1

Neha Gupta, Harish Balaga, D.N. Vishwakarma

Numerical Differential Protection of Power

Transformer using GA Trained ANN
Neha Gupta, Harish Balaga, D. N. Vishwakarma
Abstract: This paper presents the use of ANN as a pattern
classifier for differential protection of power transformer,
which makes the discrimination among normal, magnetizing
inrush, over-excitation, external fault and internal fault
currents. The Back Propagation Neural Network Algorithm and
Genetic Algorithm are used to train the multi-layered feed
forward neural network and simulated results are compared.
Comparing the simulated results of the above two cases,
training the neural network by Genetic algorithm gives more
accurate results (in terms of sum square error) and also faster
results than Back Propagation Algorithm. The GA trained ANN
based differential protection scheme provides faster, accurate,
more secured and dependable results for power transformers.
Keywords: Artificial neural network, Differential protection,
Genetic algorithm, Power transformer, Pattern classification.

I.

INTRODUCTION

In power system, transformer is one of the essential elements

and thus transformer protection is of critical importance. The
most common method of transformer protection is
differential relay which involves converting the primary
current and secondary currents in a common base and
comparing them. The main challenge in differential
protection is to discriminate between an internal fault and
other operating conditions such as magnetizing inrush
current, over excitation.
Numerical differential protection has been improved
significantly since its inception in 1969 [l]. In 1988, the first
industrial application of the differential relay was introduced
[3]. A contemporary design of hardware for a digital
differential protection algorithm is given in [5]. A number of
more or less reliable protection criteria- supporting the
traditional biased differential characteristic combined with
2nd and 5th harmonic restrain have been developed in [1].
They include direct wave shape identification, protective
algorithms based on equivalent circuit composed of inverse
inductance [4]. The basic concept of apparatus protection, in
particular, protection of power transformer by means of
differential relay can be seen in many books [17]-[19].
Several algorithms have been proposed to detect winding
faults in single phase and three phase transformers and the
results show that the algorithms perform well [7].
A recursive algorithm suitable for microprocessor based
power system relaying and measurement has been proposed
in [6]. Many authors described application of Artificial
Neural Network (ANNs) in power systems [10]-[16]. In [13]
and [14], it was suggested that a feed-forward neural network
(FFNN) has been trained to discriminate between power
transformer magnetizing inrush and fault currents.
Neha Gupta, Harish Balaga and D.N. Vishwakarma are with the
Department of Electrical Engineering, Indian Institute of Technology (BHU)
Varanasi,
UP,
India
(e-mail:
[email protected],
[email protected], [email protected])

The result show that a FFNN can be consider as an

alternative method to make the discrimination between
inrush and fault currents in a digital relay implementation.
Transformer inrush is recognized on the basis of the
second harmonic components in differential currents (above
16%), which can be obtained by the various digital filters.
Other authors, of [4], determine the transformer operating
conditions on the basis of calculations of unforeseen
operating conditions on mathematical transformer models,
while the results of the calculation are continuously
compared with actual transformer operating conditions. Such
calculations are questionable in real time execution, since the
mathematical model must be significantly simplified,
resulting in the loss of information on the content of higherharmonic components in transformer current. Considering
these factors, many researchers continued their work to
develop new algorithms for transformer protection; however,
all of these algorithms are either based on the transformer
equivalent circuit model or require some transformer data,
and this may become susceptible to parameter variations.
In the recent past, numerical protection was introduced to
improve the protection scheme. As with classical protection,
new numerical protection can also lead to unnecessary
operation due to inability of protection algorithms to
distinguish transient states from faults. To overcome this
difficulty ANN based techniques are introduced to power
system protection because of their advantages like possibility
of learning from examples, generalization ability, parallel
computation, nonlinear mapping nature etc. There has been
extensive research on applications of ANNs over the last few
years, particularly in the field of pattern recognition. The
ANN-based algorithms have been successfully implemented
in many pattern or signature recognition problems. Because
this particular protection problem can also be considered as a
current waveform recognition problem, the use of ANN
seems to be a good choice. The application of ANN to
achieve some protective relay functions is a matter of
research. The ANN-based approach can detect normal,
magnetizing inrush, and over-excitation currents based on
recognizing their wave shapes, more precisely, by
differentiating their wave shapes from the fault current wave
shapes. It gives a trip signal in the case of internal fault only
and exercises restraint under healthy, magnetizing inrush and
over-excitation conditions.
But in order to obtain still more effective and fast
discrimination, Genetic Algorithm (GA) has been introduced
to train the ANN architecture.GA has been used for training
of ANN for fault classification in parallel transmission line in
[20]. The genetic algorithm is implemented with neural
network to determine automatically the suitable network
architecture and the set of parameters from a restricted region
of space in [21]. A simple real-number string is used to
codify both architectures and weights of the networks in [22].

International Journal of Emerging Trends in Electrical and Electronics (IJETEE ISSN: 2320-9569)

Vol. 3, Issue.1, May-2013.

Neha Gupta, Harish Balaga, D.N. Vishwakarma

Fig.1. Simulated power system model

Introduction of the mechanics and real world application

of GA can be seen in many books [23, 24].
In this work, an attempt has been made to design a novel
relay using ANN which can recognize all the possible
conditions, such as normal, inrush, over-excitation, external
fault, and internal fault, and give the trip signal in case of
internal fault only. The developed ANN was trained by using
feed forward back propagation algorithm and Genetic
algorithm. Network is two layered with four neurons in the
output layer and 10 neurons in hidden layer and is fed with
full cycle data. The training data are obtained using
SimPowerSystems, MATLAB software package for all of the
possible conditions, as described above. The ANN-based
algorithm has been tested to evaluate the performance of the
proposed method in terms of accuracy and speed, and the
encouraging results have been obtained.
II.

SIMULATION OF THE POWER SYSTEM TO

PREPARE THE PATTERNS

A single-phase 110/33 kV power system included a 200 km

transmission line, as shown in Fig. 1, has been used to
produce the required test and training patterns. The
simulation was done by means of SimPowerSystems
(MATLAB) software. Table 1 represents the associated data
with this power system. The combination of condition of
system is shown in Table. 2, have been produced using this
system to train the ANN. Faults are located at different points
of transmission line. Also, they involve inrush current and
over excitation condition with different voltage angles and
with different loads. As the inputs to the network are samples
taken from the waveforms generated by creating different
operating conditions, they themselves define the
TABLE I
SIMULATED POWER SYSTEM PARAMETERS

Transformer nominal
power & frequency
Transformer winding
parameters
Transformer core loss
resistance

16 MVA, 50 Hz
R=.002 pu, L=.08 pu
500 pu

Line parameters (200

km)

R=.2568 ohm/km, L=2e-3

H/km, c=8.6e-9 F/km

TABLE II
TRAINING PATTERN DATA GENERATION

Condition
system

Voltage angle
Load(MW)

Normal: when there is no fault.

Internal fault: Transformer secondary is
shorted.
External fault: faults at different
locations of transmission line.
Inrush: At different voltage angles by
closing the breaker connected
Over-excitation: At different over
voltages
0, 10, 20..to 180 degree
200, 400, 600, 800 and 1000

range to decide the operating condition. Here secondary

current is converted to primary side to make a common base.
Breakers are connected at different positions for obtaining
data for above different conditions of power transformer.
III.

DEVELOPMENT OF THE PROPOSED

NEURAL NETWORK

The developed ANN is one of the ANNs that are biologically

inspired, and uses an approach similar to human brain to
make decision and to arrive at conclusion. Each neuron in
each layer generates one normalize output, as a function of
its normalized inputs. The used network architecture is multilayered with one input layer, one hidden layer, and one
output layer. All layers are fully connected and of the feed
forward type. The outputs are nonlinear functions of the
input, and are controlled by weights that are computed during
learning process. The Multi-Layered Feed Forward Neural
Network is most commonly used in power system protection
applications.
A. Back Propagation Algorithm
The Multi-Layered Feed Forward Neural Network Algorithm
is usually trained by an algorithm known as error back
propagation algorithm. In this method the actual response of

International Journal of Emerging Trends in Electrical and Electronics (IJETEE ISSN: 2320-9569)

Vol. 3, Issue.1, May-2013.

Neha Gupta, Harish Balaga, D.N. Vishwakarma

the network is subtracted from the desired response to

produce error signal. This error signal is then propagated
backward through the network. The synaptic weights are
adjusted so as to make the actual response of the network
much closer to the desired response. This algorithm is
developed on the basis of the least square method.
B. Overview of Genetic Algorithm
GA is an optimization technique based on the principle of
genetics and natural selection. A GA allows a population
composed of many individuals to evolve under specified
selection rules to a state that maximizes the fitness.
The classical optimization methods have limitations in
searching for global optimal point.Due to the fact that GA is
a multipoint search method rather than the conventional
single point search methods.GA promises the global
optimum point to be reached.Moreover GA uses only the
value of objective function. The derivatives are not used in
the search procedure.
In GA,the design vector is represented as string of real
numbers,each number represents a variable.This string is
called a chromosome.GA starts with a group of
chromosomes known as population.
The basic operations of natural genetics-reproduction,
crossover and mutation are implemented during numerical
optimization. Reproduction is a process in which the
individuals are selected based on their fitness value relative
to that of the population. Thus individuals with higher fitness
values have a greater chance of being selected for mating and
subsequent genetic action.
After reproduction, the crossover is implemented.
Crossover is an operator that forms a new chromosome
called offspring from two parent chromosomes by
combining part of the information from each. So, first two
individuals are selected from the mating pool generated by
the reproductin operator. Next, a crossover is performed by
taking random weighted average of two parents.
The offsprings obtained from crossover are placed in the new
population The mutation is performed after crossover.
Mutation functions make small random changes in the
individuals in the population, which provide genetic diversity
and enable the genetic algorithm to search a broader space.
C. Network Architecture and Training
The first step to formulate the problem is identification
of proper input and output set. Various architectures and
combination of input sets were attempted to arrive at the final
configuration with a goal of maximum accuracy. The
simulated model parameters are given in table.1. Differential
currents are typically represented in discrete form. These
signals are sampled at a sampling rate of 16 samples per
cycle (over a data window of one cycle). These samples are
used for training and testing the developed NNs. Inputs to
ANN is selected as 16 samples of differential currents (i.e., Id
= I1 I2). One hidden layer is taken and number of neurons is
varied from 7 to 56 in order to arrive at final architecture.
After enough experimentation, one hidden layer of 10
neurons apart from 16 inputs and 4 outputs has been selected.
Transfer function for hidden layer is chosen
as tansig and for output layer it is chosen logsig for both
architectures.

Genetic Algorithm is used for finding weights and biases

of Artificial Neural Network. In this architecture 214
variables (16*10+10*4=200 weights and 14 biases) are to be
optimized using GA. The algorithm for finding weights and
biases of ANN is divided in two sections. First part is to
START

Define Fitness Function and enter no. Of parameters in the Fitness

Function
Set Stopping Criterion
Define GA parameters and create initial population
Evaluate MSE
Sort the Population according to MSE
Select pairs for mating
Perform Crossover
Mutate the Population
Re-evaluate the MSE
Sort the MSE and the associated parameters

Is Stopping criterion
Satisfied?
No
Increment iteration count

Yes

Display
Output

STOP
Fig.2. Block diagram for Genetic Algorithm

compute fitness function and second part is to generate

weights and biases of ANN using genetic operators.
Algorithm for GA
{
i=0;
Generate the initial population Pi of real coded
chromosomes Cji each representing a weight and bias set;
While the current population Pi has not converged
{
Generate fitness value Fji for each Cji Pi using the
Fitness ();
Get the mating pool ready by terminating worst fit
individuals and duplicate high;
Using the crossover mechanism, reproduce offspring
from parents chromosomes;

International Journal of Emerging Trends in Electrical and Electronics (IJETEE ISSN: 2320-9569)

Vol. 3, Issue.1, May-2013.

Neha Gupta, Harish Balaga, D.N. Vishwakarma

Mutate the population;

i=i+1;
Call the current population Pi
Calculate fitness value Fji for each Cji;
}
}
Algorithm for Fitness function
{
Let (Ii,Ti), i=1,2,..N where Ii=(I1i,I2iIli) and
Ti=(T1i,T2i,Tni) represents the input-output pairs of the
problem to be solved by ANN with configuration l-m-n.
For each chromosome Ci=1,2,p belonging to the
current population Pi whose size is p
{
Extract weights and biases from Ci
Keeping theses weights and biases setting train the ANN
for N input instances;
Calculate error Ei for each of the input instance using the
formula
Ei=Tji-Oji;
Where Oi is the output vector calculated by ANN;
Find root mean square E of the errors Ei, i=1,2,N
Calculate the fitness value Fi for each of the individual
string of the population as Fi=1/E;
}
Output Fi for each Ci=1,2.pi
}
ANN was trained for 1000 epochs. The training functions
and parameters are given in table- 3. The authors have made
use of the multicore processor functionality by using parallel
processing toolbox available in MATLAB. All 4 cores of the
Intel core I5 processor are made to run in parallel during the
training process. This significantly improved the speed of
training.
Since the networks have to distinguish among five kinds of
signals, 5 sets of example signals (cases) have been obtained
for this purpose. These cases are normal, magnetizing inrush,
over- excitation, internal fault currents and external faults
currents. Following figure -3 show the differential line
currents of three phase transformer at different conditions.
Signal is sampled at 800 Hz (16 samples per cycle).The input
data set for the neural network is organized in the form of a
moving data window with a fixed window length of a full
cycles each input set has 16 samples.
165 training samples generated by SimPowerSystems in
MATLAB, have been used to train and test the neural
network. These training sets of samples contain 5 different
conditions of power transformer (normal, magnetizing
inrush, over-excitation, internal fault conditions and external
fault conditions). For each case, signals are sampled at the
sampling rate of 16 samples per cycle (over a data window of
one cycle).These samples are arranged in form of full cycle
data to feed as input to the ANN. The outputs of the networks
have a unique set (e.g., 0000 = normal, 1000= inrush,
0100=over excitation, 0010=internal fault and 0001 =external
fault). This network i.e. with 4 outputs monitors

No. of Variables
Population Size
Scaling Function
Selection Function
Mutation Function
Mutation Rate
Crossover Function
Generations
Function Tolerance

TABLE III
GA PARAMETERS

214
1000
Rank
Roulette Wheel
Uniform
0.01
Heuristic
1000
1e-16

Fig.3(a). Normal condition

Fig.3(b). Inrush condition

Fig.3(c). Over excitation condition

Fig.3(d). Internal fault condition

Fig.3(e). External fault condition

all the conditions occurring in the power transformer and it

issues the trip signal only in case of fault condition.For
training the network, Back Propagation and Genetic
Algorithms were used and results were compared. After
training ANN model was tested with all possible sets of data
under different operating conditions of power transformer by
using both Back Propagation and Genetic Algorithm.

International Journal of Emerging Trends in Electrical and Electronics (IJETEE ISSN: 2320-9569)

Vol. 3, Issue.1, May-2013.

Neha Gupta, Harish Balaga, D.N. Vishwakarma

Fig.4. Actual Architecture of ANN

IV.

NETWORK PERFORMANCE AND

NUMERICAL RESULTS

The Architecture of developed ANN is shown in figure 4.

While training with Back Propagation the mean square error
obtained is 0.0158 (Training). The following figure 5 shows
the mean square error during training. The tested outputs of
architecture using Back Propagation are given in Table 4.
But, while the same architecture is trained using Genetic
Algorithm, the mean square error is reduced to 3.73e-8. The
learning error graph over 1000 generation is shown in the
figure 6. The tested outputs of architecture using Genetic
Algorithm are given in Table 5 .

Fig.6. Learning error over 1000 generations.

It can be concluded from results that when ANN was trained

using BP, it was not able to distinguish between normal and
external fault conditions. It gives 0.3890 instead of 0 under
normal condition and 0.4270 instead of 1 under external
fault condition at the 4th output neuron. But when same
architecture is trained using GA the mean square error is
reduced to 10-8 and produced more nearer output to 0 or 1
depending on the condition. So it can be seen that when
ANN is trained using GA, weight adjustment is enhanced.

Best Validation Performance is 0.0071626 at epoch 63

10

Train
Validation
Test
Best

)
e
s
-1
10
m
(
r
o
rr
E
d
e
ra
u 10-2
q
S
n
a
e
M
-3

10

50

100

150

200

250

263 Epochs

Fig.5 Learning error of ANN for while training with BP

TABLE IV
OUTPUT OF ANN USING BACK PROPAGATION

Normal

Inrush

Over excitation

Internal Fault

External Fault

0.0018

0.9950

0.0012

1.27e-05

0.0017

0.0011

5.37e-07

0.9976

3.36e-08

0.0011

0.0010

2.43e-05

0.0001

0.9921

0.0010

0.3890

0.0005

0.0003

0.0057

0.4270

International Journal of Emerging Trends in Electrical and Electronics (IJETEE ISSN: 2320-9569)

Vol. 3, Issue.1, May-2013.

Neha Gupta, Harish Balaga, D.N. Vishwakarma

TABLE V
OUTPUT FOR ANN USING GA

Normal

Inrush

Over excitation

Internal Fault

External Fault

8.78e-08

0.9999

3.56e-08

1.28e-11

4.06e-12

3.06e-12

8.20e-09

0.9999

8.15e-08

5.61e-09

5.21e-11

7.41e-15

2.71e-10

0.9999

6.91e-08

1.19e-07

1.37e-13

4.16e-08

9.90e-09

0.9999

V.

CONCLUSION

In this paper a feed forward NN model for protection of

power transformer have been analyzed. The 16_10_4
architecture could correctly discriminate among the different
conditions in power transformer such as normal, magnetizing
inrush, over-excitation, internal fault and external fault. The
ANN has been trained for all the possible sets of simulated
data under different operating conditions of transformer.
ANN has been trained using Genetic Algorithm and results
are compared with back propagation training algorithm. It
was inferred that back propagation algorithm minimizes an
average sum squared error term by doing a gradient descent
in the error space. But there is a possibility of getting struck
or oscillating around a local minimum. So the effectiveness
of this method is not satisfactory. In addition, the
convergence of the back propagation algorithms during
training is sensitive to the initial values of weights. If the
initial values of the weights are not properly selected, the
optimization will be trapped in a local minimum or
maximum.
Thus the GA trained ANN is very efficient in
solving detection and classification problem. Use of parallel
processing toolbox of MATLAB is suggested to improve the
training speed of GA. The GA trained ANN based
differential relaying for power transformer shows promising
security, accuracy and speed.
VI.
[1]
[2]

[3]

[4]

[5]

[6]

[7]
[8]
[9]

[10]
[11]

[12]

[13]
[14]
[15]

REFERENCES

A.J. Degens, Microprocessor Implemented Digital Filters

for Inrush Current Detection, Electrical Power and Energy
Systems, vol.4,no.3, pp.196-205, July 1982.
H. K. Verma and A. M. Bosh, A microprocessor-based
inrush restrained differential relay for transformer protection,
Journal of Microcomputer Applications, no. 9, pp-313-318,
1986
Michel Habib, Miguel A. Marin, A comparative Analysis of
Digital Relay in Algorithms for the differential protection of 3
phase Transformers, IEEE Transactions on power systems,
Vol.3, No.3 pp.1376-1384 August 1988
Keizo Inagaki, Masaru Higaki, Yoshiaki Matsui and Kentaro
Kurita, Digital Protection Method for Power Transformer
Based on an Equivalent Circuit Composed of Inverse
Inductance, IEEE Transactions on Power Delivery, vol.3,
no.4, pp.1501, Oct 1988
J. C. Maim, Hardware and software design of a digital
protection relay for power transformers, Conference on
Computer Relaying-Future Directions and Impacts, pp. 1- 19,
1988.
M. s. Sachdev, and M. Nagpal, A Recursive Least Error
Squares Algorithm for Power system Relaying and

[16]
[17]
[18]
[19]
[20]

[21]
[22]

Measurement Application, IEEE Trans. On Power Delivery,

vol.6, no.3, pp-1008-1015, July 1991.
Limin Fu. Neural Network in Computer Intelligence.
McGrew-Hill, Inc, New Delhi 1994
SimPowerSystems and MATLAB ( 7.0.1) Simulink help
manual Guide.
T.S. Sidhu, M.S. Sachdev, H.C. Wood, M. Nagpal, Design
Implementation and Testing of Microprocessor Based High
Speed Relay for Detecting Transformer Winding Faults,
IEEE Tran. On Power Delivery, vol.7, no.1, pp.108-117,
January 1992.
J.Pihler, D. Dolinar, Improved Operation of Power
Transformer Protection Using Artificial Neural Network,
IEEE Transctions on Power Delivery, vol.12, no.3, July 1997.
BKasztenny
and
E.Rosolowski,
Multi-Objective
Optimization of a Neural Network based Differential Relay
for Power Transformers IEEE conference on Transmission
and Distribution, 1999, pp 476 - 481 vol.2.
M.R. Zaman and M.A. Rahman , Experimental Testing of
the Artificial Neural Network Based Protection of Power
Transformers, IEEE Transactions on Power Delivery,
vol.13, no.2, April 1998
Zahra Moravej, D. N. Vishwakarma and S. P. Singh, ANN
Based protection Scheme for Power Transformer, Electric
Machine and Power Systems, 28:875-884, 2000.
Zahra Moravej, and D. N. Vishwakarma, ANN Based
Harmonic Restraint Differential Protection of Power
Transformer, IE(I) Journal-EL, Vol 84, June 2003
Harish Balaga, D. N. Vishwakarma, Amrita Sinha,
Numerical Differential Protection of Power Transformer
using ANN as a Pattern Classifier, International Conference
on Power, Control and Embedded Systems (ICPCES2010),IEEE, Nov 28-Dec 1, 2010.
Enio C. Segatto and Denish V. Coury A Differential Relay
For Power Transformers Using Intelligent Tools, IEEE
Transactions on power system, vol. 21, No. 3, August 2006
B. Ram and D. N. Vishwakarms, Power System Protection
and Switch Gear, Tata Mcgrew-Hill Publishing Company
Limited, New Delhi,1995
C. L. Wadhwa, Electrical Power Systems, New Age
International(P) Limited, Publishers
T.S. Madhava Rao, Power System Protection Static Relays
With Microprocessor Applications, Tata Mcgrew-Hill
Publishing Company Limited New Delhi,1989.
R.Rajeswari, N.Kamaraj, Application of Genetic Algorithm
in
Artificial Neural Network for fault classification in
parallel transmission lines, IET-UK International
Conference on Information and Communication Technology
in Electrical Sciences (CTES2007).
YasAbba Alsultanny, MusbahM.Aqel , Pattern recognition
using multilayerneural-genetic algorithmElsevier Science
B.V. May 2002.
Daniel Rivero, Julian Dorado, Enrique Fernndez-Blanco and
Alejandro Pazos, A Genetic Algorithm for ANN Design,

International Journal of Emerging Trends in Electrical and Electronics (IJETEE ISSN: 2320-9569)

Vol. 3, Issue.1, May-2013.

Neha Gupta, Harish Balaga, D.N. Vishwakarma

7

Training and Simplification,IWANN 2009, Part I, LNCS

5517, pp. 391398, 2009.
[23] Randy L. Haupt, Sue Ellen Haupt, Practical Genetic
Algorithms, Wiley-Interscience, a John Wiley & Sons Inc.
Publication.
[24] Olympia Roeva, Real-world applications of Genetic
Algorithms,Published by InTech 2012.

Neha Gupta is currently pursuing

M.Tech at the Indian Institute of
Technology (BHU) Varanasi, India.
She obtained her B.E. (Electrical
Engineering) from Jai Narain Vyas
University, Jodhpur.

Devendra Nath Vishwakarma is

currently Professor of Electrical
Engineering at the Indian Institute of
Technology (BHU) Varanasi, India.
He obtained his B.Sc. (Engineering),
M.Sc. (Engineering) and Ph.D. from
Patna University, Patna. He had earlier
served as associate professor of electrical engineering at
the Bihar College of Engineering, Patna (Presently
National Institute of Technology Patna). He has a
teaching and research experience of over 35 years and
has contributed about 65 papers to various national and
international journals and conferences. He is a senior
member of IEEE (USA), a fellow of the Institution of
Engineers (India) and a fellow of the Institution of
Electronics and Telecommunication Engineers.

Harish Balaga is currently pursuing

Ph.D at the Indian Institute of
Technology (BHU) Varanasi, India. He
obtained his B.E. (Electronics and
Instrumentation Engineering) from
Andhra University and M.Tech
(Control Systems) from the National Institute of
Technology Kurukshetra.

International Journal of Emerging Trends in Electrical and Electronics (IJETEE ISSN: 2320-9569)

Vol. 3, Issue.1, May-2013.