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

net/publication/317428704

Power System Fault Detection, Classification and Location using Artificial

Neural Networks

Conference Paper · May 2017

DOI: 10.1007/978-3-319-71321-2_8

CITATIONS READS

9 7,715

4 authors, including:

Tatjana Konjic Admir Jahić

Institute for Systems and Computer Engineering, Technology and Science (INESC … JP Elektroprivreda BiH d.d. Sarajevo
53 PUBLICATIONS 227 CITATIONS 12 PUBLICATIONS 46 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Admir Jahić on 09 June 2017.

The user has requested enhancement of the downloaded file.

Power System Fault Detection, Classification and Location using
Artificial Neural Networks

Almin Karić1, Dr.Sc. Tatjana Konjić2, Mr.Sc. Admir Jahić3

1 Almin Karić
Faculty of Electrical Engineering, University of Tuzla
Tuzla, Bosnia and Herzegovina
e-mail: [email protected]

1 Dr.Sc. Tatjana Konjić

Faculty of Electrical Engineering, University of Tuzla
Tuzla, Bosnia and Herzegovina
e-mail: [email protected]

1 Mr.Sc. Admir Jahić

Faculty of Electrical Engineering, University of Tuzla
Tuzla, Bosnia and Herzegovina
e-mail: [email protected]

Abstract

This article focuses on detecting, classifying and locating faults in power system using Artificial Neural Net-
works (ANNs). Feed-forward neural networks have been employed and trained with back-propagation algo-
rithm.
The model of WSCC 9 bus test system has been modelled in Matlab/Simulink, and used to validate the pro-
posed fault detection system. First, normal state of the model was observed. After determination of normal state,
different types of faults have been simulated on all nine buses and on all lines of the model. Voltage and current
magnitudes, obtained by the fault simulation, are used as inputs of the ANN. Output of the ANN should provide
information about the fault type and location in case the fault occurs.
Number of hidden layers and neurons per hidden layer is determined by testing performance and training time
of each developed neural network.

1. Introduction

Power systems all over the world are experiencing huge and rapid expansion. End users who are very sensi-
tive to power outages are demanding reliable and uninterrupted supply of electric power [1]. On the other side,
the appearances of large generations and highly interconnected systems are making early fault detection and rap-
id equipment isolation the most important functions to maintain system stability. One of the factors that hinder
the continuous supply of electricity and power is a fault in the power system [2]. Power lines encounter various
faults due to tree contact, animal contact or other natural causes such as thunderstorms. These faults cannot be
avoided, since they occur because of natural reasons which humans cannot control. It is very important to have
protection system that can detect any abnormal flow of current in the power system, identifies the type of fault
and accurately locates the position of the fault in the power system [6]. The automatic location of the fault can
greatly enhance the systems reliability because faster the power is restored, more money and time is saved.
Conventional algorithms based on deterministic computations and well-defined model of power lines, are re-
sulting in the late detection and inaccurate results [1-5]. Conventional distance relays consider power swing as a
fault. Such malfunctioning could lead to serious disturbance of power system stability.
 Artificial intelligence based methods are being used in the process of fault detection and location to accelerate
fault detection and to improve performance of protection system. There are three major artificial intelligence
based techniques that are widely used in the power and automation industry:
 Expert System Techniques [7-9],
 Artificial Neural Networks [6, 10, 11, 15], and
 Fuzzy Logic Systems [12-14].
ANN based methods do not require a knowledge base for the detection of faults. The purpose of fault detec-
tion is to detect, locate and classify the fault in the power system as fast and accurate as possible [15]. Special
features of ANN for fault tolerant systems:
 ANN consists of layers of neurons which are connected with each other. In this formation it is possible to car-
ry out parallel processing of the information. This helps processing huge amount of data in real time.
 ANN can solve any degree of nonlinearity and the fault detection problems are nonlinear.
 ANN approach is non-algorithmic and there is no need for knowing functions that are relating the problem
variables. Also, there is no need for approximations unlike the most of the mathematical models.
 ANN is capable of handling situations of incomplete information, corrupt data and thus is highly fault toler-
ant.
In this work ANN based approach is used to detect, locate and classify faults that occur in the network. De-
veloped Artificial Neural Networks have three layers: input, hidden and output. The input layer is receiving data
from the power system, and then hidden layers are processing that data so we can get the result on the output
layer. Neural networks are capable of working in real time environment and responding to the changes immedi-
ately, which makes the system for the fault detection fast and flexible.

2. Artificial Neural Networks

Artificial neural network theory has served to better identify how the neurons in the brain function and to pro-
vide the basis for efforts to create artificial intelligence [15]. Neural network can be defined as a composition of
elements for receiving, processing and transferring information. Those elements are called neurons. The neurons
use activation function that converts neuron's weighted input to its output activation. The activation function can
be divided into linear and nonlinear functions [16].
There are various activation functions which are chosen depending on the type of problem to be solved by the
network. The most common activation function includes linear function, tangent hyperbolic function, threshold
function and sigmoidal function [16]. Figure 1 shows a simple neuron.

Fig. 1. A single neuron model

Values w1j, w2j,…, wnj are weights to determine the strength of input vector X=[x1, x2,….., xn]T of neuron j.
Each input is multiplied with its associated weight of the neuron. To generate the final output yj, the sum zj is
passed on to a nonlinear filter f(z) called activation function or transfer function, which releases the output.
In ANN applications, the challenge is to find the right values for the weights, based on the known inputs and
outputs. Various numbers of hidden layers and numbers of neurons in each layer, transfer functions and training
algorithms can be employed depending on the complexity of the problem to be solved [15].
 3. Faults in Power System

There are two types of faults in the power system: short circuit and open circuit. A short circuit is a fault in
which a current travels along an unintended path with no or very low electrical impedance. Opposite of a short
circuit is an open circuit fault which is an infinite resistance between two nodes. Information gained from fault
studies are used for proper relay setting and coordination [17, 18]. Fault studies are also used to obtain the rating
of the protective switchgears. The magnitude of the fault currents depends on the internal impedance of the gen-
erator plus the impedance of the intervening circuit [18].
Short circuit faults are divided into three-phase balanced faults and unbalanced faults. Different types of un-
balanced faults are single line-to-ground fault, line-to-line fault and double line-to-ground fault.
The most often faults, around 70-80%, are single line-to-ground faults. Three-phase faults are very rare,
around 5-10% [17].
In this paper, the short circuit faults will be simulated on WSCC 9 bus test system.

3.1. WSCC 9 bus test system

WSCC bus systems are used by researchers to implement new ideas and concepts. The system is shown in
Figure 2 [19].
This test system consists of 9 buses, 3 generators, 3 two-winding power transformers, 6 lines and 3 loads. The
bus 1 is a slack bus, which is used to regulate any changes in the power system. The buses 5, 6 and 8 are repre-
senting a PQ buses, or loads. All values are given in per units and are reduced to base power of 100 MVA.

Fig. 2. WSCC 9 bus test system

3.2. Simulation of WSCC 9 bus test system in Simulink

Simulink model of WSCC 9 bus test system is shown in Figure 3.

Diagrams of the simulation can be viewed from the Data Acquisition subsystem, shown in Figure 4. Each
SCOPE element has two inputs, data of three-phase voltages and data of three-phase currents.
Fig. 3. WSCC 9 bus test system in Simulink

Fig. 4. Data Acquisition subsystem

3.2.1 Simulating the test system in normal state

Starting simulation it’s possible to get voltages and currents in the system in normal state.
Voltage and current diagrams of bus 1 with no fault, as well as of bus 8 are shown in Figure 5.
a) b)
Fig. 5. Diagrams of voltages and currents: a) on bus 1 with no fault, b) on bus 8 with no fault

3.2.2. Simulating the test system with faults

Three-Phase Fault element from Simulink library is used for simulating any type of fault. From its settings it
is possible to select any phase (A, B, C) and ground (G) that will be „connected“. By selecting the phase A and
G, the single line-to-ground fault will be simulated, and by selecting phases A and B the two-phase fault will be
simulated. Voltage and current diagrams for both faults obtained by the simulation are shown in Figure 6.

a) b)
Fig. 6. Diagrams of voltages and currents: a) Single line-to-ground fault on bus 1 b) Two-phase fault on bus 1

4. Artificial neural network for fault detection, classification and location

Developed artificial neural network consists of three layers: input, hidden and output. Input layer depends on
the system and data types. For WSCC 9 bus test system input data will be three values of each phase voltage
magnitude and three values of each phase current magnitude, for all buses in the system, which means there will
be a total of 54 inputs in a network. Input values are obtained by simulating the system during the short circuit,
and then placed in a table that could be used to train the neural network.
The output layer depends on the function of a neural network. Neural network must be able to locate and clas-
sify fault. Location and classification of fault can be realized with 5 outputs from the neural network, four of
which are used for classification (three phases and one ground), and the last output provides information about
the location of the fault on the transmission line. Developed model of the neural network is presented in Figure
7.

Fig. 7. Model of the neural network

Activation function at the input is log-sigmoid because input data are in per units, while the transfer function
at the output is saturating liner because the distance where the fault occurs can receive a value between 0 and 1.
The neural network should be able to find a location of the fault even in the locations it is not trained for. This
will be done in a way to first simulate all the faults at the beginning of the line (single-phase, two-phase and
three-phase), the resulting voltage and current magnitudes will be placed in the input data table. For the output
layer first four values depending on the type of failure are selected, and the last output is set to 0 (which repre-
sents the start of the line).
In the same way it is necessary to simulate all types of faults at the end of the line. In this case, the output
which represents the distance is set to 1. For additional precision the line is separated in two and all kinds of
faults are simulated in the middle of the line. Output of the neural network representing the distance is set to 0.5
and other outputs in accordance with the type of fault.
This way a table of input and output data is obtained. Table 1 provides information needed to train ANN to
locate and classify single-phase short circuit.
 Table 1. Input and output data for single-phase short circuit
Output Data
A 0 1 0 0 1 0 0 1 0 0
B 0 0 1 0 0 1 0 0 1 0
C 0 0 0 1 0 0 1 0 0 1
G 0 1 1 1 1 1 1 1 1 1
Distance 0 0 0 0 0.5 0.5 0.5 1 1 1
Input Data
No fault AG - 0 BG - 0 CG - 0 AG - 50m BG - 50m CG - 50m AG - 1 BG - 1 CG -1
VB1_a 0.8555 0.5022 0.8280 0.8438 0.5042 0.8280 0.8485 0.4854 0.8266 0.8266
VB1_b 0.8558 0.8477 0.5066 0.8312 0.8474 0.5012 0.8354 0.8510 0.4848 0.4848
VB1_c 0.8557 0.8279 0.8466 0.4941 0.8262 0.8473 0.5153 0.8281 0.8493 0.8493
IB1_a 1.1330 1.9994 1.2914 1.0342 1.9846 1.2745 1.0389 2.0729 1.2721 1.2721
IB1_b 1.1326 1.0997 1.7153 1.2215 1.0978 1.6847 1.2088 1.1006 1.7625 1.7625
IB1_c 1.1322 1.2887 1.1074 1.3733 1.2734 1.1045 1.3711 1.2726 1.1049 1.1049
VB2_a 0.8655 0.3157 0.8805 0.9057 0.3616 0.8530 0.8982 0.4109 0.8424 0.8424
VB2_b 0.8656 0.8928 0.3383 0.9347 0.8906 0.3641 0.8779 0.8693 0.4093 0.4093
VB2_c 0.8654 0.8955 0.9024 0.4721 0.8716 0.8976 0.3824 0.8435 0.8701 0.8701
IB2_a 0.8916 2.3079 0.8996 0.9044 2.0421 0.9019 0.8711 1.8907 0.9617 0.9617
IB2_b 0.8913 0.9318 2.2401 0.8776 0.8815 1.8962 0.8903 0.8869 1.7138 1.7138
IB2_c 0.8914 0.9000 0.9456 1.4243 0.8968 0.8822 1.2652 0.9555 0.8858 0.8858
VB3_a 0.8730 0.2944 0.8679 0.9270 0.2491 0.8707 0.9079 0.0702 0.8768 0.8768
VB3_b 0.8727 0.9015 0.2877 0.9040 0.9092 0.2533 0.8850 0.9332 0.0702 0.0702
VB3_c 0.8730 0.8908 0.9149 0.3233 0.8733 0.9080 0.3171 0.8768 0.9345 0.9345
IB3_a 0.6242 1.5739 0.6373 0.6152 1.7669 0.6384 0.6425 2.0900 0.6483 0.6483
IB3_b 0.6236 0.6355 1.5201 0.6207 0.7018 1.7338 0.5802 0.7852 2.0867 2.0867
IB3_c 0.6238 0.6364 0.6330 0.9586 0.6388 0.7046 1.0481 0.6488 0.7866 0.7866
VB4_a 0.8535 0.4347 0.8261 0.8348 0.4361 0.8243 0.8447 0.4193 0.8223 0.8223
VB4_b 0.8536 0.8385 0.4424 0.8300 0.8406 0.4396 0.8375 0.8422 0.4174 0.4174
VB4_c 0.8533 0.8251 0.8352 0.4331 0.8248 0.8399 0.4653 0.8228 0.8463 0.8463
IB4_a 0.6563 1.1387 0.7223 0.6289 0.7592 0.6967 0.6316 0.4518 0.6800 0.6800
IB4_b 0.6560 0.6543 0.9795 0.7052 0.6428 0.6798 0.6800 0.6380 0.4501 0.4501
IB4_c 0.6557 0.7220 0.6629 0.7982 0.6982 0.6469 0.5974 0.6781 0.6379 0.6379
VB5_a 0.8353 0.3635 0.8244 0.8163 0.3977 0.8127 0.8264 0.3944 0.8093 0.8093
VB5_b 0.8354 0.8245 0.3564 0.8274 0.8169 0.3950 0.8248 0.8172 0.3944 0.3944
VB5_c 0.8350 0.8193 0.8158 0.4071 0.8232 0.8212 0.4075 0.8110 0.8176 0.8176
IB5_a 0.4470 1.0735 0.4602 0.4431 0.6677 0.4556 0.4407 0.4105 0.4492 0.4492
IB5_b 0.4467 0.4843 0.9703 0.4299 0.4814 0.6511 0.4262 0.4756 0.4052 0.4052
IB5_c 0.4469 0.4584 0.4910 0.6029 0.4547 0.4777 0.3596 0.4483 0.4760 0.4760
VB6_a 0.8483 0.3732 0.8309 0.8388 0.3462 0.8362 0.8426 0.2820 0.8331 0.8331
VB6_b 0.8485 0.8287 0.3811 0.8513 0.8446 0.3524 0.8333 0.8488 0.2822 0.2822
VB6_c 0.8484 0.8439 0.8328 0.3757 0.8313 0.8412 0.3718 0.8331 0.8477 0.8477
IB6_a 0.3683 0.8410 0.3576 0.3684 1.2138 0.3568 0.3813 1.6536 0.3538 0.3538
IB6_b 0.3683 0.4110 0.7677 0.3296 0.4137 1.0656 0.3121 0.4271 1.3959 1.3959
IB6_c 0.3683 0.3579 0.4087 0.4565 0.3531 0.4199 0.7568 0.3528 0.4317 0.4317
VB7_a 0.8650 0.2346 0.8811 0.9221 0.2918 0.8521 0.9009 0.3424 0.8414 0.8414
VB7_b 0.8652 0.8922 0.2591 0.9484 0.8990 0.2934 0.8666 0.8688 0.3425 0.3425
VB7_c 0.8649 0.9123 0.9200 0.4154 0.8724 0.9015 0.3150 0.8409 0.8707 0.8707
IB7_a 0.5119 3.4711 0.5460 0.5261 2.7337 0.5831 0.5153 2.2502 0.6283 0.6283
IB7_b 0.5117 0.5464 3.1509 0.5302 0.5426 2.4617 0.5493 0.5470 2.0081 2.0081
IB7_c 0.5118 0.5404 0.5461 2.0100 0.5816 0.5427 1.6478 0.6287 0.5466 0.5466
VB8_a 0.8575 0.0001 0.8783 0.9088 0.1345 0.8601 0.8872 0.2150 0.8475 0.8475
VB8_b 0.8577 0.9052 0.0001 0.9222 0.8858 0.1343 0.8641 0.8656 0.2196 0.2196
VB8_c 0.8578 0.8850 0.9013 0.0001 0.8609 0.8847 0.1456 0.8498 0.8695 0.8695
IB8_a 0.5224 3.7556 0.5394 0.5826 2.8028 0.5794 0.5531 2.3027 0.6301 0.6301
IB8_b 0.5221 0.5702 3.4475 0.5868 0.5690 2.5055 0.5628 0.5770 2.0565 2.0565
IB8_c 0.5222 0.5375 0.5727 2.5846 0.5824 0.5732 1.7059 0.6301 0.5789 0.5789
VB9_a 0.8739 0.2429 0.8698 0.9309 0.1929 0.8680 0.9106 0.0001 0.8737 0.8737
VB9_b 0.8739 0.9076 0.2347 0.8989 0.9134 0.1982 0.8829 0.9409 0.0001 0.0001
VB9_c 0.8736 0.8928 0.9186 0.2742 0.8718 0.9143 0.2944 0.8755 0.9421 0.9421
IB9_a 0.6224 1.5742 0.6357 0.6134 1.7678 0.6368 0.6410 2.0913 0.6465 0.6465
IB9_b 0.6219 0.6339 1.5203 0.6189 0.7004 1.7345 0.5785 0.7837 2.0879 2.0879
IB9_c 0.6221 0.6346 0.6314 0.9581 0.6370 0.7030 1.0478 0.6471 0.7852 0.7852

Next problem is to determine the number of neurons in the hidden layer. This can only be determined by test-
ing, so the characteristics of the network are tested for 5, 10, 20, 30, 35 and 40 neurons in the hidden layer.
Figure 8 shows the results of the neural network with different number of neurons in the hidden layer. The
network is capable of locating the faults for places it is trained for (at the begging, in the middle and at the end of
the line) without a problem. Negligible deviation can be seen when locating the fault in the middle of the line.
Fig. 8. Characteristics of the ANN for single-line fault

Figure 9 shows the characteristics of neural networks with different number of neurons in the hidden layer for
the fault at the location of 25 km from the begging of the line. Neural network was not trained for this location so
there are certain deviations in the output for the fault location, between the real location of the fault and the loca-
tion obtained with the neural network.

Fig. 9. Locations of different faults obtained by ANN with different number of neurons

Network with 5 neurons in the hidden layer can be used only for single-phase faults, while for other types of
faults obtained results are not acceptable. Two-phase faults with or without the ground can be located if the net-
work has 10 or more neurons in the hidden layer. However, for the three-phase faults, neural network has to have
at least 30 neurons in the hidden layer to be able to relatively well locate the fault. From the Figure 9, it is obvi-
ous that the network with 35 neurons in the hidden layer is the most accurate for all types of faults.

Conclusion

Obtained results clearly indicate that ANNs can be used for detection, classification and location of faults.
Detection, location and classification of faults for the points where the network was trained, is an easy task for
ANN. Also, results are showing that ANN can relatively accurately detect, locate and classify the fault even for
the places it is not trained for. Depending on the type of fault, networks with different number of neurons in hid-
den layer can be used. Single line-to-ground fault can be detected and located with the smallest number of neu-
rons in hidden layer, only 5. Network for the detection, location and classification of two-phase and two-phase-
to-ground faults needs to have minimum of 10 neurons in the hidden layer. The three-phase fault requires neural
network with 30 to 35 neurons in hidden layer while further increase of neurons in the hidden layer does not lead
to improvement of results.
 References

[1] Das R., Novosel D., “Review of fault location techniques for transmission and sub-transmission lines”, Proceedings of 54th An-
nual Georgia Tech Protective Relaying Conference, 2000
[2] IEEE guide for determining fault location on AC transmission and distribution lines. IEEE Power Engineering Society Publ.,
New York, IEEE Std C37.114, 2005
[3] Saha M.M., Izykowski J., Rosolowski E., “Fault Location on Power Networks”, Springer publications, 2010
[4] Magnago F.H., Abur A., “Advanced techniques for transmission and distribution system fault location”, Proceedings of CIGRE –
Study committee 34 Colloquium and Meeting, Florence, 1999, paper 215
[5] Tang Y., Wang H.F., Aggarwal R.K., “Fault indicators in transmission and distribution systems”, Proceedings of International
conference on Electric Utility Deregulation and Restructuring and Power Technologies – DRPT, 2000, pp. 238-243
[6] Ayyagari S. B., “Artificial neural network based fault location for transmission lines“, University of Kentucky Master's Thesis,
2011
[7] Hsu Y. Y., Lu F. C., Chien Y., Liu J. P., Lin J. T., Yu H. S., Kou R. T. “An Expert System for Locating Distribution System
Faults”, IEEE Transaction on Power Delivery, Vol. 6, No. 1, January 1991
[8] Yongli Z., Yang Y.H., Hogg B. W., Zhang W.Q., Gao S., “An expert system for power systems fault analysis”, IEEE Trans. on
Power Systems, Vol. 9, No. 1, Feb., 1994, pp 503- 509
[9] Chang C.S., Chen J.M., Liew A. C., “An expert system approach for fault diagnosis considering uncertainties,” Proceedings of
International Conference on Intelligent Manufactureing’95 (ICIM’95), Wuhan, China, June, 1995
[10] Butler K. L. & Momoh J. A., “Detection and classification of line faults on power distribution systems using neural networks”,
Midwest symposium on circuit and systems, p368-371, 1993
[11] Eisa B. M. T., “Faults Detection in Power Systems Using Artificial Neural Network”, American Journal of Engineering Re-
search (AJER), Volume-02, Issue-06, pp-69-75, 2013
[12] Frank P.M., Seliger B.K., “Fuzzy logic and neural network applications to fault diagnosis”, International Journal of Approxi-
mate reasoning, 1997
[13] Mor V., Vaghamshi A., “Review on fault detection, identification and localization in electrical networks using fuzzy-logic”, In-
ternational Journal of Applied OF Research in Science and Engineering, November 2016
[14] Adhikari S., Sinha N., Dorendrajit T., “Fuzzy logic based on-line fault detection and classification in transmission line”, Avail-
able at: http://link.springer.com/article/10.1186/s40064-016-2669-4
[15] Seema S., Mamatha K. R.,Thejaswini S., “Intelligent Fault Identification System for Transmission Lines Using Artificial Neural
Network”, IOSR Journal of Computer Engineering (IOSR-JCE), 2014
[16] Tatjana K., Goran Š., “Odlučivanje i optimizacija”, Repro Karić d.o.o., Tuzla, 2010
[17] Nikola R., Dragan T., Gojko S., “Distributivne i industrijske mreže”, ETF Akademska misao, Beograd 2004
[18] Haadi S., “Power System Analysis”, R.R.Donnelley & Sons, Boston, 1999
[19] Manitoba HVDC Research Centre a division of Manitoba Hydro International Ltd. “IEEE 09 Bus System”, 2014

View publication stats