
BCG 20.00 72170.00 456.70
ACG 20.00 72170.00 437.90
ABCG 20.00 7.65 0.04
AG 30.00 72920.00 460.70
BG 30.00 72910.00 442.00
CG 30.00 72730.00 435.50
ABG 30.00 67150.00 419.70
BCG 30.00 67420.00 427.70
ACG 30.00 67440.00 409.70
ABCG 30.00 6.39 0.03
AG 40.00 69070.00 435.50
BG 40.00 69060.00 417.50
CG 40.00 69070.00 428.30
ABG 40.00 63120.00 394.20
BCG 40.00 63350.00 402.10
ACG 40.00 63360.00 384.90
ABCG 40.00 0.38 0.24
AG 50.00 65150.00 412.70
BG 50.00 65170.00 395.50
CG 50.00 64780.00 405.60

18
ABG 50.00 59860.00 371.50
BCG 50.00 59810.00 379.20
ACG 50.00 59800.00 362.80
ABCG 50.00 15.88 0.01
AG 60.00 62020.00 392.10
BG 60.00 62020.00 375.50
CG 60.00 62020.00 385.10
ABG 60.00 56260.00 351.20
BCG 60.00 56450.00 358.80
ACG 60.00 56480.00 343.10
ABCG 60.00 11.14 0.01
AG 70.00 58960.00 373.30
BG 70.00 58970.00 357.40
CG 70.00 58880.00 366.30
ABG 70.00 53750.00 332.90
BCG 70.00 53640.00 340.30
ACG 70.00 53630.00 325.40
ABCG 70.00 2.11 0.01
AG 80.00 56350.00 356.30
BG 80.00 56340.00 340.80
CG 80.00 56350.00 349.30
ABG 80.00 51300.00 316.40
BCG 80.00 51070.00 323.70
ACG 80.00 51030.00 309.30
ABCG 80.00 2.13 0.00
AG 90.00 53700.00 340.60
BG 90.00 53710.00 325.70
CG 90.00 53540.00 333.70
ABG 90.00 48320.00 301.40
BCG 90.00 48480.00 308.60
ACG 90.00 45100.00 294.70
ABCG 90.00 3.43 0.01
AG 100.00 51500.00 326.20
BG 100.00 51490.00 311.80
CG 100.00 51500.00 319.40
ABG 100.00 46490.00 287.70
BCG 100.00 46400.00 294.80
ACG 100.00 46380.00 281.50
ABCG 100.00 10..57 0.00487500

19
For this neural fitting tool, the number of hidden neurons
chosen is 10. After each training of fault classifier and
locator, the value of Mean Squared Error (MSE) and
Regression, R were recorded, and the trained model was
saved. From 20 times of training for both fault classifier
and locator, the trained model which has the lowest value
of MSE, and R was used as the final NN based fault
classifier and locator. Below figure shows an example of
trained model which can be generated after each finished
training.

Generated neural model

The best trained model for fault classifier and locator
then were inserted and connected into the transmission
line system model as shown in circuit diagram. The
model was used to the test the accuracy of the NN based
fault classifier and locator for various fault types, fault

20
resistance values and fault locations differ from the fault
conditions used previously to generate the training data.

21
4.3 CIRCUIT DIAGRAM

22
CIRCUIT DIAGRAM AT RUNNING
CONDITION

23
CHAPTER- 5

24
5.1 Results and Discussion
The results of error for neural network training of fault
classifier and locator are shown below. The results in
both figures represent the best trained model from 20
times of training for both fault classifier and locator. The
number of samples used for training, validation and
testing are 25, 3 and 3 samples and 16, 2 and 2
respectively. The MSEs of training and testing are very
low for both trained fault classifier and locator. While the
regressions of training and testing are nearly one for both
fault classifier and locator.
Results of errors for the best trained model of fault
classifier-

Results of errors for the best trained model of fault

locator-

25
The last test is the most important test which used to test
the accuracy of the trained fault classifier and locator.
This test used the test data from different simulated fault
conditions other than the data used during the training
stage. The results of fault detection and location for this
test is shown
Performances of training state for: (a) Fault
classifier, (b) fault locator-

(a)
26
 (b)
Results of fault classifier and locator for different fault
types, fault resistance values and fault locations.

27
 5.2 RESULTS OF FAULT CLASSIFIER
AND LOCATOR FOR DIFFERENT
FAULT TYPES AND FAULT
LOCATIONS-
NO. FAULT TYPE ESTIMATED FAULT CORRECT FAULT ACTUAL FAULT LOCATION AND
TYPE VALUE FOR TYPE DETECTION? ESTIMATED FAULT LOCATION
EVERY FAULT (KM)
1 1-0-0-1 YES 10 7.914

2 1-0-0-1 YES 30 29.91

3 AG 1-0-0-1 YES 60 58.66

4 1-0-0-1 YES 80 74.7

5 1-0-0-1 YES 100 98.34

1 0-1-0-1 YES 10 16.91

2 0-1-0-1 YES 30 29.9

3 BG 0-1-0-1 YES 60 60.73

4 0-1-0-1 YES 80 77.92

5 0-1-0-1 YES 100 101.8

1 0-0-1-1 YES 10 10.87

2 0-0-1-1 YES 30 30.16

3 CG 0-0-1-1 YES 60 58.25

4 0-0-1-1 YES 80 81.19

5 0-0-1-1 YES 100 99.31

1 1-1-0-1 YES 10 26.65

2 1-1-0-1 YES 30 30.38

3 ABG 1-1-0-1 YES 60 73.99

4 1-1-0-1 YES 80 102

5 1-1-0-1 YES 100 99.8

28
From the above table, the NN based fault classifier
successfully classified all the fault types which have
different fault resistance values and locations. All the
estimated fault type values are very near to the actual
fault type values. Then for the fault location, the NN
based fault locator accurately located the faults where the
average error of fault location is 0.705 % (less than 1%).
However, there are several fault conditions which have a
slightly higher error than 1%. The accuracy of the fault
location estimation can be enhanced by training the fault
locator with higher number of samples which have
higher number of different fault resistance and fault
location values. Besides that, the number of trainings
also can be increased until the lowest MSE and
regression errors been found. From the table, the fault
classifier and locator still can classify and locate the AG,
BC and ACG faults although the fault classifier and
locator been trained with different SLG, LL and LLG
faults respectively.

29
CHAPTER- 6

30
6.1 ADVANTAGES

- Accuracy: Neural networks can achieve high levels of

accuracy in fault detection and location, reducing false
alarms and downtime.
- Speed: They can quickly analyse data and react to
faults in real-time, minimizing the impact on the power
system.
- Adaptability: Neural networks can adapt to changing
conditions in the power grid, making them versatile and
robust.

- Handling Complex Data: Neural networks can adapt

to changes in the power system and learn from new
data, improving over time.

- Reduced Outages: Faster fault detection and

location lead to quicker repairs, reducing the
duration and frequency of power outages.

- Cost Effective: By preventing extensive damage

and reducing downtime, neural networks can lead
to significant cost and savings in maintenance and
repairs.

31
- Scalability: Neural network can be scalded to
monitor large and complex power systems
effectively.

-Integration with Modern Technologies: Easy to

integrate with other modern technologies, such as
lot devices and smart grids, for enhanced
functionality.

32
6.2 CONCLUSION

Neural network-based fault classifier and locator have

been successfully developed to classify and locate
faults at a transmission line. Both fault classifier and
locator were trained with the training data which
represent the positive, negative and zero sequence
values for different fault types, fault resistance values
and fault locations. From the results of the test on the
simulated faults which differ from the simulated faults
used to train the networks, all fault types and their
locations have been successfully classified and located.
The advantage of using sequence values as the inputs
is the number of samples used to train the network can
be significantly reduced compared with the method
which use the phase values as the inputs. This is
because, only one fault type needs to be simulated for
each SLG, LL and LLG fault types for the same fault
resistance value and fault location.

33
CHAPTER- 7

34
7.1 BIBLIOGRAPHY
 M. Dashtdar, M. Esaeilbeig, M. Najafi, and M. E.
N. Bushehri, “Fault Location in the Transmission
Network Using Artificial Neural Network,” Autom.
Control Comput. Sci., vol. 54, no. 1, pp. 39–51,
2020.
Bhuyan, B. K. Panigrahi, K. Pal, and S. Pati,
“Convolutional Neural Network Based Fault
Detection for Transmission Line,” Int. Conf. Intel.
Controll. Comput. Smart Power, 2022.
 M. Nithyavelam and Siddappanaidu, “Novel fault
detection on three-phase transmission line using
modified artificial neural network with wavelet-
based analysis,” ARPN J. Eng. Appl. Sci., vol. 14,
no. 2, pp. 556–564, 2019.
 Pandey, P. S. Gadekar, and C. B. Khadse,
“Artificial neural network-based fault detection
system for 11 kV transmission line,” 1st Int. Conf.
Adv. Electr. Comput. Commun. Sustain. Technol.,
2021.
You tube channel: Learn MATLAB Simulink
Link: https://youtu.be/RLZFjXA7tiY?si=yDmu1NqoKYOm6-
x-

35