ISSN 0146-4116, Automatic Control and Computer Sciences, 2020, Vol. 54, No. 1, pp. 39–51.

© Allerton Press, Inc., 2020.

Fault Location in the Transmission Network

Using Artificial Neural Network
M. Dashtdara, *, M. Esmaeilbeiga, M. Najafia, and M. Esa Nezhad Bushehrib
a
Electrical Engineering Department, Faculty of Engineering, Bushehr Branch, Islamic Azad University, Bushehr, Iran
b
Computer Engineering Department, Bushehr Branch, Islamic Azad University, Bushehr, Iran
*e-mail: dashtdar.masoud@gmail.com
Received May 24, 2019; revised October 5, 2019; accepted October 21, 2019

Abstract—In this paper, in order to locate the fault in the transmission network, a discrete wavelet
transform is used to extract the fault characteristics from the zero-sequence current, in order to train
the artificial neural network. In fact, the basis of the work is based on the information recorded after
the fault at the beginning and at the end of the line, received by the relay. In the following, with the
help of Fortescue’s transform, the current of zero sequence seen from both terminals is calculated and
by the transform of the wavelet of stored information at high frequency is extracted in the horizontal
components of the zero sequence current from both terminals, and finally calculating the energy
stored in horizontal components, as well as extracting the maximum scales of horizontal components
can reveal certain features of the fault that are suitable for training the neural network. Simulation
results show that the maximum scales of horizontal components and the energy stored in these com-
ponents strongly depend on the fault resistance, type of fault, fault angle and fault location. Therefore,
the training data should be selected in such a way that these changes are well represented so that the
neural network does not encounter problem in its diagnosis. Finally, the proposed method has been
tested on a transmission network of 735 kV at different distances of the transmission line. And results
indicate that the proposed algorithm can estimate fault distance depending on the type of fault in dif-
ferent conditions.

Keywords: fault location, zero-sequence current, wavelet transform, WEE, ANN

DOI: 10.3103/S0146411620010022

1. INTRODUCTION
The ultimate goal of a power system is the continuous transmission of energy to consumers. Todays,
due to the strong dependence of industries and urban consumers on electrical energy, electricity discon-
nection will cause heavy economic damages to industries and also disruptions in the daily-life of users. On
the other hand, there may be also several accidental and unpredictable faults in the total power network;
which by minimizing the detection time of fault location, the damages caused by power loss can be
decreased. Various methods have been proposed so far to locate the fault in order to accelerate the network
repair and improve the reliability [1]. Further, in addition to fault location methods, some methods are
proposed to determine the section or fault distance, especially in distribution networks [2]. Fault location
techniques can be categorized firstly to the impedance methods, traveling waves, components of high-fre-
quency and fault voltage, and secondly, training-based intelligent methods. Traveling wave methods have
a complex structure and require the peripheral equipment [5]. On the other hand, high-frequency com-
ponent methods will cost a high cost due to the use of high-rate sampling filters [6]. Impedance methods
have a relatively simple structure and their use in protection relay is easily possible. As a result, the most
common way of finding a fault is the impedance method. These methods use the phasor components of
voltage and current frequency to locate the fault [7]. Impedance methods are divided into two categories
of single-terminal and two-terminal methods. In single-terminal methods, only measured data (at the ter-
minal of which the relay is installed) are used, which is since of the simplicity of equations. The precision
of these methods depends on the size of fault resistance, the homogeneity of the line (impedance con-
stancy per unit length of the line) and the identification of fault type [8]. In the two-terminal method, the
measured data are used at both local and remote terminals. This method will be more precise than single-
terminal methods because of independency from the fault resistance and type of fault [9–11]. Along with
the existing methods, Training-based algorithms can also be used as an alternative to locating the faults in

39
40 DASHTDAR et al.

transmission lines. Training-based strategies can represent an acceptable level of flexibility and perfor-
mance in various conditions and against the disruptions in the system, if executed properly. Extracting
efficient features and applying an appropriate training algorithm are two main and influential issues in the
formation of training-based methods [2, 3]. One of the most commonly used functions in signal process-
ing is the wavelet transform that allows the extraction of high and low frequency components of signal,
which can also play an important role in fault location. For instance, by applying the wavelet transform on
high frequency samples of the traveling fault signal (which is recorded at the beginning and end points of
the medium-pressure feeders), the exact location of the fault can be determined. It was first shown in [13]
that by recording the fault at the two ends of a simple transmission line, the precise location of the fault
can be determined by wavelet transform through traveling waves method, and regardless of the load along
the line. In [4] and [17], a method is proposed that uses a wavelet transform to determine the fault location
using the traveling waves in the power system; which makes possible the conversion of a high-resolution
wavelet for high-frequency components of fault transmissions. Further, it is possible to determine the
location of several faults by converting the three-phase voltage signals into the modal components, and
performing the wavelet transform on the modal signals from the partial component of the regional mode
at high frequencies [4]. In the current paper, the wavelet transform and artificial neural network are used
to fault location in the transmission line. It should also be noted that in traveling waves methods, registers
are required at high cost to record the signals sent from the fault location; but in the proposed scheme, by
easily transforming the wavelet and extraction of a series of specific features of the obtained horizontal
components which is only based on the current data recorded by the relay and by training the neural net-
work, one can estimate the exact location of the fault. In the second section, the proposed method is
described based on the wavelet transform and the neural network; while the network implementation and
fault locator algorithm are presented in the third section. In Section 4, the results of this approach are pre-
sented, and finally, the conclusions are presented in the fifth section.

2. SUGGESTED METHOD
In the following paper, the basis of method is based on the analysis of the current of the zero-sequence
obtained from the information recorded after the fault at the beginning and end of the line, received by
the relay. Thus, in the first step, by Fortescue transform of three-phase current recorded by the relay, the
zero-sequence current is obtained at both terminals, according to Fig. 1.
And in order to extract a series of specific characteristics of the fault, we need to obtain the wavelet
transform from zero sequence current at both terminals, which will result in four components (approxi-
mate, horizontal, vertical and diagonal) for each decomposition level. In current research, we focus on the
horizontal components of the 2nd level of signal decomposition. In the next step, maximum scales of hor-
izontal components should be extracted on both sides of the line; and by calculating the energy stored in
each horizontal component we have four input data to the neural network, consisting of two energy stored
in the horizontal components and two maximum scales of horizontal component on both sides of the line.
And since these training data are strongly dependent on the resistance changes, angle, and type of fault;
thus, the data should be selected in an intelligent way so that these changes can be traced to the neural
network, making the network to recognize and not get into trouble.
The asymmetric faults of transmission lines such as short-circuits, impedance between lines, the
equivalent impedance of one or two lines with ground, or the interruptions of conductors of a line can be
analyzed using the Fortescue method. Fortescue’s method proves that each n-phase unbalanced system
can be divided into n balanced phase systems. According to this, any unbalanced three-phase system can
be divided into three balanced three-phase systems as follows:
• The system of positive sequence components consists of three phases of vector with a phase differ-
ence of 120 degrees and the phase arrangement similar to the main unbalanced phases.
• The system of negative sequence components consists of three phases of vector with a phase differ-
ence of 120 degrees and the phase arrangement opposite of the main unbalanced phases.
• The system of zero sequence components consists of three phases of vector that do not have any
phase differences.
To perform this transform, the matrix of symmetric components is used, which is usually represented
by relation (1). Where, we focus only on the components of the zero sequence in this study.

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020

 FAULT LOCATION IN THE TRANSMISSION NETWORK 41

Zero sequence current B1 (A-g fault) Zero sequence current B2 (A-g fault)
60 60
40 40
20 20
0 0
–20 X = 50 km –20 X = 50 km
Rf = 0.5 ohm Rf = 0.5 ohm
–40 L = 100 km –40 L = 100 km

0 1000 2000 3000 4000 0 1000 2000 3000 4000

Post-fault Post-fault

Bus 1 Bus 2
Source
Transmission Circuit Three
Circuit phase
Breaker 1 Breaker 2 load
Fault

Fig. 1. The zero-sequence current seen from the both sides of line after the fault in the sample network.

I 0  1 1 1  I a 
 I  = 1 1 α α2  ⋅ I  . (1)
 1 3    b
I 2  1 α α   I c 
2
 
In order to extract the characteristics of fault, simultaneous processing was done on the three-phase
current received by the relay, having the advantage of reducing the training data for different angles of fault
[2, 3]; but since of the simultaneous processing of three phases, the rate of produced training data is high. In
this paper, the transformation process of three-phase current into the zero-sequence current is used to reduce
the amount of data and improve the algorithm accuracy by processing zero sequence current. Thus, the atten-
tion is on the stream of zero sequence on the both sides of line in this scheme; i.e. I0(B1) and I0(B2).

3. WAVELET TRANSFORM—WAVELET ENERGY ENTROPY

Wavelet transform is able to display the signals by a technique of fenestration areas with different sizes,
taking advantage of using the major time-windows in case of need the accurate information at low fre-
quencies and also use of smaller time windows when the need for data of high-frequency. One of the major
improvements made by wavelet transform is the Local Analysis of signal, which makes it possible to ana-
lyze a small area of wide signal. In fact, the wavelet is a form of a wave with limited time and zero effective
value. The wavelets are arbitrary and asymmetric, and their energy is concentrated, limited, and around a
point. A wavelet transform involves breaking a signal into transmitted signals and changing the scale of the
original wavelet signal (Mother Wavelet), and therefore analyzing signals with sharp changes by means of
wavelets is better than the Fourier analysis with sinusoidal waves. It is concluded that local variations are
made better with wavelets. The wavelet transform has a characteristic called Mother Wavelet with different
types. This transform initially separates the waveform that is more similar to the Mother Wavelet among
the all waves, and then arranges the separated wave based on the types of frequency components (from
high to low), so that they can be combined all to obtain the initial wave. As an example, the separation of
the waves forming the signal in the zero-sequence current from terminal 1 is shown in Fig. 2 with one and
two levels of decomposition using db4, respectively. As one can see, two components of approximate and
detail (including horizontal, vertical, diagonal) are obtained at each decomposition level, which are
related to the low frequency and high frequency information of signal, respectively. Therefore, the wavelet
transform can be used to make reveal the hidden information of fault, where we pay more attention to the
horizontal component of the 2nd level of decomposition (H2) in the following study.

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020

42 DASHTDAR et al.

Zero sequence current B1 Approximation A1 Approximation A2

50 100 200
Post-Fault
Pre-Fault
50 0
–200
0 1000 2000 3000 4000 0 500 1000 1500 2000 0 200 400 600 800 100
A-g Fault Horizontal H1 Horizontal H2
0.5 2
0 0
–0.5 –2
0 500 1000 1500 2000 0 200 400 600 800 100
100 u10 2 u10
–14 –14 Vertical V2
Vertical V1
50 0
0 –2
0 500 1000 1500 2000 0 200 400 600 800 100
1 u10 Diagonal D1 1 u10
–14 –14 Diagonal D2

0 0
–1 –1
0 500 1000 1500 2000 0 200 400 600 800 100
Level 1 Level 2

Fig. 2. Two-level decomposition of the zero-sequence current seen from the terminal 1.

The wavelet transform is a linear transformation that maintains the time allocation in the various fre-
quency components of the given signal. Then, we can write the relation (2) by considering a function of
scale x and a wavelet function of ψ:
k − n0 b0 am 
DWTψ f ( m.k ) = 1  n x ( n) ψ  , (2)
m
a0  a0m 
in which, ψ is the Mother Wavelet, a0m is the scale parameter; am , b0 and n0 are transmission parameters.
In addition, the db4 (daubechies-4) mother wavelet is employed in this research. Using the following
command, one can implement the wavelet transform process from the zero-sequence current, in the
MATLAB software:
[A1, H1, V1, D1] = dwt2(I0,'db4'); % level-signal decomposition
[A2, H2, V2, D2] = dwt2(A1,'db4'); % level-Two decomposition, where in;
A1, A2: Approximate component of level one and two of signal decomposition. H1, H2: Horizontal com-
ponent of level one and two signal decomposition. V1, V2: Vertical component of the level one and two of
signal decomposition. D1, D2: Diagonal component of the level one and two of signal decomposition.
I0: zero sequence current. db4: Mother wavelet. Therefore, according to Fig. 2 one can write the current
of zero sequence in the form of sum of components in (3):
I 0 = A1 + ( H1 + V1 + D1 ) = ( A2 + H 2 + V2 + D2 ) + ( H1 + V1 + D1 ) . (3)
The feature of the wavelet’s detail component is the existence of high frequency information in it, and
since high frequency components are created when the fault occurs, then in this scheme we pay more
attention to the horizontal component of level two of zero-sequence current decomposition. Now the
question is, what features can be extracted from the horizontal components so that the neural network can
be trained well in order to prevent problems in its diagnosis. One of these ways is the use of wavelet energy
entropy to calculate the energy of horizontal components. Meanwhile, surveys have shown that the peak
(or maximum scale) of horizontal components varies from one fault to another. The wavelet energy is
obtained from the sum of detail coefficients and its value varies from one stream to another. In this study,
the wavelet energy of horizontal component of zero-sequence current on the scale j and the moment k is
obtained in the form of (4): [14–16]
E jk = H j ( k ) .
2
(4)

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020

 FAULT LOCATION IN THE TRANSMISSION NETWORK 43

Horizontal H2
2.0
Maximum scale
1.5
1.0
0.5
0
–0.5
–1.0
–1.5 WEE calculation

–2.0
0 100 200 300 400 500 600 700 800 900 1000

Fig. 3. The horizontal component of level-two decomposition of zero-sequence current.

A signal can be evaluated in a variety of ways. An easy algorithm for optimal signal decomposition is
entropy. An entropy shows the amount of information stored in the signal. Assume that on the scale j, with
k = 1, 2, 3,…, N; N is the number of moments (coefficients) on the scale j and L is the number of decom-
position levels. In sum, the wavelet energy spectrum of the horizontal component on the j scale can be
written as (5):
N
Ej = E
k =1
jk , j = 1, L. (5)

The relative wavelet energy is given in [15] is represented in the form of relation (6), which represents
the distribution of energy:
E jk
Pjk = , j = 1, L. (6)
Ej
Therefore, the wavelet energy entropy (WEE) of the horizontal component of zero-sequence current
can be obtained as (7): [15, 16]
WEE j 0 = −Σ k Pjk 0logPjk 0 , j = 1, L. (7)
Thus, in accordance with Fig. 3, we extract two properties of the horizontal component in this scheme:
(1) Maximum scale of horizontal component; (2) calculate the wavelet energy entropy (WEE).

3.1. Design of Artificial Neural Network

The artificial neural network is a modern computing method for machine learning, displaying knowl-
edge, and, finally, for applying the obtained knowledge to estimate the output responses from complex
systems, which its feature is creating a new structure for information processing system. An artificial neu-
ral network consists of three layers of input, output and hidden. Each layer contains a group of neurons
that are generally associated with all other neurons in the other layers, unless the user restricts the com-
munication between the neurons, but the neurons of each layer are not related to other neurons of the
same layer. In fact, the basis of the neural network is based on training and learning. Therefore, in order
to train the neural network in the field of fault location, it should be possible to extract a series of fault
properties so that the neural network correctly detects the fault and location of the fault in the system. As
mentioned in the previous section, two features are used in this paper: the maximum scale of horizontal
component and the wavelet energy entropy of the horizontal component as the input of neural network.
In the [2], in order to extract the fault characteristics, the simultaneous processing was carried out on three
phases of the current. While in this research, the processing is performed only on the zero-sequence cur-
rent seen from both sides of the transmission line, in order to reduce the number of training data.
In Fig. 4, the proposed artificial neural network structure is shown. As indicated, the input of neural
network consists of two entropies of wavelet energy and two maximum scales of horizontal components

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020

44 DASHTDAR et al.

tansig(n) purelin(n) purelin(n)

WEE0B 1
1 1
Max(H2B1) 2 1 1 X
2 2
WEE0B 3 2 2 Rf

4 N N
Max(H2B2)

Input Hidden layers Output

Fig. 4. Structure of artificial neural network.

obtained from the two sides of line. And since the results indicate a strong dependence of neural network
inputs on the fault resistance, thus we assign the fault resistance (Rf) in the output of neural network in
this plan to increase the accuracy of algorithm, in addition to fault location (X). Meanwhile, there are
three hidden layers in this network, which the size of first and second layers can be changed and the size
of third layer is equal to the output of neural network, due to the existence of two outputs in neural network
that restricts the size changing of third layer; thus, by placing the first and second layers, we can easily
increase the accuracy of algorithm by resizing these layers, where in this scheme, the size of first layer is
12, and the second and third layer are considered 8 and 2, respectively. It should also be noted that we used
a neural network with three hidden layers to locate the faults in this scheme, instead of using three separate
neural networks for single-phase, two-phase and three-phase faults, where in:
WEE0B1: wavelet energy entropy of horizontal component in level-two decomposition of zero-
sequence current seen from terminal 1.
Max(H2B1): maximum scale of horizontal component in level-two decomposition of zero-sequence
current seen from terminal 1.
WEE0B2: wavelet energy entropy of horizontal component in level-two decomposition of zero-
sequence current seen from terminal 2.
Max(H2B2): maximum scale of horizontal component in level-two decomposition of zero-sequence
current seen from terminal 2.

4. NETWORK IMPLEMENTATION
In the current study, the proposed algorithm is implemented on a 735 kV transmission network with a
frequency of 60 Hz, a 0° source phase angle, an end-to-end load of 30,000 MV A with a –28992-degree
phase angle. Meanwhile, the length of the line is assumed 100 km and all models are done with MATLAB
software. Table 1 presents information on the structure of artificial neural network. And finally, the full
flowchart of suggested location finder algorithm is shown in Fig. 5.

5. SIMULATION RESULTS
Initially, in order to extract and produce training data, we have investigated the transmission line of
Figure (1) in the event of several faults. The results of studies show that the maximum scale of horizontal
components and WEE are different at the beginning and end of the line, and are highly dependent on the
fault resistance and the fault type. Which here, the results of studies on Terminals 1 and 2 are presented
briefly in Figs. 6 to 11.
In Figs. 6 to 8 the WEE variations, and in Figs. 9 to 11 the variations of maximum scale of horizontal
component for several faults are shown. The WEE variations and the maximum scale are represented
respectively in Figs. 6, 7, 9, and 10 for fault resistance of 0.5 to 125 Ohms at 25 km fault distance from the
line; and in Figs. 8 and 11, the WEE variations and the maximum scale are represented respectively for
fault distances from 15 to 85 km. Meanwhile, the angle of a, b and c phases is –40, –160 and –280 degrees
at the time of fault. It was mentioned in [2, 3] that if the processes are performed simultaneously on a

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020

 FAULT LOCATION IN THE TRANSMISSION NETWORK 45

Table 1. Parameters of artificial neural network

Parameters ANN
Input 4
Output 2
Training algorithm function Traingdm
Hidden layer 1 function Tansig(n)
Hidden layer 2 function Purelin(n)
Hidden layer 3 function Purelin(n)
Size of hidden layer 1 12
Size of hidden layer 2 8
Size of hidden layer 3 2
Weight update method Batch-mode
Train parameter goal 1e-4
Learning rate 0.05
Performance function MSE (mean squared error)

three-phase current, due to the 120 degrees difference between phases, the symmetry of phases entropy is
not changed for different angles of fault, thus there is no need to extract the training data per various
angles. In addition, due to the simultaneous processing of the three-phase current in this scheme and its
conversion into zero-sequence current, it is practically not necessary to calculate the entropy for different
angles, and only three phase angles are sufficient at the time of fault, provided that simultaneous process-
ing on three phases will take place.
As can be seen from the results, variations in WEE and maximum scale of horizontal components are strongly
dependent to the fault phase, fault type, fault distance, and the value of fault resistance. In Figs. 6 and 9, training
data changes for single phase faults are shown. As one can see, in this type of fault, the highest value of
horizontal component occurs in the B-g fault, whose angle is –160 degrees; but the highest WEE is cre-
ated in the A-g fault, whose angle is –40 degrees. And the reason can be obtained by observing Fig. 12,
which shows the variation of surface under the current curve in the event of a fault occurring in different
angles. Figure 12 shows that the WEE calculated from the angle of –40 degrees will have the highest value.
But this is not always the case, because the distance and the fault resistance can reduce the WEE, such
that in the C-g fault, the calculated WEE obtained from the base 1 is less than of base 2. In Figs. 7 and 10,
training data changes are shown for faults in two phases and three phases to the ground. In this case, as
one can see, the training data has the highest value for the three-phase fault to the ground, which is the

Recording the 3-phase current after fault in bus 1 and 2

Calculating the zero-sequence current seen at bus 1 and 2

Two-level decomposition of the zero-sequence current with db4 wavelet

Extracting the horizontal component seen from bus 1 and 2

Extracting the maximum scale and calculating the WEE of horizontal components

Design and training of artificial neural network

Estimation of resistance and distance of fault

Fig. 5. Flowchart of proposed algorithm to fault location.

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020

46 DASHTDAR et al.

L = 100 km, x = 25 km, L = 100 km, x = 25 km,

Single-phase-g fault, BUS 1 Single-phase-g fault, BUS 2
WEE0(A-g) WEE0(A-g)
4.5 4.5 WEE0(B-g)
WEE0(B-g)
WEE0(C-g) WEE0(C-g)
4.0 4.0

WEE 0 (B2)
WEE 0 (B1)
3.5 3.5
3.0 3.0
2.5 2.5
2.0 2.0
0 20 40 60 80 100125 0 20 40 60 80 100125
Rf, Ohm Rf, Ohm

Fig. 6. WEE for single-phase fault to the ground.

L = 100 km, x = 25 km, L = 100 km, x = 25 km,

Two-three-phase-g fault, BUS 1 Two-three-phase-g fault, BUS 2

4.5 WEE0(A-B-g) 4.5

WEE0(B-C-g)
4.0 WEE0(C-A-g)
4.0
WEE 0 (B1)

WEE0(A-B-C-g)
WEE 0 (B2)
3.5 3.5
3.0 3.0
WEE0(A-B-g)
2.5 WEE0(B-C-g)
WEE0(C-A-g)
WEE0(A-B-C-g)
2.0
0 20 40 60 80 100125 0 20 40 60 80 100125
Rf, Ohm Rf, Ohm

Fig. 7. WEE for two-phase and three-phase fault to the ground.

L = 100 km L = 100 km
Two-three-phase fault, BUS 1 Two-three-phase fault, BUS 2

4.5 4.5 WEE0(A-B)

WEE0(B-C)
WEE0(C-A)
WEE 0 (B1)

WEE 0 (B1)

4.0 4.0 WEE0(A-B-C)

3.5 3.5
WEE0(A-B)
3.0 WEE0(B-C) 3.0
WEE0(C-A)
WEE0(A-B-C)
20 40 60 8090 20 40 60 8090
X, km X, km

Fig. 8. WEE for two-phase and three-phase fault.

most severe type of fault. In Figs. 8 and 11, training data changes are shown for two-phase and three-phase
faults. In this case, the training data of three-phase fault is the highest, as in the previous state. Meanwhile,
the comparison of training data in two-phase and three-phase faults with two-phase and three-phase
faults to the ground shows the effect of resistance on data, so that the fault resistance reduces training data.
Therefore, in order to generate training data, the type of fault, fault location and range of resistances
should be selected so that the data change can be accurately presented to the neural network; and the neural
network can also be able to design an appropriate algorithm for fault detection, through the transform functions
presented in Table 1. Thus, using the information of Table 2, the training data presented in Figs. 6 to 11 can be
generated and using the transform functions shown in Table 1, a relationship is created between input and

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020

 FAULT LOCATION IN THE TRANSMISSION NETWORK 47

L = 100 km, x = 25 km, L = 100 km, x = 25 km,

Single-phase-g fault, BUS 1 Single-phase-g fault, BUS 2

Maximum scale (B1)

Maximum scale (B2)

3.0 3.0 Maximum
scale(A-g)
Maximum scale(A-g)
2.5 Maximum scale(B-g)
2.5 Maximum
scale(B-g)
Maximum scale(C-g) Maximum
2.0 2.0 scale(C-g)
1.5 1.5
1.0 1.0
0.5 0.5
0 20 40 60 80100125 0 20 40 60 80100125
Rf, Ohm Rf, Ohm

Fig. 9. Maximum scale of horizontal component for single-phase fault to the ground.

L = 100 km, x = 25 km, L = 100 km, x = 25 km,

Two-three-phase-g fault, BUS 1 Two-three-phase-g fault, BUS 2
4.0 4.0
Maximum scale (B1)

Maximum scale(A-B-g)

Maximum scale (B1)

3.5 Maximum scale(B-C-g) 3.5
Maximum scale(C-A-g)
3.0 Maximum scale(A-B-C-g) 3.0
2.5 2.5
2.0 2.0
1.5 1.5
1.0 1.0
0.5 0.5
0 20 40 60 80100125 0 20 40 60 80100125
Rf, Ohm Rf, Ohm

Fig. 10. Maximum scale of horizontal component for 2-phase and 3-phase fault to the ground.

L = 100 km L = 100 km
Two-three-phase fault, BUS 1 Two-three-phase fault, BUS 1
3.5 3.5
Maximum scale (B1)

Maximum scale(A-B)
Maximum scale (B1)

3.0 Maximum scale(B-C) 3.0

Maximum scale(C-A)
2.5 Maximum scale(A-B-C) 2.5
2.0 2.0
1.5 1.5
1.0 1.0
0.5 0.5
20 40 60 8090 20 40 60 8090
X, km X, km

Fig. 11. Maximum scale of horizontal component for two-phase and three-phase fault.

output to train the neural network. In Table 3, the generated data is presented in order to train the neural
network for various faults.
As you can see in Table 3, the training data for faults to the ground is considered greater, due to the
fault resistance. Finally, the total data provided in this plan are 146 items, which are suitable for a trans-
mission line for various faults and only using an artificial neural network. In the following, the data is pre-
sented for training on the neural network. And the neural network performance results show its high accu-
racy in fault location.

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020

48 DASHTDAR et al.
Occur of fault at angle – 40 (A-g Fault) Occur of fault at angle – 160 (B-g Fault) Occur of fault at angle – 280 (C-g Fault)

–160 X = 25 km
50 50 50 –280
Rf = 20 ohm
0 0 0
X = 25 km X = 25 km
–50 –40 Rf = 20 ohm –50 –50 Rf = 20 ohm

1000 1500 2000 2500 3000 1000 1500 2000 2500 3000 1000 1500 2000 2500 3000
Current A Current B Current C
30 30 30
20 20 20
10 10 10
0 0 0
–10 –10 –10
1000 1500 2000 2500 3000 1000 1500 2000 2500 3000 1000 1500 2000 2500 3000
Zero sequence current B1 (A-g Fault) Zero sequence current B1 (B-g Fault) Zero sequence current B1 (C-g Fault)

Fig. 12. Changes in the area under the three-phase current curve and zero-sequence current during fault occurrence for
angles of –40, –160, and –280 degree.

In MATLAB software, there are various diagrams to examine and demonstrate the performance of
artificial neural network. Figure 13 shows the performance diagram of neural network. This diagram
shows the number of epochs and also the mean squared error (MSE) of the network, which in this scheme
has reached a value of 9.874e-0.05 at 1544 epochs. In Fig. 14, the Training State diagram of network is
presented. In this chart, we observe the errors. The operation mode is such that if the “val fail” term does
not improve 6 times continuously and not ascending, the training process stops, which was in improve
mode in this plan, and the faults reached 0.0001 in 1544 epochs.
Figure 15 shows the network’s regression graph. This diagram draws on the output of neural network
and the real value, and if these two coincide, it indicates the good quality of neural network. In this
scheme, the output of neural network with a good approximation of 1*Target – 1e-5 is close to real value.
In Fig. 16, the confusion matrix of network is presented. In the confusion matrix, the rows are related to
the Output Class and the columns are for the Class Target. The diagonal cells are related to the observa-
tions that are properly categorized. And non-diagonal cells are classified as incorrect observations. Both
the number of observations and the percentage of total observations are shown in each cell. The right col-
umn of graph shows the percentage of all predicted samples belonging to each class that are correctly and
incorrectly categorized. These criteria are often referred to as precision (or positive predictive value) and
discovery rate, respectively. The end rows of graph represent the percentage of all instances that are clas-
sified for each class correctly and incorrectly. These criteria are often referred to as a call (or real positive
rate) and a negative false rate. The right cell bottom of the graph shows the overall accuracy of neural net-
work. In this scheme, the overall accuracy of network was 98.6% and only two errors were made, which in
these two cases only made errors in the estimation of fault resistance, and accurately predicted the fault
location. In the following, after training of neural network, the proposed algorithm is tested for various
faults at different intervals of transmission line, the results of which are shown in Table 4. Plus, the error

Table 2. Parameters considered for generating training data

Parameters affecting the fault

Resistance, Angle, Fault Type

Location, km
Ohm degree 1 Ph.g 2 Ph.g 2 Ph 3 Ph.g 3 Ph
15 0.5 –40 A.g A.B.g A.B A.B.C.g A.B.C
25 20 –160 B.g B.C.g B.C
50 50 –280 C.g C.A.g C.A
75 75
85 100
125

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020

 FAULT LOCATION IN THE TRANSMISSION NETWORK 49

Table 3. The data generated for training the neural network

Fault type Number of training data = NlocationNresistanceNfault
1 Ph.g 3 × 6 × 3 = 54
2 Ph.g 3 × 6 × 3 = 54
3 Ph.g 3 × 6 × 1 = 18
2 Ph 5 × 3 = 15
3 Ph 5×1=5
Total 146

percentage of estimation can be calculated from equation 8, the results of which are given in Table 4. In
addition, results in Table 4 as compared with the prevalent impedance method.
Actual location − Estimated location
RE = × 100%. (8)
Length of line

Best training performance is 9.874e-05 at epoch 1544

104
Train
Best
Mean squared error, mse

Goal
102

100

10–2

10–4

0 500 1000 1500

1544 Epochs

Fig. 13. Performance diagram of neural network.

Gradient = 0.6565, at epoch 1544

105
Gradient

100
10–5
105 Mu = 0.001, at epoch 1544
Mu

100
10–5
Validation checks = 0, at epoch 1544
1
Val fail

0
–1
0 500 1000 1500
1544 epoch

Fig. 14. Training state diagram of neural network.

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020

50 DASHTDAR et al.

Regression: R = 1
120 Data
Fit

Output ~= 1*Target + 1e-05

Y=T
100

80

60

40

20

0 20 40 60 80 100 120
Target
Fig. 15. Regression diagram of neural network.

Confusion matrix

47 0 100%
1 32.2% 0% 0%
Output class

2 97 98.0%
2 1.4% 64.4% 2.0%

95.9% 100% 98.6%

4.1% 0% 1.4%

1 2
Target class
Fig. 16. Confusion matrix of neural network.

6. CONCLUSIONS
In this paper, signal processing techniques and artificial neural networks are used to fault location in
the transmission network. In the first, the data recorded by the relay are transformed by Fortescue method
and the zero-sequence current is obtained, and then the training data are extracted by transforming the

Table 4. Results of the proposed algorithm test

Actual values Estimated values Impedance
RE, % RE, %
fault type Rf, Ohm X, km Xe, km Rf, Ohm method
A-g 0.4 40 40.0543 0.3981 0.0543 40.1252 0. 1252
A-B-g 7 20 20.1315 6.8639 0.1315 20. 2696 0.2696
A-B-C – 50 50.0147 – 0.0147 50. 1181 0.1181
C-g 85 65 65.0605 84. 9200 0.0605 66. 0120 1.0120
A-C-g 100 80 80.0354 99.9410 0.0354 81. 3021 1.3021
B-C – 10 9.9601 – 0.0399 10.1312 0.1312
B-g 75 40 40.1666 75.0341 0.1666 41.1321 1.1321
B-C-g 150 80 79.9487 157.1231 0.0513 81.5329 1.5329
A-C – 30 29.8389 – 0.1611 30.1023 0.1023
A-B – 90 90.0972 – 0.0972 89.9754 0.0246
A-g 15 15 15.1298 14.9964 0.1298 15.1387 0.1387
A-C-g 95 90 90.0623 95.0576 0.0623 91.2715 1.2715
A-B-C-g 20 35 35.1207 20.4321 0.1207 35.8312 0.8312

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020

 FAULT LOCATION IN THE TRANSMISSION NETWORK 51

wavelet from zero sequence current and calculating the WEE and the maximum scale of horizontal com-
ponents of zero sequence current. Further, studies show that training data is highly dependent on type,
resistance and location of fault. Therefore, in this project, the training data is selected on purpose, so that
the fault characteristics can be accurately mapped to the neural network.
In addition, instead of using three separate neural networks for single-phase, two-phase and three-
phase faults, an artificial neural network with three hidden layers is used in this algorithm, and designed
in such a way that for different faults with different resistances, the performance is good and is able to esti-
mate the fault resistance in addition to the location. The simulation results show that the maximum esti-
mate error in this design is 0.1666 percent.
CONFLICT OF INTEREST
The authors declare that they have no conflicts of interest.
ADDITIONAL INFORMATION
Masoud Dashtdar was born in Bushehr, Iran, on September 20, 1986. He received the M.Sc. in power electrical engi-
neering of Islamic Azad University of Bushehr, Iran, in 2013. he is currently pursuing the Ph.D. degree with Islamic Azad
University, Bushehr, Iran. His research interests are distribution system, include fault diagnosis in power systems, neural
network computing, power electronics, Electric machines, renewable energy, and harmonic analysis.

REFERENCES
1. De Andrade, L., De Leão, and Ponce, T., Impedance based fault location analysis for transmission lines, Trans-
mission and Distribution Conference and Exposition (T&D), IEEE PES, 2012.
2. Masoud Dashtdar, Rahman Dashti, and Hamid Reza Shaker, Distribution network fault section identification
and fault location using artificial neural network, 2018 5th International Conference on Electrical and Electronic
Engineering (ICEEE), 2018.
3. Masoud Dashtdar, Fault location in distribution network based on fault current analysis using artificial neural
network, J. Electr. Comput. Eng., 2018, vol. 1, no. 2, pp. 18–32.
4. Majid Dashtdar and Masoud Dashtdar, Fault location in the transmission network using a discrete wavelet
transform, Am. J. Electr. Comput. Eng., 2019, vol. 3, no. 1, pp. 30–37.
5. Lopes, F.V., Kusel, B.F., and Silva, K.M., Traveling wave-based fault location on half-wavelength transmission
lines, IEEE Latin Am. Trans., 2016, vol. 14, no. 1.
6. Bo, Z.Q., Weller, G., and Redfern, M.A., Accurate fault location technique for distribution system using fault-gen-
erated high-frequency transient voltage signals, IEE Proc. Gener. Transm. Distrib., 1999, vol. 146, no. 1, pp. 73–79.
7. Rao, A. and Bogale, B., Accurate fault location technique on power transmission lines with use of phasor mea-
surements, Int. J. Eng. Res. Technol., 2015, vol. 4, no. 2.
8. Xu, Z. and Zhang, Z., What accuracy can we expect from the single-ended fault locator?, 2015 68th Annual Con-
ference for Protective Relay Engineers, 2015.
9. Venugopal, M. and Tiwari, C., A novel algorithm to determine fault location in a transmission line using PMU
measurements, 2013 IEEE International Conference on Smart Instrumentation, Measurement and Applications
(ICSIMA), 2013.
10. Elkalashy, N.I., Kawady, T.A., Khater, W.M., and Taalab, A.M., Unsynchronized fault-location technique for
double-circuit transmission systems independent of line parameters, IEEE Trans. Power Delivery, 2015.
11. Gaurav Kapoor, A fault-location evaluation method of a 330 kV three-phase transmission line by using discrete
wavelet transform, Int. J. Eng. Des. Anal., 2018, vol. 1, no. 1, pp. 5–10.
12. Magnago, F.H. and Abur, A., Fault location using wavelet, IEEE Trans. Power Delivery, 1998, vol. 13, no. 4,
pp. 1475–1480.
13. Samantaray, S.R., Panigrahi, B.K., and Dash, P.K., High impedance fault detection in power distribution net-
works using time-frequency transform and probabilistic neural network, IET Gener. Transm. Distrib., 2008,
vol. 2, no. 2, pp. 261–270.
14. Zhengyou, H., Ling, F., Sheng, L., and Zhiqian, B., Fault detection and classification in EHV transmission line
based on wavelet singular entropy, IEEE Trans. Power Delivery, 2010, vol. 25, pp. 2156–2163.
15. Ekici, S., Yildirim, S., and Poyraz, M., Energy and entropy-based feature extraction for locating fault on transmission
lines by using neural network and wavelet packet decomposition, Expert Syst. Appl., 2008, vol. 34, pp. 2937–2944.
16. Majid Dashtdar and Masoud Dashtdar, Fault location in the transmission network based on the analysis of the record-
ed current by the relay using a discrete wavelet transform, Am. J. Electr. Comput. Eng., 2019, vol. 3, no. 1, pp. 30–37.
17. Dashtdar, M. and Dashtdar, M., Fault location in distribution network based on phaser measurement units
(PMU), The Scientific Bulletin of Electrical Engineering Faculty, 2019, vol. 19, no. 2, pp. 38–43.
18. Dashtdar, M. and Dashtdar, M., Detecting the fault section in the distribution network with distributed gener-
ators based on optimal placement of smart meters, The Scientific Bulletin of Electrical Engineering Faculty, 2019,
vol. 19, no. 2, pp. 28–34.
19. Dashtdar, M. and Dashtdar, M., Fault location in the transmission network based on extraction of fault com-
ponents using wavelet transform, The Scientific Bulletin of Electrical Engineering Faculty, 2019, vol. 19, no. 2,
pp. 1–9.

AUTOMATIC CONTROL AND COMPUTER SCIENCES Vol. 54 No. 1 2020