See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/336916610

Artificial Neural Network Based Fault Classification and Location for

Transmission Lines

Conference Paper · October 2019

DOI: 10.1109/CPERE45374.2019.8980173

CITATIONS READS

0 255

3 authors, including:

Ahmed Elnozahy Khairy Sayed

Assiut University Sohag University Faculty of Engineering
15 PUBLICATIONS   72 CITATIONS    92 PUBLICATIONS   393 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

A Hybrid Fuel Cell-Battery Generation for Electric Vehicles View project

resilient DC microgrid View project

All content following this page was uploaded by Ahmed Elnozahy on 31 October 2019.

The user has requested enhancement of the downloaded file.

 2019 IEEE Conference on Power Electronics and Renewable Energy (CPERE)

Artificial Neural Network Based Fault

Classification and Location for Transmission Lines

Ahmed Elnozahy Khairy Sayed Mohamed Bahyeldin

Electrical Engineering department Electrical Engineering department. Control and communication sector
Faculty of Engineering Faculty of Engineering Middle Egypt electricity distribution
Assiut University Sohag University Assiut, Egypt
Assiut, Egypt Sohag, Egypt [email protected]
[email protected] http://orcid.org/0000-0001-8871-5358

Abstract—Due to various faults occur to transmission lines network [14] [15] [16], neuro-fuzzy [17] [18] [19] and
and because it was necessary to find and recover these faults genetic [20].
quickly as possible. This paper discussing fault detection, The advance in the field of power system protection
classification and determining fault location as fast as possible technology is going faster with the advancements in
via Artificial Neural Network (ANN) algorithm. The software
electronics and the development of new high speed designed
used for modeling the proposed network is a
MATLAB/SIMULINK software environment. The training, microprocessors like DSPs [21], which enable Power
testing and evaluation of the intelligent locator processes are systems protective relays to feel and separate fault area
done based on a multilayer Perceptron feed forward neural quickly.
network with back propagation algorithm. Mean Square Error Additionally, the ease of use of power system application
(MSE) algorithm is used to evaluate the performance of the software packages, like MATLAB software enhanced the
detector/classifier as well as fault locator. The results show that study of power systems using blocks instead of using very
the validation performance (MSE) for the fault complicated equation models [22].
detector/classifier is 2.36e-9 and for fault locator is 2.179e-5. Application of ANN to discover, classify fault type, and
The system can detect if there is a fault or not, can classify the
discover fault place in transmission lines is proposed by
fault type and determine the fault location very precisely.
[23]. This approach composed of two parts. The first part,
Keywords— Artificial Neural Network, Fault location, the fault detected and classified, and the second part, the
Transmission Lines, Detector/classifier, Mean Square Error. fault position is determined. Joorabian et al presented an
RBF (radial basis function) neural networks based fault
I. INTRODUCTION classifier and locator. The results indicated that the
Electric power transmission lines systems play a vital maximum error of the fault location algorithm is 0.5 % [24].
role in delivering electrical power to customers. The size of Banu et al developed a new fault location method for a
transmission network is increasing as demand for power transmission line. The method uses one ANN based on the
increases. Although the transmission lines are planned to Levenberg Marquardt optimization technique. However, the
stand the weather conditions, the variation of temperature fault detection and the fault classification are not indicated
and the mechanical stresses, a fault can happen on power in this research. However, the error of this method was is
network for various reasons; the majority of faults causes still below 0.65 % [25]. Another neural network design to
are natural such as winds, falling trees, birds and lightning. detect, determine fault type and find fault place in
Sometimes cars and planes collides the transmission towers. transmission lines are suggested by [26]. However, these
Others may be technical like over load, broken insulators, approaches detected only the faults happened in 80 % of the
cut of conductors and malfunction of any power system transmission line [26].
components such as transformers and generators. That M. B. Hessine et al proposed ANN-based-fault classifier
means faults are unavoidable. Faults cause a very high technique based on input data of each phase currents and the
current, which can cause great damage to the power network zero sequence current component (four pre-fault and four
components, Disconnection of power feeding which means post-fault for each phase and zero sequence currents
de-energizing factories and vital establishments and components) instead of pre and post phase voltages.
customers and voltage level drop which is harmful to Moreover, the time domain of the four samples was used
electrical equipment. For these reasons, discovering faults instead of commonly used FFT [27].
and determining fault location is vital for both electrical The main object in this paper is to find a new accurate
networks and customers. The protection relays actually play method to detect and determine the fault type and find the
this role in discovering and disconnecting power to the fault. fault place in power transmission lines. MATLAB/
The faster the fault location is determined, the quicker the SIMULINK based ANN is used to detect, determine the fault
system can be restored and disconnection time can be type, and find the place of faults in long transmission lines.
shorter. Various fault location procedures are used to define The ANN is a good online and offline method to achieve fast
fault detection, determining fault types, and finding fault
the fault position accurately [1] [2] [3] [4] [5]. The most
location in overhead transmission lines.
dominant techniques being used is impedance-based method
[6-8], travelling wave method [9] [10] [11], and Artificial II. POWER SYSTEM UNDER CONSIDERATION
intelligence methods [12] [13] such as artificial neural
Model of the proposed 132kV, and 100km length,
transmission line using MATLAB/SIMULINK is illustrated

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

978-1-7281-0910-7/19/$31.00 ©2019 IEEE 140

 2019 IEEE Conference on Power Electronics and Renewable Energy (CPERE)

in Fig. 1. The system composed of two sections of pi model

blocks for the overhead transmission line, and a three-phase
fault block as shown in Fig. 1. The specifications of the
proposed transmission line are illustrated in TABLE I. The
type and position of the fault as well as the fault resistance
and inception angle can be varied via the
MATLAB/SIMULINK fault block. Consequently, the data
exported to the workspace. The phase voltage Va waveform
during fault illustrated in Fig. 2.
A. Proposed Fault Location system Fig. 2 Single phase to ground fault phase voltage, Va.
The proposed system composed of two parts; the first
part is data processing. In this part, The obtained data from
the measuring devices is filtered then the data is sampled and Input Data
Sampling
from Anti-aliasing
transformed to frequency components to obtain the measuring filter
process at 2 FFT process
kHz
fundamentals and then data is normalized. devices
The second part is the neural network fault detector/classifier
and the fault locator. Fig. 3 illustrates the proposed system
block diagram. ANN
Determination
detector/ Normalization
B. Data processing of fault location
classifier
The output Data of the MATLAB/SIMULINK model
includes three phase voltages and currents before and during
fault are being used as the input data to the system.
1) Anti-Aliasing Filter: The anti-aliasing filter removes ANN algorithm
the unwanted frequencies from a sampled waveform. A Fig. 3 Proposed system block diagram.
simple 10th order low-pass Butterworth filter with a cut-off
frequency of 400 Hz is integrated.

Fig. 4 Sampled phase voltage Va during A-G fault

3) Fast Fourier Transform: Tow full cycle Fast Fourier

Transform (FFT) is used to calculate the magnitudes of
fundamental components of three-phase currents and
voltages before and after the fault. The fundamentals of Va
before and during a fault in the frequency domain using FFT
Fig. 1 Simulink model of the studied system are illustrated in Fig. 5.
2) Sampling: The three phase currents and voltages
signals are sampled at a sampling frequency of 2 kHz,
which is very Suitable for power systems. As illustrated in
Fig. 4.

TABLE I Specifications of proposed transmission line.

Specification Value
Length 100 Km
Positive sequence resistance 0.044965 Ohms/km
Zero sequence resistance 0.11241 Ohms/km.
Positive -sequence inductances 1.01e-3 H/km
zero-sequence inductances 2.02e-3 H/km Fig. 5. Pre fault and post fault frequency components of Phase voltage, Va.
Positive -sequence capacitance 7.471e-9 F/km
zero-sequence capacitance 4.394e-9 F/km C. Normalization
The input signals samples have to be normalized as the input
data must be in the range of (±1) this step is very important
for the neural network to find suitable weights in short time

141
 2019 IEEE Conference on Power Electronics and Renewable Energy (CPERE)

as data changes in a small range (±1). These pre-processed If the ANN keep changing W, the MSE will increase again,
data were normalized to suit the ANN Input data. so the values of W at the lowest MSE is the best values, and
saved as the final values of W. and so the training process is
III. ARTIFICIAL NEURAL NETWORK completed. In addition, the ANN can predict in the future
An Artificial Neural Network (ANN) is biologically any output at a given input. The Backpropagation learning
inspired architectures from the biological neural cells in algorithm is used to train the fault detector/classifier and
human brains. It can be described as a set of neurons that are locator.
interconnected and arranged in several layers. MSE used to measure error and to measure the performance
The structure of a feed-forward ANN, also called as the of the ANN. The Levenberg-Marquardt training algorithm is
perceptron. A neuron mathematical model has a much used, which is also known as the damped least-squares
simpler structure comparing with a biological neuron [28]. method, which designed to work with loss functions, which
Fig.6 shows multi- input neuron [29], and Fig. 7 illustrates takes the sum of squared errors.
three layers perceptron.
A neuron j can be described mathematically with the IV. DETECTOR/CLASSIFIER ANN
following equation: Model in Fig. 1 is used to simulate 10 types of faults (A-
φ ∑ ∗ (1) G, B-G, C-G, A-B, B-C, C-A, A-B-G, B-C-G, C-A-G, A-B-
C-G). A 100 case each (every 1 km) to get 1000 cases of
Where: w ∗ a is the threshold value (polarization), f φ is different fault types (fault resistance is changed between
the neuron activation function, Φ is the summation output 0.25 ohm and 125 ohms and inception angle between (20,
signal and y is the neuron output. 40, 70) degrees at each case. In addition, 100 cases of no
φ W ∗A (2) fault. The output of every case is three phase currents and
Where A and W are the inputs and weights vectors, voltages.
respectively, which defined as: After data processing, the normalized ratio of pre-fault and
W w w w … . . w , A a a a … . a post-fault fundamental frequency component of each output
There are several types of activation functions. Such as, step (Vaf2/Vaf1, Vbf2/Vbf1, Vcf2/Vcf1, Ia2/Ia1, Ib2/Ib1, and
function, piece wise linear function, sigmoid, unipolar Ic2/Ic1) are used as an input of (1100×6) matrix to the ANN.
function and sigmoid bipolar function. However, the corresponding code of each case is used as
An activation function decides the effect of the on the sum (4× 1100) target data matrix.
of the inputs on the outputs. Depending on the best After several trays of changing, the number of neurons and
performance, the appropriate activation function is chosen. number of hidden layers and trying different activation
The basic operation of ANN is to find the relation between functions the structure of 6-12-6-4 used for the detector/
input data matrix A, and target data matrix Y, by classifier with transfer functions (tan -pure line -tan –tan)
determining weights. In the beginning, ANN assumes the which gave the best validation performance of 2.3e-9, which
initial values of W, and calculates the error between outputs is very satisfactory.
and targets and changes the values of W, until reach the The training performance and the best validation
suitable MSE. This is called error backpropagation. performance are shown in Fig. 8 & Fig. 9. Where Fig. 10
shows the training regression of the detector/classifier.
TABLE II shows that the accuracy of the detector/classifier
is 100%.

Fig. 6 Typical multi-input neuron

Fig. 7 Three layer perceptron.

Fig. 8 Training performance of the detector/classifier.

142
 2019 IEEE Conference on Power Electronics and Renewable Energy (CPERE)

tan) which gave the best validation performance 2.179e-5,

which is very accurate, Fig. 11 Shows the best structure for
ANN locator.
Mean Squared Error (mse)

VI. RESULTS AND DISCUSSION

The neural network divides input data into three categories
training set, validation set and test set, is the actual dataset
that we use to train the model. The validation set is used
infrequently evaluating the model during the training
process. The test dataset is used for evaluating the model
after it completes the training process as a final evaluation.
The training performances are shown In Fig. 8, while the
MSE performance for the three datasets plotted in different
colors are shown in Fig. 9 & Fig.12. Another important
Fig. 9 The training regression of the detector/classifier.
factor in evaluating the ANN is the best linear regression
TABLE II the output of the detector/classifier. that relates the targets to the outputs as shown in Fig. 10 &
Fig.13.
Fault Target The linear regression indicates the ability of the ANN to
Detector/classifier output
type predict outputs in the future. The correlation coefficient (r)
A-C-G 1 0 1 1 1.000 0.000 1.0000 1.000
is a factor indicates how strong the relationship between
targets and outputs (0 being no correlation at all and 1 being
A-C-G 1 0 1 1 1.000 0.000 1.0000 1.000
0.000 1.000 1.0000 0.000
complete correlation). Fig 10 & Fig. 13 show that the
B-C 0 1 1 0
1.000 1.000 0.0000 0.000
relationship between outputs and targets are linear as it can
A-B 1 1 0 0
A-B 1 1 0 0 1.000 1.000 0.0000 0.000 be expressed as:
A-B 1 1 0 0 1.000 1.000 0.0000 0.000 Y ax b
A-G 1 0 0 1 1.000 0.000 0.0000 1.000 The correlation coefficient is found to be 0.99969 for test
A-G 1 0 0 1 1.000 0.000 0.0000 1.000 data, which is very satisfactory. The final step is testing the
ANN using a separate set of data. TABLE III shows the
outputs of the ANN and the real data. It shows a very
satisfactory error percentage.

Fig. 11 The best structure for ANN fault locator.

Fig. 10 Regression of detector/classifier ANN.

Fig. 12 The best validation performance for fault locator.

V. FAULT LOCATOR ANN
VII. CONCLUSIONS
The model is used to simulate the 10 types of faults, at
distances (10, 20, 30, 40… 100 km). The fault resistance is A fault classifier/locator based on multilayer perceptron
changed between (0.25, 0.5, 5, 10, 25, 50, 75, 100) ohms, feed forward neural network with back propagation
and inception angles is changed between (20 , 40 ,70) A 240 algorithm is proposed in this paper to classify and locate
cases are obtained as an input matrix (240×6), while the faults in power system transmission lines. It uses the
corresponding distances as a target matrix. After several magnitude of the fundamental frequency for fault
trays of changing the number of neurons and hidden layers classification and location. MATLAB/ SIMULINK is used
and trying different activation functions, the structure of 6- to model and simulate the transmission line and the neural
16-6-6-1 is used with transfer functions (tan -log -tan –tan- network. A separate detector/classifier was developed to
detect and classify ten different shunt faults.

143
 2019 IEEE Conference on Power Electronics and Renewable Energy (CPERE)

. [7] T. Takagi, Y.Yamakoshi, M. Yamaura, R. Kondow, and T.

Matsushima,” Development of a New Type Fault Locator Using the
Training: R=0.99995 Validation: R=0.99984 One-Terminal Voltage and Current Data”, Power Apparatus and
1 1 Systems, IEEE Transactions on, vol. 101: pp. 2892-2898, 1982.
Output ~= 1*Target + 0.00053

Output ~= 1*Target + 0.00021

Data Data
Fit
0.8
Fit [8] J. Izykowski, E. Rosolowski, and M.M. Saha.” Locating faults in
0.8 Y=T Y=T
parallel transmission lines under availability of complete
0.6 0.6
measurements at one end. Generation, Transmission and
Distribution”, IEE Proceedings, vol. 151, pp. 268-273, 2004.
0.4 0.4 [9] S. Lin S, Z. Y. He, X. P. Li, and Q. Q. Qian, “Travelling wave time-
frequency characteristic-based fault location method for transmission
0.2 0.2 lines,” IET Gener. Transm. Distrib., vol. 6, 2012.
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
[10] Z. Yi-ning, L. Yong-hao, X. Min, and Cai Ze-xiang, “A novel
Target Target algorithm for HVDC line fault location based on variant travelling
wave speed,” in: 4th Int. Conf. Electric Utility Deregulation and
Test: R=0.99969 All: R=0.99991 Restructuring and Power Technologies, 2011.
1 1
Output ~= 0.99*Target + 0.0025

[11] A. L. Kulikov and V. V. Ananev, “Adaptive wave determination of

Output ~= 1*Target + 0.0007

Data Data

0.8
Fit
Y=T 0.8
Fit
Y=T
fault locations in electric power transmission lines,” Vestnik
Ivanovsk. Gos. Énerget. Univ., vol. 4, 2014.
0.6 0.6 [12] S. K. Kumar, M. Swamy, and V. Venkatesh, Artificial Neural
Network Based Method for Location and Classification of Faults on a
0.4 0.4 Transmission Lines, Int. J. Sci. Res. Publ., vol.4, pp. 2250–3153,
2014.
0.2 0.2
[13] J. C. S. Souza, M. A. P. Rodrigues, M. T. Schilling, and M. B. D. C.
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 Filho, “Fault location in electrical power systems using intelligent
Target Target systems techniques,” IEEE Trans. Power Deliv., vol. 16, pp. 59–67,
2001.
Fig.13 The linear regression for training, validation and test between target [14] S. B. Ayyagari,” Artificial neural network based fault location for
and output. transmission lines”, University of Kentucky Master's Theses,2011.
[15] T.Dalstein, B. Kulicke, “Neural network approach to fault
TABLE III the outputs of the ANN and the real data. classification for high speed protective relaying”, IEEE Transactions
on Power Delivery, vol. 4, pp. 1002 – 1009, 1995.
Fault Fault Inception Actual ANN Error% [16] M. El-Sharkawi, D. Niebur, “A tutorial course on artificial neural
type Resistance angle distance distance networks with applications to Power systems”, IEEE Publ., vol. 96TP
A-G 0.3 20 65 65.08 0.08% pp. 11-20, 1996.
B-G 15 40 47 47.16 0.16% [17] W. H. Chen, C. W. Liu, and M. S. Tsai, “On-line fault diagnosis of
C_G 5 20 37 36.99 0.01% distribution substations using hybrid cause–effect network and fuzzy
A_G 12 20 92 92.08 0.08% rule-based method,” IEEE Trans. Power Delivery, vol. 15, pp. 710–
B-C 25 70 24 24.13 0.13% 717, Apr. 2000.
A-B 0.5 40 64 64.06 0.06% [18] O. Babayomi et al,”Neuro-fuzzy Based Fault Detection Identification
ABG 5 20 32 32.09 0.09% and Location in a Distribution Network Cent. for Space Transport &
BCG 25 40 67 0.6708 0.08% Prop. Nat. Space Res. & Dev. Agency Nigeria,2016.
[19] R. Syahputra,” A neuro-fuzzy approach for the fault location
estimation of unsynchronized two-terminal transmission
Another separate network for fault locator was lines”,Department ,Universitas Muhammadiyah Yogyakarta,
developed. The performance of the detector/classifier and Yogyakarta, 55183, INDONESIA,2013.
locator was measured using Mean Square Error (MSE) a [20] Liu D, Du X, Li Y, Huang Y,” Fault location using two terminal data
validation performance of 2.3e-9 is achieved and accuracy for HV& long transmission line based on genetic Algorithm”. High
of 100% for the detector/classifier. A validation Voltage Engineering; vol.33, pp. 21–25, 2007.
performance MSE was achieved of 2.179e-5 for the fault [21] Arun G. Phadke and James S. Thorp, “Computer Relaying For Power
Systems,” New York: Wiley, 1988.
locator, which is a very good performance. That proves that
[22] "Using MATLAB,” The Mathworks Inc., Natick, MA, January 1999.
the neural network fault classifier/locator is a powerful and
reliable tool to classify and locate faults in power [23] Tahar, B.: Fault location in EHV transmission lines using artificial
neural networks. Int.J. Appl. Math. Comput. Sci., vol. 14, pp. 69–780,
transmission lines. 2004.
[24] Joorabian, S.M.A., Taleghani, A.S.L., Aggarwal, R.K” Accurate fault
REFERENCES locator for EHV transmission lines based on radial basis function
[1] J. Iżykowski , “Fault location on power transmission lines”, Oficyna neural network”. Electr. Power Syst. Res., vol. 71, pp. 195–202,2004.
Wydawnicza Politechniki Wrocławskiej Wrocław,vol.1, pp10-15, [25] Banu, G., Suja, S,” ANN based fault location technique using one end
2008. data for UHV lines”, Eur. J. Sci. Res., vol. 77, pp. 549–559, 2012.
[2] S. Ghimire “Analysis of fault location methods on transmission [26] Hassan, K.Z., Zuyi, L,” An ANN based approach to improve the
lines”, University Of New Orleans, vol. 2, pp. 16-19, 2014. distance relaying algorithm”, Turkish J. Electr. Eng. Comput. Sci.,
[3] W. P. Davis, “Analysis Of faults In Overhead Transmission Lines, vol. 14, pp. 345–354, 2006.
California State University, Sacramento,vol. 1, pp.13-16, 2012. [27] M. Hessine, H. Jouini and S. Chebbi, “Neural Network Approach to
[4] IEEE Guide for Determining Fault Location on AC Transmission and Fault Location for High Speed Protective Relaying of Transmission
Distribution Lines. IEEE Std C37.114-2004, 2005: p. 1-36. Lines”, Springer International Publishing Switzerland, 2015.
[5] P. M. Anderson, “Analysis of Faulted Power Systems”. ISBN 0-7803- [28] Jang, J.S.R.: ANFIS: adaptive- network-based fuzzy inference
1145-0, The Institute of Electrical and Electronics Engineers, 1973. system. IEEE Trans. Syst. Man Cybern., vol. 23, pp. 665–684, 1993.
[6] K. Zimmerman and D. Costello. “Impedance-Based Fault Location .
Experience”. Schweitzer Engineering Laboratories, Inc, SEL Journal
of Reliable Power, vol. 1, July 2010.

View publication stats

144