IRJET- A Simple Approach to Identify Power System Transmission Line Faults using PNN

INTERNATIONAL RESEARCH JOURNAL OF ENGINEERING AND TECHNOLOGY (IRJET) E-ISSN: 2395-0056
VOLUME: 06 ISSUE: 04 | APR 2019 WWW.IRJET.NET P-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 1008
A Simple Approach to Identify Power System Transmission Line Faults
using PNN
A. Acharya1, A. Mukherjee2, S. Jana3
-----------------------------------------------------------------------***------------------------------------------------------------------------
Abstract - Large interconnected overhead transmission lines
are very often subjected to various types of faults leading to
mal-operation or even discontinuity of good quality power
flow. Hence, accurate and prompt identification and
classification of different power systemfaultshave become
one of the major areas of research these days to ensure
reliability, protection and system stability.This paper
presents a Probabilistic Neural Network (PNN) based simple
and novelapproachto distinguish and classify different types
of power system faults. Electromagnetic Transient Program
(EMTP) simulation software has been used to simulate a
frequency dependent model and extract the receiving end
voltage data only for various faults carried out at different
distances. Only two sets of simulated data corresponding to
two different locations are fed directly to the PNN algorithm
so designed for training, instead of preprocessing the data by
filtering tools like Wavelet Transform, component analysis
etc. which is done most commonly, thus requiring lesser
computation burden, thus saving time. Experimental
verification has been carried out to validate the proposed
algorithm.
Index Terms—Probabilistic Neural Network (PNN),
Electromagnetic Transient Program (EMTP),
I. INTRODUCTION
Electrical power transmissionsystem is one of the most
widespread and largest engineering networks meant for
energy transfer. The lion’s share of this power transmission
is done by overhead transmission lines which remain
directly in contact with the atmosphere, and hence, quite
often are subjected to several atmospheric hazards leading
to minor to severe power system faults. Sometimes these
faults are temporary in nature, and sometimes, sustain for a
long time. The relays connected to the transmission lines and
the Current Transformer (CT) and Potential Transformer
(PT) sense the line currents and voltages and send the actual
signals to the controller unit which senses the fault, if any,
and sends an actuating (or tripping) signal to the respective
circuit breakers which operate accordingly to interrupt the
fault current at the exact location or bus specified. The
situation complicates further when it is a three phase
system, since the fault could occur in any one or two or may
be among all the three lines, may be involving ground in
some cases. Isolation of only the faulty phase is the most
essential in order to ensure a reliable and continuous power
supply, even at a reduced power flow through the healthy
phases. Therefore, quickidentification and classification of
faults is of most importance to ensure power system stability
and reliability. The proposedresearch work is intended to
serve the same purpose, i.e. to detect power system fault, if
any, and categorize the fault in different fault types for a 400
kV, three phase, single circuit, 150 km long overhead
transmission line.
The primary objective of power system fault analysis is to
extract some key features from the fault waveforms using
the different identifier techniques and provide these data to
the detection mechanism to find the true cause of the fault
and restore the normal operation in least possible
time.Techniques normally used to develop such power
system protection algorithm are Wavelet Transformation
based Artificial Neural Network (ANN) analysis, Fuzzy
Inference system, Principal Component Analysis (PCA),
support vector analysis etc.and differenthybrid models of
these individual techniques.The tripping circuit of the circuit
breakers and other protective relaying mechanisms are
given proper signals at the correct instants during any mal-
operation, to interrupt the power flow immediately.Hence,
fast and accurate detection of fault along with precise fault
location have been practiced by researchers to ensure
safety.Researchers have devoted much time for the purpose
of fault identification and classification and developed
different mathematical and computational tools for the
detection, classification and localization of electrical power
system faults to activate the tripping elements at proper time
and proper locations[8].
Artificial Neural Network (ANN) along with Neuro Fuzzy
System and Wavelet based fault pattern recognition
techniques have been extensively used these days [1]-[6].
But the ANN implementation requires a large number of
training data as well as training cycles, thus requiring
heavier amount of data and computational burden [7]-
[11].Other techniques include wavelet transform with other
methods such as Probabilistic Neural Network (PNN),
adaptive resonance theory, adaptive neural fuzzy inference
system, and support vector machines [3], [12]-[14]. Fuzzy
logic has also been combined with discrete Fourier
transform, adaptive resonance theory, principles of
estimation and independent component analysis to enhance
performance [15],16].
II. SYSTEM DESIGN
The proposed algorithm has been designed based on the

pattern recognition capability of the Artificial Neural
Network (ANN). The design has been carried out to identify
faults and find out the particular fault type if any. The phase
voltages of the three phases have been given as inputs to the
proposed neural network. The power system network has
been taken to be a 150 km long transmission line which is
segmented into 15 blocks each of 10 km line. The simulation
has been carried out in EMTP-ATP simulation software and
the data have been taken therefrom.The simulation model is
of frequency independent type so that we get a more realistic
approach to the real time situation.The entire simulation
schematic has been shown in fig. 1.
Fig. 1. Single line diagram of power system network.
Fig. 4.EMTP simulated radial power system network
A simple technique for fault type classification in a radial
system has been developed here. EMTP-ATP simulation
software has been used to simulate the transmission line
model. The faults were considered as a part of the
transmission line which has been simulated at every 10 km
in a 400 kV-150 km long three phase single circuit
transmission line. The simulation was followed by analysis of
the three phase voltage waveform for classification of fault
types using Artificial Neural Network in the MATLAB
environment. The proposed protection scheme is tested
under different fault types and varying fault locations. The
entire test results of the automatic recognition achieved by
the ANN, show that the exact fault types have been detected
and classified within half cycle after the occurrence of the
fault.
A. Radial Power System Network
To obtain the necessary information about the fault types
of a radial power system network, a 400kV 150 km long
three phase transmission line model is experimented. The
one line EMTP simulation diagram of the studied system is
shown in fig 1. The circuit is mainly designed to detect the
types of faults in the transmission line. The voltage
waveforms are monitored from the receiving end. Since we
have used a single side fed power system model, the
transmission system consists of only one three phase AC
voltage source having a rating of 400 kV 50 Hz.Fifteen LCC
three phase blocks of 10 km are taken for simulating the 150
km long overhead transmission line. The frequency
dependent ‘JMarti’ model is taken here as the basic building
block of the power system network. The line resistance is
taken as 20 Ohm/km and the fault resistance being 10 Ohm.
For the various types of faults occurred at different locations,
the faulty voltage signals only are collected and analyzed in
ATP analyzer. The sampling frequency is taken to be 2000
samples/cycle, yielding better accuracy.
III. NEURAL NETWORK CLASSIFIER
An Artificial Neural Network (ANN) can be defined as a
modern data processing system. The ANN architecture is
basically a mathematical model intended for computation of
especially a very large number of simplesby, usually a large
number of highly interconnected processing elements called
artificial neurons. These neurons can compute values from
inputs by feeding information through the network. This
ANN technique is highly useful for pattern recognition as it
can be programmed or designed to identify the identical
types of data from a very large number of information, i.e.
pattern recognition is its one of the most useful applications.
Now-a-days, this particular feature of these neural networks
is used extensively in power system analysis for fault
classification process. Usually, the voltage or current
waveforms of the system of the corresponding fault are
collected and matched with the prototypes using suitable
algorithm to obtain the correct pattern matching.
A. Probabilistic Neural Network
Probabilistic neural network (PNN) is a special category of
ANN structure differing from the conventional ANN
architecture in only a few areas.It is a kind of radial basis
network suitable especially for classification problems. The
architecture is feed forward in nature which is similar to
back propagation algorithm adopted for ANN, but differs in
the learning process, i.e. how the training is performed
before testing. Fig. 2 describes a prototype PNN model which
has been adopted in this work. Similar to the ANN
architecture, PNN also hasan input layer, an exemplar layer,
a summation layer or class layer and an output layer or
decision node.Probability Density Functions (PDFs) are
statistically estimated based on training waveform patterns
and are used as the activation function of a neuron to train
the network.The input layer represents the input pattern and

takes the feature vector as an input. Although in our
experiment, we have simply given the original sampled data
regarding the three phase voltage waveform taken at the
receiving end without any modification or feature extraction.
Next layer is called exemplar or more commonly referred to
as hidden layer. This layer consists of several nodes, usually
much higher in number compared to the input layer and it is
highly interconnected with all the nodes or inputs of the
input layer. Thus, each point of this particular layer receives
the information regarding all the inputs. This layer contains
the example vectors (the training set for the PNN), thus the
actual example vector serves as the weights as applied to the
input layer. Finally, there is an output layer consisting of
several decision-making nodes, thus representing each of the
possible cases for which the input data can be classified. In
the proposed work, the total number of possible cases is
eleven, i.e. one healthy condition and ten different types of
faults.However, the hidden layer is not fully interconnected
to the output layer. The example nodes for a given class
connect only to that class's output node and none
other.Thus, it is evident that PNN is a supervised learning
algorithm but includes no weights in its hidden layer. Instead
each hidden node here represents an example vector, which
acts as the weights to that hidden node.
Fig. 2.PNN Architecture showing different layers
B. Back Propagation Algorithm
Among the several algorithms used in neural networks,
Back Propagation has been adopted in the proposed work.
Sigmoid function, similar to the step function,is used in the
proposed work to train the neurons in between the nodes in
the hidden layer, the transfer function and characteristic
curve of which is given by σ(x) and fig. 3 as follows:
= ;
Fig. 3. Characteristic curve of Sigmoid Function used for
PNN
Hence,
If input to the neuron isξ, say and the corresponding weight
issay ω, then output O is given by,
And considering a threshold or bias term, say θ, the output
becomes as,
Minimizing the Mean Square Error (MSE) is the primary
objective of the given method. Hence, for a given set of target
data points tjand layer output Oj, theMSE is given as,
The summation is taken over all the nodes in the output
layer (K). This method calculates the rate of change of MSE
with respect to the connective weights of each respective
path, i.e. and updates the weights in consecutive
iterations in a direction of convergence of MSE towards zero.
For the output layer, this rate of change of MSE is given by:
Where, , and is the
weight from node “i” of (l-1)th layer to node “j” of the layer l
and for the hidden layer node the weight is to be calculated
from input layer to hidden layer, and rate of change of mean
square error is given can be formulated as,

Where
Even incorporating of the bias term θ, it can be observed
that,
Hence, the bias term is looked upon as output from a node
which is always equal to 1 and it holds for every layer.
Hence, it can be said that,
.
A prototype Neural Network classifier is used here for
different types of faults recognition and classification. A total
of thirteen sets of data are taken at thirteen different
locations, each separated by 10 km of the 150 km long
transmission line, excepting for the generating station
andthe load point. Each set of data contains quarter cycle
pre-fault and half cycle of post-fault sampled voltage data
corresponding to the receiving end for the three phases.Each
set consists ofone pure and ten different types of faulty
signals for the training as well as testing purpose of the
neural network for the radial power system transmission
network. Since we have used 2000 samples/sec as the
sampling rate, (1/4) cycle pre-fault and (1/2) cycle post-fault
waveform counts for a total of 1500 sample data.
C. Training of PNN
The PNN architecture constructed by connecting the
nodes layer by layer has to be trained before testing, i.e. the
neural network must be provided with some specific inputs
and the corresponding known outputs in order to make the
network understand the relationship between the input and
the output and modify the weights in each layer accordingly.
This is done so that the network becomes adaptable to
similar unknown inputs to predict or determine the
corresponding output.
In our work, the power system model simulated using the
PNN feature is trained in two ways. At first, the model is
trained using only one set of data corresponding to a
particular location in the network. This location is taken to
be at 70 km away from the sending end i.e. almost the
midpoint of the transmission line. Thus, a total of eleven sets
of signals containing ten types of faults and one healthy
signal are taken here for the training. Thus the dimension of
the training data matrix becomes 1500 × 11 and the results
were taken using the remaining twelve sets of data
corresponding to remaining locations and tested whether
the designed model could successfully identify the faults or
not. We have deliberately taken the fault location at almost
the mid-point of the transmission system to make the system
more adaptable to identify faults on both sides of the
midpoint.
In the second run, the designed PNN model is trained by
twenty two sets of signals i.e. eleven sets of data
corresponding to two different locations, viz. 70 km and 10
km from the sending end, for achieving better accuracy. In
this proposal, one set of data is taken at the midpoint as
before and the other training set has been taken almost at
the generating end, i.e. at one side of the transmission
system. The size of the training input vector for this run
becomes 1500 × 22 this time to ensure better accuracy for
faults occurring at any point all throughout the entire length.
In this case, the training is done more accurately, but at the
cost of higher processing time.
D. Target of PNN
Target of PNN is modeled here as numbers like 1, 2, 3, 4, 5,
6, 7, 8, 9, 10 and 11. Each number representany particular
type of fault, e.g. 1 represent healthy condition 2, 3, 4
represent pure, single line to ground fault for line A, B and C
respectively, 5, 6, 7 represent line to line fault in between
lines AB, BC and CA respectively, 8, 9, 10 represent double
line to ground fault in between linesAB-G, BC-G and CA-G
respectively and 11 represents three phase fault ABC.
E. Testing of PNN
Pure signal along with ten different types of faults are
considered in each set to test the simulated PNN based
protection algorithm.The three phase receiving end voltage
signals for different types of faults at
thirteendifferentlocations for the first part of the experiment
and corresponding totwelve different locations for the
second part of the experiment are used here for testing the
radial network to classify the types of faults. The PNN model
has been tested successfully as the results have matched the
target pattern.
IV. SIMULATION AND RESULT ANALYSIS
After training the neural network based fault detector and
classifier, it is tested extensively using independent data sets
consisting of fault scenarios not used previously for training
the network. The fault type, fault locations are changed to
investigate the effects of these factors on the performance of
the classifier and validate the proposed scheme.
The proposed classification system uses only the three
phase voltage signals. As has been stated before, we have

performed the training in two ways, first by taking only one
set of training data of the receiving end three phase voltages
for a fault located at 70 km from the sending end in the first
case and noted the results. In the second case, the training
data has been doubled to achieve higher accuracy of the fault
type identification. Table I and table II give the classifier
performance for each type of fault for the two types of
experiments respectively. We have also incorporated table
III here which shows the variation of the fault detection and
classification accuracy for the thirteen different fault
experiment locations, each separated by 10 km. Observation
of table III yields the fact that accuracy of the PNN is very
poor when considering the distances close to the sending
end, especially at 10 km distance the classifier cannot detect
any fault at all, and concludes it to be a healthy system. The
accuracy level increases as we move towards the receiving
end as can be seen from table IV. Hence, in the second case,
we have deliberately taken the second fault data
corresponding to 10 km from the sending end as our second
set of training data.
TABLE – I
PNN RESULTS SHOWING THE FAULT CLASSIFIER PERFORMANCE WITH
ONLY ONE SET OF TRAINING DATA TAKEN AT 70 KM FROM SENDING
END
Fault
Type
P
U
R
E
A
G
B
G
CG
A
B
BC CA
A
B
G
BC
G
CA
G
AB
C
PURE 13 0 0 0 0 0 0 0 0 0 0
AG 1 12 0 0 0 0 0 0 0 0 0
BG 1 0 12 0 0 0 0 0 0 0 0
CG 1 0 0 12 0 0 0 0 0 0 0
AB 2 0 0 0 11 0 0 0 0 0 0
BC 2 0 0 0 0 11 0 0 0 0 0
CA 1 0 0 0 0 0 12 0 0 0 0
ABG 2 0 0 0 0 0 0 11 0 0 0
BCG 1 0 0 0 0 0 0 0 12 0 0
CAG 1 0 0 0 0 0 0 0 0 12 0
ABC 1 0 0 0 0 0 1 0 0 0 11
Overall Accuracy: 90.21 %
TABLE – II
PNN RESULTS SHOWING THE FAULT CLASSIFIER PERFORMANCE
WITH TWO SETS OF TRAINING DATA TAKEN AT 70 KM AND 10 KM FROM
SENDING END
Fault
Type
P
U
R
E
A
G
B
G
CG
A
B
BC CA
A
B
G
BC
G
CA
G
AB
C
PURE 12 0 0 0 0 0 0 0 0 0 0
AG 0 12 0 0 0 0 0 0 0 0 0
BG 0 0 12 0 0 0 0 0 0 0 0
CG 0 0 0 12 0 0 0 0 0 0 0
AB 0 0 0 0 12 0 0 0 0 0 0
BC 0 0 0 0 0 12 0 0 0 0 0
CA 0 0 0 0 0 0 12 0 0 0 0
ABG 0 0 0 0 0 0 0 12 0 0 0
BCG 0 0 0 0 0 0 0 0 12 0 0
CAG 0 0 0 0 0 0 0 0 0 12 0
ABC 0 0 0 0 0 0 0 0 0 0 12
Overall Accuracy: 100 %
TABLE – III
FAULT CLASSIFIER PERFORMANCE WITH VARYING DISTANCES FOR
FIRST PHASE OF EXPERIMENT
Location of
the Fault
(km)
Total
Data
Taken
Correct
Results
Wrong
Results
% Accuracy
10 13 0 13 0
20 13 9 4 69.23
30 13 13 0 100
40 13 13 0 100
50 13 13 0 100
60 13 13 0 100
80 13 13 0 100
90 13 13 0 100
100 13 13 0 100
110 13 13 0 100
120 13 13 0 100
130 13 13 0 100
140 13 13 0 100

TABLE – IV
FAULT CLASSIFIER PERFORMANCE WITH VARYING DISTANCES FOR
SECOND PHASE OF EXPERIMENT
V. CONCLUSIONS
 The purpose of this work is to find an effective
technique suitable for the fault identification and
classification of power system overhead long
transmission lines.A simple PNN model has been
implemented here to classify the different kinds of
faultsalong with affected phases within half cycle
after the occurrence of the faults using, at best, only
two sets of receiving end voltage data at two
different locations, without considering the current
waveforms or any other parameters which may
reveal the existence of any fault, thus making the
training as well as testing convenient.
 Conventional classifier methods normally use entropy
or energy based Waveletdecompositionor other
filtering techniques to extract some key features
from the waveforms and then use the PNN model to
recognize the different fault patterns, where as in
the proposed work, the sampled voltage data is fed
directly to the PNN classifier model without any
modification of the original data, thus reducing
computational burden arising from the conventional
filtering processes and minimizing analysis
time.Hence the operation of this classifier is very
simple and here lies the novelty of the proposed
scheme.
 The results presented in this paper show that the
classifier accuracy reaches 90.21% while using only
REFERENCES
[2] Kim. K.H. et al. (2007), “Wavelet and Neuro-Fuzzy Fault
Location for Combined Transmission Systems” Electrical
Power and Energy Systems 29 pp. 445-454
[3] Jung. H et. al. (2007), “Novel Technique for Fault location
estimation on parallel transmission lines using wavelet”
Electrical Power and energy Systems 29 pp.76-82
[5] Fernando H. M. &Abur. A. (October 1998), “Fault
Location Estimation on Parallel Transmission Lines
using Wavelet” IEEE Trans. on Power Delivery, Vol. 13,
No 4.
[6] Ekici S., Yildirim S., Poyraz M., (2008) “Entropy and
energy based feature extraction for locating fault on
transmission lines by using neural network and wavelet
packet decomposition” Expert Systems with
Applications 34 2937-2944
[7] Qais H. Alsafasfeh et al., “Fault Classification and
Localization in Power Systems Using Fault Signatures
and Principal Components Analysis”, Energy and Power
Engineering, 2012, 4, 506-522.
Location of
the Fault
(km)
Total
Data
Taken
Correct
Results
Wrong
Results
% Accuracy
10 13 13 0 100
20 13 13 0 100
30 13 13 0 100
40 13 13 0 100
50 13 13 0 100
60 13 13 0 100
80 13 13 0 100
90 13 13 0 100
100 13 13 0 100
110 13 13 0 100
120 13 13 0 100
130 13 13 0 100
140 13 13 0 100 [1] S. El. Safty& A. El-Zonkoly (2009), “Applying wavelet
entropy principle in fault classification” Electrical Power
and Energy Systems 31 pp.604-607
one set of training data. In the second phase of the
experiment where receiving endvoltage data
corresponding to two different locations have been
used for training the network,i.e. the training data is
doubled using suitable locations, the accuracy
improves to as high as 100% which convey the
possibility of developing an accurate fault classifier
which can be implemented to develop a reliable
transient-based power system protection
scheme.Further investigations are being carried out
to validate the robustness and flexibility of the PNN
based proposed classifier performance under
varying conditions like changes in the network
configurations, different fault inception angles,
varying fault resistances etc. The proposed work is
also expected to be carried out to detect and classify
faults for ring main type system with more than one
source, as well as for double circuit long
transmission lines.
[4] Chunju. F et Al. (2007), “Application of Wavelet Fuzzy
Neural Network in Location Single Line to Ground Fault
(SLG) in Distribution Lines” Electric Power and Energy
Systems, 29 pp. 497-503.

[8] K. Ferreira, “Fault Location for Power Transmission
Systems Using Magnetic Field Sensing Coils,” Master
Thesis, Worcester Polytechnic Institute, Worcester, April
2007.
[9] A. Jain, A. Thoke and R. Patel, “Fault Classification of
Double Circuit Transmission Line Using Artificial Neural
Network,” International Journal of Electrical and Elec-
tronics Engineering, Vol. 1, No. 4, 2008, pp. 230-235.
[10] A. Jain and A. Thoke, “Classification of Single Line to
Ground Faults on Double Circuit Transmission Line Us-
ing ANN,” International Journal of Computer and Elec-
trical Engineering, Vol. 1, No. 2, 2009, pp. 196-203.
[11] M. Sanaye-Pasand and H. Khorashadi-Zadeh,
“Transmis- sion Line Fault Detection & Phase Selection
Using ANN,” International Conference on Power Systems
Transients— IPST, New Orleans, 2003, pp. 1-6.
[12] K. S. Swarup, N. Kamaraj and R. Rajeswari, “Fault Di-
agnosis of Parallel Transmission Lines Using Wavelet
Based ANFIS,” International Journal of Electrical and
Power Engineering, Vol. 1, No. 4, 2007, pp. 410-415
[13] J. Upendar, C. P. Gupta and G. K. Singh, “Discrete
Wavelet Transform and Genetic Algorithm Based Fault
Classification of Transmission Systems, National Power
Systems Conference (NPSC-2008),” Indian Institute of
Technology, Bombay, December 2008, pp. 223-228.
[14] S. El Safty and A. El-Zonkoly, “Applying Wavelet En-
tropy Principle in Fault Classification, International,”
Journal of Electrical Power & Energy Systems, Vol. 31, No.
10, 2009, pp. 604-607. doi:10.1016/j.ijepes.2009.06.003
[15] K. Razi, M. Hagh and G. Ahrabian, “High Accurate Fault
Classification of Power Transmission Lines Using Fuzzy
Logic,” International Power Engineering Conference IPEC,
Singapore, 3-6 December 2007, pp. 42-46.
[16] S. Vaslilc and M. kezunovic, “Fuzzy ART Neural Net-
work Algorithm for Classifying the Power System
Faults,” IEEE Transactions on Power Delivery, Vol. 20, No.
2, 2005, pp. 1306-1314
AUTHORS
Arnab Acharya has completed B.Tech from the Department
of Electrical Engineering, Techno Main, Salt Lake ( Formerly
Techno India), EM 4/1, Kolkata 700 091, India, in August,
2014.
Alok Mukherjee is an Assistant Professor of Electrical
Engineering, in Government College of Engineering &
Ceramic Technology, 73, Abinash Chandra Banerjee Lane,
Kolkata-700010, India.
Corresponding Author: Soumyadip Jana is with the
Department of Electrical Engineering, Techno Main
(Formerly Techno India), Salt Lake, EM 4/1, Kolkata 700
091, India as an Assistant Professor.

IRJET- A Simple Approach to Identify Power System Transmission Line Faults using PNN

More Related Content

What's hot (17)

Similar to IRJET- A Simple Approach to Identify Power System Transmission Line Faults using PNN (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- A Simple Approach to Identify Power System Transmission Line Faults using PNN