Electr Eng (2012) 94:125–134

DOI 10.1007/s00202-011-0218-2

ORIGINAL PAPER

An alternative approach to fault location on power distribution

feeders with embedded remote-end power generation
using artificial neural networks
Yilmaz Aslan

Received: 7 February 2008 / Accepted: 5 June 2011 / Published online: 13 September 2011
© Springer-Verlag 2011

Abstract In this paper, the design and implementation of patrols are relied upon, particularly, on relatively long lines
a feed-forward artificial neural network (ANN)-based fault and when the visibility is very poor because of tree growth,
locator to classify and locate shunt faults on primary over- etc. Also, the locators can help in the case that maintenance
head power distribution lines with load taps and embedded is shared by many companies or divisions within a company.
remote-end power generation is presented. In the ANN algo- Moreover, the weak spots that are not obvious may be found
rithm, the standard back-propagation technique with a sig- and a more focused inspection can be performed within a
moid activation function is used. The fault locator utilizes limited area defined by the fault locator [1].
fault voltage and current samples obtained at a single loca- Although, so far, the majority of the distribution sys-
tion of a typical radial distribution system. The ANNs are tems comprise feeders with remote ends open, due to the
trained with data under a wide variety of fault conditions and privatisation of the electricity supply industries and also
used for the fault type classification and fault location on the increasing environmental and global concerns, more and
distribution line. A 34.5 kV distribution system is simulated more non-conventional energy sources such as wind energy,
using electro-magnetic transients program and their results biogas, solar and small hydro, etc. are being added to the
are used to train and test the ANNs. The ANN-based fault existing distribution systems. This trend is likely to continue
locator gives high accuracy for the vast majority of the prac- in the near future [2]. An important problem due to increase
tically encountered systems and fault conditions, including in complexity of the distribution system arises from the fact
the presence of load taps and the remote-end in-feed source. that any fault will also have remote in-feed. It is well known
that a remote-end in-feed can adversely affect the accuracy of
Keywords Fault location · Distribution lines · Artificial the conventional fault locators. A number of the impedance-
neural networks · Embedded generation based fault location techniques for the single-ended overhead
distribution systems have been proposed in [3–5]. In such
techniques, the effects of pre-loading the presence of remote-
1 Introduction end source and errors originating from the interface and the
quantization are not taken into consideration. The technique
In recent years, as a result of the competition with the pri- developed in [6] is based on superimposing the components
vatisation of the power distribution companies, it has been which is very efficient in distribution lines with remote-end
become essential to provide the customers a high-quality ser- generation, but it requires exact knowledge of the feeder
vice without any outages. Consequently, it is needed for these configuration and the load data. A technique based on the
companies to locate permanent and transient faults on distri- travelling waves and high-frequency (HF) components pre-
bution feeders as quickly as possible. The importance of the sented in [7,8] requires specifically tuned filters, and its initial
fast fault locators is more obvious for the cases where foot costs would be prohibitively high in longer radial distribu-
tion lines. In Refs. [9,10] the knowledge-based approaches
Y. Aslan (B)
which often require external information such as substation
Department of Electrical and Electronics Engineering, Faculty of
Engineering, Dumlupinar University, 43100 Kutahya, Turkey and feeder switch status, feeder measurements, load voltage
e-mail: [email protected] and current sensors etc. are presented. The techniques devel-

123
126 Electr Eng (2012) 94:125–134

oped in [11–16] require a dedicated communication medium. Threshold

θk
In addition, the recorded fault data gathered from the ends x1 wk1
of the transmission system are needed to be synchronised. Activation
x2 function
Although these techniques offer very accurate results, they wk2
Input uk Output
are not tested for an overhead distribution system and also signals Σ ϕ (.) yk
their initial investment and operational costs are too high. In
Summing
Ref. [17], an intelligent two-port numerical algorithm for the xn wkn junction
fault location, the adaptive autoreclosure, the detailed dis- Synaptic
weights
turbance record analysis and the fault data management is
presented. The proposed algorithm does not require a syn- Fig. 1 A basic structure of a neuron
chronized sampling of the data taken from the line terminals
and also the fault arc is included in the complete fault model.
This algorithm, in contrast to the approaches that do not take methods and algorithms, ANNs can learn the key informa-
the fault arc into consideration, gives a high accuracy in fault tion from a multi-dimensional database. Moreover, since the
location and has the ability to determine both the arc and ANNs can process the missing and noisy data, they have
the fault/tower-footing resistance. But, the algorithm has not higher error tolerances. Another outstanding advantage of
been tested for a practical low voltage overhead distribution the ANN is that once trained they can produce fast results
feeder with load taps, yet. for nonlinear problems [18–20]. As seen in Fig. 1, the math-
This paper introduces a new approach based on ANNs ematical model of a basic ANN neuron has a much simpler
to accurately locate shunt faults on primary overhead power structure comparing with a biological neuron. Basically, a
distribution systems with load taps and also in the presence neuron k may be mathematically described with the Eqs. (1–
of small-scale distributed generation system connected at the 3) [21]:
 n 
remote-end of the network. The latter has a particular impor- 
tance regarding the fact that the future distribution networks uk = wki xi + θk (1)
will increasingly become active (i.e. bi-directional power i=1
flow) due to the penetration of the renewable energy-based and
generation systems. Our technique is based on utilising the
fault voltage and current samples obtained at a single location yk = ϕ(u k ) (2)
of a typical radial distribution line. In order to attain a high The data received from outer environment (or other neurons)
degree of accuracy in the location of a fault, the post-fault xn are transferred to the neuron k through weights wkn which
currents and voltages obtained from one end of the distri- adjust the effect of inputs. The neuron calculates a weighted
bution system are filtered using Discrete Fourier Transform average (u k ) using the summation function and then uses
(DFT) techniques to extract the required power frequency some activation function ϕ to produce an output yk [21].
components. The extracted voltage and current phasors are The activation function is used to express the nonlinear rela-
then fed to the ANN which has been trained off-line with tionship process between the input and the output data. The
the data under a wide variety of fault conditions to classify selection of the activation function may vary depending on
and locate the faults on the distribution line. The practical the problem studied. In general, one of the activation func-
limitations originating from the hardware such as interface tions of step, sign, sigmoid or linear is used. In this study, the
and quantization errors are taken into account in the design log-sigmoid which is a nonlinear continuous function from 0
process. The proposed algorithm is practically tested in a to 1 and differentiable everywhere for other x values is pre-
34.5 kV overhead distribution system through the simula- ferred as an activation function. The mathematical expression
tions using the electro-magnetic transients program (EMTP). of the log-sigmoid function is given in Eq. (3) [20]:
The accuracy of the technique is evaluated for the fault type,
1
the location and the resistance, the presence of remote-end ϕ(x) = (3)
in-feed, the fault inception angle and the fault cycle. The sim- 1 + e−x
ulation results clearly show that the proposed fault location Figure 2 depicts the architecture of a typical feed-forward
technique presented here is highly accurate. multilayered neural network which consists of an input layer,
hidden layers (one or more) and an output layer. In the net-
2 ANNs work, xn shows the inputs from outer environment to the first
layer; w kn and wmk are the weights applied to the second
Artificial neural networks are widely accepted and recogni- and the third layers, respectively. The number of the hid-
sed method for the solution of the complex and ill-defined den layers and the number of neurons in the layers are sub-
problems. Instead of using complicated mathematical ject to the problem studied and decided upon the trial-error.

123
Electr Eng (2012) 94:125–134 127

34.5 kV Distribution Line

w1,1 1 w1,1
x1 1 w2 ,1 CT VT
w1,n w1,2 1 y1 End P
Substation
w1,2 w 2 w1,k
2 ,2
Currents Voltages
x2 2 Interface modules (input
w2 ,n wm ,2
wm ,1 transformers and low-pass filters)

wk ,1 Sample and hold + multiplexer

wk ,2 A/D Conversion
wm ,k m ym
xn n
Digital filtering and phasor
wk ,n k Outputs extraction (DFT)
Inputs
Hidden Layer
ANN for fault type classification
Fig. 2 A feed-forward multi layer ANN

A B C ABC
The input layer receives the signal from outer environment ANN1
a-phase
ANN2
b-phase
ANN3
c-phase
ANN10
3-phase
and distributes it to the neurons in the hidden layer. The num-
ber of hidden layers having computational neurons depends Fig. 3 The complete fault location algorithm
on the functions to be used. Since each additional layer
increases the computing load exponentially, 3-layered ANNs
are usually preferred in practice [18,19].
ted. The computer scans the voltage and the current samples,
The training stages in the ANNs comprise both the feed-
storing the data and the oldest information being overwritten.
forward and back-propagation networks. After the training
is completed, only the feed-forward algorithm is used in nor-
mal operation. Therefore, the training stage of the ANN is
3.1 Fault inception time identification
more time consuming when it is compared with the utiliza-
tion stage.
Before applying the fault location algorithm, any changes in
stored current and voltage samples should be identified. In
the digital fault recorder (DFR), after the digitisation stage,
3 ANN-based fault location algorithm the microprocessor continuously executes a monitoring rou-
tine. In this process, the current and voltage samples taken
The complete fault locator scheme is shown in Fig. 3. The from the near end of the distribution line are measured and
three-phase currents and voltages enter through the input stored in the RAM memory of the computer. In the presence
transformers which provide a galvanic isolation from the of a fault, the current and the voltage waveforms are distorted
instrument transformers as well as transforming the signals to and therefore, their magnitude and phase angle may change
±10 V reference voltage level. The input signals taken from with respect to the pre-fault conditions. To identify the fault
voltage transformers (VTs) and current transformers (CTs) inception time, an adaptive approach is used and the first
may contain HF components under fault conditions. In order three samples of the second cycle are compared with the
to prevent aliasing, second-order Butterworth filters with a corresponding three samples of the previous cycle for the
cut-off frequency of 1.5 kHz are used. The filter outputs are current waveform. Any significant change more than a pre-
switched in sequence by the multiplexer and fed into the sam- defined threshold level indicates the time at which the fault
ple and hold circuit in preparation for digital conversion. The has occurred. If these criteria are not satisfied for the cur-
analogue to digital conversion is achieved via a 12-bit A/D rent samples, then the same process is applied to the voltage
converter and a sampling frequency of 4 kHz (80 sample win- samples. Upon the inception of a fault, the fault recorder is
dow) is used throughout the process. The ±211 conversion triggered and the captured fundamental values of currents
process leads to a quantisation level of approximately 4.8 mV. and voltages provide three cycles (or till the circuit breaker
Analogue to digital conversion introduces further errors due operates) of the fault data. In the development of the DFR, the
to quantisation. After the digitisation stage, the voltage and important aspects of the practical fault recorders such as VT
current data is acquired and stored in a circular buffer in the and CT responses, analogue interface effects and quantisa-
RAM memory before being printed, processed or transmit- tion errors are taken into account. This is made to ensure that

123
128 Electr Eng (2012) 94:125–134

(a) Since the DFT can be evaluated for any particular frequency
40
component, it is more preferred in protection system appli-
30
cations [22]. The DFT is very efficient in rejecting the HF
20
Voltage (kV)

components and thus effectively attenuating the DC offset. It

10
is applied to digitised voltage and current samples based on
0
-10
0 0.01 0.02 0.03 0.04 one cycle information which gives both the magnitude and
-20
the phase of the fundamental phasor X v,i (ω) as [23,24]:
-30   n=N−1
2
-40
Time (s)
X v,i (ω) =
N
n=0
 
(b) × xv,i (n) {cos(ωnt) − j sin(ωnt)} (4)
2
In Eq. (4), N is the number of the samples per cycle, t is
1
the sampling time, ω is the frequency of the phasor to be
Current (kA)

0
extracted and xv,i (n) represents the sampled voltage or cur-
0 0.01 0.02 0.03 0.04 rent waveforms. The phasors extracted from the DFT filter
-1 are used as the input to the fault type classifying ANN as
shown in Fig. 3. When the classification of the fault type is
-2 achieved, to have a good generalization, the separate ANNs
Time (s)
are used in accurately locating the types of the faults on the
Fig. 4 ‘a’-phase-earth fault. a The three-phase voltages. b The three- distribution system [25].
phase currents

V, I
4 The practical considerations
34.5 kV Distribution Feeder
10 km 10 km 10 km 10 km
End P End Q
As mentioned previously, the fault location technique in this
34.5 kV Embedded work is based on utilising the voltage and the current phasors
Substation Remote Source
100 MVA SCL Load 1 Load 2 Load 3 (if present) at the fault locator end of the line. Therefore, the training data
is determined applying the following steps:
Fig. 5 A 34.5 kV distribution system

Step 1: The simulation of the distribution system with

the performance obtained is considerably close to a real-life the EMTP to generate faulted voltage and current
situation. waveforms.
Step 2: The extraction of voltage and current phasors using
3.2 The extraction of voltage and current phasors a DFT filter.
Step 3: The application of the phasors to the designed best
Figure 4 shows the voltage and current waveforms for an ANN topology.
‘a’-phase-earth fault at 3 km away from end P in the system Step 4: The calculation of the error between the output vec-
shown in Fig. 5. As can be seen in the waveforms, after the tor and the desired output and adjusting the weights
fault, the voltage waveforms are distorted due to HF compo- to reduce the error.
nents while DC off-set is more prominent in current wave- Step 5: Repeating the process until the error criteria is
forms. satisfied.
In order to achieve a high degree of accuracy, in the fault
location algorithm, after the A/D conversion it is vitally In the testing stage, the weights are used to find the output
important to extract voltage and current phasors of the power values corresponding to the input values. In this stage, the
frequency component from the post-fault waveforms which errors are not propagated backward (from the output to the
can contain transients ranging from high frequencies down input) since it is not possible to compare the output data
to the DC levels. Although the Fast Fourier Transform (FFT) set with the target data set. A computer program written in
is superior to DFT in terms of computing performance, the Delphi is used to implement the ANN [26]. In the program,
computing time is relatively long since it calculates the entire a three-layered ANN model with feed-forward and BP algo-
frequency spectrum. In power protection systems, usually, rithms was used and some certain variables such as the learn-
only the power frequency (50 Hz) component is of interest. ing rate, momentum and number of neurons in the layers can

123
Electr Eng (2012) 94:125–134 129

Table 1 The tested ANN parameters 4.1 The simulation of a practical overhead power system
Neurons in the Number of Learning rate (η) Momentum (α)
hidden layer iterations The practical distribution system studied in this work is a
34.5 kV, 50 Hz overhead system with three phase loads that
10,000, 35,000, 0.1 0.2
each rated as 2 MVA with a power factor of 0.95 (lagging),
k 50,000, 100,000 0.3 0.5
tapped off at various locations as shown in Fig. 5. The over-
0.5 0.6 head distribution lines employed in this work are aluminium–
alloy conductors with no earth wires, based on horizontal
line configuration currently used in the Turkish power dis-
Table 2 The output logic of the fault type classification ANN tribution system. The overhead power distribution system is
Fault type A B C G simulated with the EMTP software and the line elements are
considered as distributed. The relevant data used are:
No fault 0 0 0 0
‘a’-phase earth fault 1 0 0 1 1. Earth resistivity (assumed homogeneous) = 100 m.
‘a’-‘b’-phase fault 1 1 0 0 2. Source X/R ratio = 10; Z s0 /Z s1 = 0.5
‘a’-‘b’-phase-earth fault 1 1 0 1 3. 34.5 kV feeder impedance = (0.18 + j0.34) /km.
Three phase fault 1 1 1 0
4.2 The training data for ANNs

w1,1 1 In the development stage of an ANN-based algorithm, it is

w1,1
very important to train and test the network. The training
x1 1 w2 ,1 w2 ,1
w1,12 1 A data are obtained from the simulation of different types of
w1,2 faults at various points of a typical distribution system both
w1,2 w2 ,2 w2 ,2 with and without any remote-end source as shown in Fig. 5.
2
x2 2 B
The desired outputs of the ANN for fault type classification
2 w4 ,2
w2 ,12 w3,1 are defined by variables A, B, C and G. A value of unity
w1,k 3 C for any of the first three variables corresponds to the ‘a’, ‘b’
wk ,1 w2 ,k or ‘c’ phases being faulty and a value of G near unity indi-
w4 ,1
cates that the ground is involved in a fault. Table 2 shows
wk ,2 w3 ,k 4 G the output logic of the fault type classification ANN. In the
x12 12 developed algorithm, the voltage and current phasors are used
wk ,12 w4 ,k Outputs
k as the training data representing the different fault and sys-
Inputs
Hidden Layer tem conditions. The performance of the ANN was tested
using the both patterns within and outside the training set.
Fig. 6 The ANN for the fault type classification For a better accuracy and performance of the algorithm, the
ANN was trained for the fault type classification; the separate
networks are used for locating the faults on the distribution
be changed. In the development stage of the ANN model, system. The desired output of the fault locating ANNs was
the proposed model has been tested using different parame- defined as a specific point on the line in km [25]. Figure 6
ters and number of iterations to find the best structure and to shows the structure of the ANN for the fault type classifica-
minimise the errors as shown in Table 1. tion with four outputs and Fig. 7 shows one of the ANNs for
In the ANNs, although the estimation errors are reduced the fault location which has one output. After the training, the
for a higher number of iterations, this results in a longer networks were tested with a separate set of test data unseen
training time. In the development stage of the ANN model, by the ANNs before.
it is seen after 35,000 iterations that the error is negligible
and does not vary significantly. In this study, the number
of neurons (k) in the hidden layer was varied as 8, 10, 12 5 Results
and 14 neurons and the best results were achieved for 12
neurons in the hidden layer. Finally, a network with 12 neu- The effectiveness of the developed technique was tested
rons in the hidden layer and with log-sigmoid activation for different fault locations and types, fault resistance, fault
function was used. In the training stage, the learning rate inception angle and fault cycle. In all of the results presented
(η) and the momentum (α) were set to 0.5 and 0.6, respec- here, the percentage error relating to fault location is based
tively. on the following equation:

123
130 Electr Eng (2012) 94:125–134

w1,1 1 faults created at 30 km on the system operated as radial (with

x1 1 w1,2 no embedded source). These errors may be explained by the
w2 ,1 w1,12 w1,1
presence of laterals which lead to different fault currents for
w2 ,2 2 w1,2 the faults occurring at different locations of a radial distri-
x2 2 bution feeder. It is evident from the results shown in Table 4
w2 ,12 1
that the performance of the fault location algorithm, com-
Output pared with the radial feeder, is not affected by the presence
wk ,1 w1,k
of the remote-end source.
wk ,2
x12 wk ,12
The algorithm is also tested for the faults closer to the line
12
k ends. Various faults were created at 2.5 and 37.5 km on the
Inputs radial operating system and with the 10 MVA source con-
Hidden Layer nected to the end of the system and the results are shown in
Fig. 7 The ANN for the fault location Table 4. As can be seen, the results have not been signifi-
cantly affected from the location and therefore the error has
remained as less than 2% for all of the faults studied.

actual location − desired location

error (%) = × 100 (5) 5.2 The effect of the fault resistance
length of the line
In order to show the effectiveness of the proposed technique, The ANN was trained for the phase-earth faults for the fault
the training and the test data were generated using the 34.5 kV resistance of 2, 5, 10, 15, 20, 30, 40, 50, 60, 80 and 100  in
distribution system shown in Fig. 5. At every 2.5 km, the var- the system shown in Fig. 5. The effect of the fault resistance
ious faults were created and 80% of these cases were used on accuracy was tested on the same distribution system for
for the training of the network and 20% were used for the the unseen cases and the results are given in Table 5. As seen
subsequent testing. Table 3 shows the test results for the fault from the table, the fault locator gives satisfactory results in
type classification. It is evident from the results that the ANN the presence of fault resistance for the faults created at various
can accurately classify faults by giving the values around ‘1’ locations. From the results it is observed that in contrast to
and ‘2’ since, in practice, the fluctuations are always exist the analytical methods the proposed method does not suffer
[25]. from a high value of fault resistance (R f ). The algorithm was
also tested for various fault resistances for an ‘a’-phase-earth
5.1 The effects of the fault type and the location fault created at 30 km for the radial system (no remote-end
in-feed) and the results are given in Fig. 8. As seen from the
Table 4 also classifies the effect of fault type for the faults results, although the accuracy is varied with different fault
in a 34.5 kV system shown in Fig. 5. The types of faults resistances it had a variation of less than 3 % for the fault
investigated are (AG) ‘a’-phase-earth fault, (BG) ‘b’-phase- resistances studied.
earth fault, (CG) ‘c’-phase-earth fault, (AB) ‘a’-‘b’-phases In the overhead power distribution lines, the presence of
fault, (ABG) ‘a’-‘b’-phases-earth fault, (ABCG) ‘a’-‘b’-‘c’- the fault arc resistance which is a nonlinear phenomenon
phases-earth fault. From the results, an increase in the error introduces HF transients particularly in the voltage wave-
due to the phase-earth faults occurring near any tapping point forms during the shunt faults. In essence, these enhance the
can be observed. In these faulty conditions, the algorithm has magnitudes of the transients present due to travelling waves.
yielded to a maximum error smaller than 3% for phase-earth Since the proposed technique is based on the extraction of

Table 3 The results of the fault

classification Fault type Desired output Actual output

A B C G A B C G

AG 1 0 0 1 1.008658 −0.00597 −0.00347 0.997666

BG 0 1 0 1 −0.00049 0.997298 0.001566 1.008502
CG 0 0 1 1 0.005824 −0.00332 1.00085 1.000501
AB 1 1 0 0 1.001126 0.994477 0.004847 −0.00224
ABG 1 1 0 1 1.001272 0.99184 −0.00062 1.000538
ABCG 1 1 1 1 1.004145 0.932686 1.010033 1.055676

123
Electr Eng (2012) 94:125–134 131

Table 4 The effect of the fault type and the location

Fault type Remote source capacity (MVA) Fault resistance () Fault location ( km) Error (%)
Desired output Actual output

AG 0 2 2.5 2.418 −0.21

AG 0 50 2.5 2.651 0.38
AG 0 100 2.5 2.787 0.72
AG 0 2 12.5 12.480 −0.05
AG 0 50 12.5 12.445 −0.14
AG 0 100 12.5 12.202 −0.75
AG 0 2 30.0 29.348 −1.63
AG 0 50 30.0 29.477 −1.30
AG 0 100 30.0 28.822 −2.95
AG 0 2 37.5 37.188 −0.78
AG 0 50 37.5 36.992 −1.27
AG 0 100 37.5 36.887 −1.53
BG 0 2 12.5 12.274 −0.57
BG 0 50 12.5 11.967 −1.33
BG 0 100 12.5 12.141 −0.90
BG 0 2 30.0 29.602 −1.00
BG 0 50 30.0 29.058 −2.36
BG 0 100 30.0 30.141 0.35
CG 0 2 12.5 12.690 0.48
CG 0 50 12.5 13.067 1.42
CG 0 100 12.5 12.638 0.35
CG 0 2 30.0 29.432 −1.43
CG 0 50 30.0 29.874 −0.32
CG 0 100 30.0 30.328 0.82
AG 10 2 12.5 12.493 −0.02
AG 10 50 12.5 12.445 −0.14
AG 10 100 12.5 12.664 0.41
AG 10 2 30.0 29.779 −0.56
AG 10 50 30.0 29.929 −0.18
AG 10 100 30.0 29.995 −0.01
AB 0 0 2.5 2.619 0.30
AB 0 0 12.5 12.542 0.11
AB 0 0 30.0 29.998 −0.01
AB 0 0 37.5 37.289 −0.53
AB 10 0 2.5 3.370 −0.33
AB 10 0 12.5 12.520 0.05
AB 10 0 30.0 30.004 0.01
AB 10 0 37.5 37.883 0.96
ABG 0 0 2.5 2.412 −0.22
ABG 0 0 12.5 12.486 −0.04
ABG 0 0 30.0 29.961 −0.10
ABG 0 0 37.5 37.249 −0.63
ABG 10 0 12.5 12.502 0.01
ABG 10 0 30.0 29.970 −0.08
ABCG 0 0 12.5 12.482 −0.01

123
132 Electr Eng (2012) 94:125–134

Table 5 The effect of the fault

resistance Fault type Source capacity (MVA) Fault resistance () Fault location ( km) Error (%)

End P End Q Desired output Actual output

AG 100 0 2 22.5 22.607 0.27

AG 100 0 20 22.5 22.812 0.78
AG 100 0 50 22.5 22.178 −0.81
AG 100 0 100 22.5 22.685 0.46
AG 100 10 2 22.5 22.587 0.22
AG 100 10 20 22.5 22.516 0.04
AG 100 10 50 22.5 22.523 0.06
AG 100 10 100 22.5 22.408 −0.23

3
the power frequency phasors from the voltage and current
waveforms, the transients present due to fault arc will only
2.5
have little secondary effects on the accuracy. Of course, if the
proposed technique is to be based on the transient compo-

Error (%)
2
nents, then the integration of the accurate models for the fault
arcs into the system model will be critical. In this work, only 1.5
the linear arc resistances were considered. But importantly,
1
of higher values as the latter has a significant influence on
the magnitude of the power frequency components.
0.5
2 5 10 15 20 25 35 50 65 85 100

5.3 The effect of the fault inception time Fault Resistance (Ω)

Fig. 8 The effect of the fault resistance

In practice, the faults can occur at any point in the wave, i.e.
the fault inception angle cannot be defined in advance. It is
thus important to ascertain the algorithm’s performance for is slightly affected for the defined faults using two cycles.
the faults at inception angles rather than those near the max- But, in the case of the faults supplied with the data with a
imum voltage. The tests were performed on the distribution length of only one cycle, the accuracy is significantly affected
system shown in Fig. 5 for the different inception angles of for ‘a’-phase-earth-faults created at 12.5 and 25 km with the
phase-earth faults and the results are summarised in Table 6. estimation errors of 3.16 and 4.86%, respectively.
It is clearly evident from the results that the algorithm main-
tains a high degree of accuracy which is virtually independent
of the fault inception angle.
6 Conclusion
5.4 The effect of the fault cycle
In this paper, an ANN-based fault location algorithm to
In the fault location algorithm, the DFT technique ignores classify and locate shunt faults on overhead power distribution
the first cycle of post-fault data since the transients are most feeders was presented. Our algorithm is based on utilising the
prominent during this period. However, there can be situa- fault voltage and current samples obtained at a single location
tions particularly under high-speed fault clearance that only of a typical radial distribution system with load taps and
the first cycle of information is available related to the fault remote-end source. The main advantage of the algorithm over
location so it is no longer possible to ignore it. In these cases, the other known fault location techniques is that there is no
all the post-fault information must be taken into account and requirement for communication from the remote-end source
it is thus important to ascertain the effect on the performance and the exact knowledge of pre-fault loading. The proposed
of the algorithm. A comparison of accuracy attained between algorithm is less affected by the fault resistance, the fault
utilising the first and second cycle of the data following an type, the fault location and the embedded remote-end source,
‘a’-phase-earth-fault on the distribution system is shown in and the errors attained being less than 3 % for the majority of
Fig. 5 (no remote-end in-feed) and the results are summa- system and fault conditions studied here. Although the algo-
rised in Table 7. The results clearly show that the accuracy rithm has been tested using CAD simulations, the practical

123
Electr Eng (2012) 94:125–134 133

Table 6 The effect of the fault

inception angle Fault type Remote source Fault Fault angle (◦ ) Fault location ( km) Error (%)
capacity resistance
(MVA) () Desired output Actual output

AG 0 2 90 25.0 25.031 0.78

AG 0 2 135 25.0 24.977 −0.08
AG 0 2 180 25.0 25.386 0.97
AG 0 2 225 25.0 25.089 0.23
AG 0 2 270 25.0 24.537 −1.16

Table 7 The effect of the fault

cycle Fault type Fault resistance () Number of cycles Fault location ( km) Error (%)
used
Desired output Actual output

AG 2 3 25.0 25.031 0.08

AG 2 2 25.0 25.041 0.10
AG 2 2 12.5 12.582 0.20
AG 2 1 12.5 13.762 3.16
AG 2 1 25.0 26.942 4.86

limitations of the hardware such as interface and quantization 4. Sachdev MS, Das R, Sidhu TS (1997) Determining locations of
errors are taken into account in the design process. faults in distribution systems. In: Developments in power systems
protection, 25–27th March 1997 Conference publication no: 434,
The results show that the overall performance of the algo-
IEE, pp 188–191
rithm is significantly better compared with to the conven- 5. Zhu J, Lubkeman DL, Girgis AA (1996) Automated fault location
tional techniques, particularly those based on impedance to and diagnosis on electric power distribution feeders. IEEE Winter
fault measurements which give unacceptably high errors in Meet 12:1–8
6. Aggarwal RK, Aslan Y, Johns AT (1997) A new concept in fault
the presence of the remote-end embedded generation and location for overhead distribution systems using superimposed
load taps. Although the algorithm was developed for the over- components. IEE Proc Gener Transm Distrib 144(3):309–316
head power distribution systems, it can be applied to the com- 7. El-Hami M, Lai LL, Daruvvala DJ, Johns AT (1994) A new trav-
posite systems consist of overhead lines and underground elling-wave based scheme for fault detection on overhead power
distribution feeders. IEEE Trans Power Deliv 17(4):1825–1833
cable sections. Another major advantage of the algorithm is 8. Johns AT, Lai LL, El-Hami M, Daruvvala DJ (1991) A new
that the ANNs can be trained off-line with the data reflecting approach to directional fault locator for overhead power distribu-
modifications on the existing system such as demographic tion feeders. IEE Proc Gener Transm Distrib 138(2):351–357
data of the area and seasonal or daily energy demand and 9. Jarventausta P, Verho P, Partanen J. (1994) Using fuzzy sets to
model the uncertainty in the fault location process of distribution
generation levels. Once the training is over, on-line fast and networks. IEEE Trans Power Deliv 9(2):954–960
accurate fault clearance and diagnosis can be achieved in the 10. Hsu YY, Lu FC, Chien Y, Liu JP, Lin JT, Yu HS, Kuo RT (1991) An
distribution system with embedded electric power genera- expert system for locating distribution system faults. IEEE Trans
tion. Power Deliv 6(1):336–372
11. Kezunovic M, Perunicic B, Mrkie J (1994) An accurate fault loca-
tion algorithm using synchronised sampling. Electr Power Syst Res
29(3):161–169
12. Luo S, Kezunovic M, Sevick DR (2004) Locating faults in the
References transmission network using sparse field measurements, simulation
data and genetic algorithm. Electr Power Syst Res 71:169–177
1. Eriksson L, Saha MM, Rockefeller GD (1985) An accurate fault 13. Gopalakrishnan A, Kezunovic M, Mckenna SM, Hamai DM
locator with compensation for apparent reactance in the fault resis- (2000) Fault location using the distributed parameter transmission
tance resulting from remote-end infeed. IEEE Trans Power Appa- line model. IEEE Trans Power Deliv 15(4):1169–1172
ratus Syst 104(2):424–436 14. Kezunovic M, Perunicic B (1996) Automated transmission line
2. Singh GK (2004) Self-excited induction generator research—a fault analysis using synchronised sampling at two ends. IEEE Trans
survey. Electr Power Syst Res 69:107–114 Power Syst 11(1):441–447
3. Girgis AA, Fallon CM, Lubkeman DL (1993) A fault location 15. Radojević Z, Terzija VV (2007) Effective two-terminal algorithm
technique for rural distribution feeders. IEEE Trans Ind Appl for overhead lines protection. Electr Eng (Archiv für Elektrotech-
29(6):1170–1175 nik), 89:425–432

123
134 Electr Eng (2012) 94:125–134

16. Radojević Z, Terzija VV (2006) Two terminals numerical algo- 22. Moore PJ (1997) Power system protection. Digital protection and
rithm for fault distance calculation and fault analysis. In: Power signalling. The Institution of Electrical Engineers, UK, vol 4,
systems conference and exposition, 29th October–1st November pp 32–36
2006, IEEE PES, pp 1037–1042 23. Gu JC, Yu S (2000) Removal of dc offset in current and voltage
17. Radojević Z, Terzija VV (2008) Intelligent two-port numerical signals using a novel filter algorithm. IEEE Trans Power Deliv
algorithm for transmission lines disturbance record analysis. Electr 15(1):73–79
Eng (Archiv für Elektrotechnik) 90:323–330 24. Purushothama GK, Narendranath AU, Thukaram D, Parthasarathy
18. Beccali M, Cellura M, Lo Brano V, Marvuglia A (2004) K (2001) ANN applications in fault locators. Electr Power Energy
Forecasting daily urban electric load profiles using artificial neural Syst 23:491–506
networks. Energy Convers Manag 45:2879–2900 25. Joorabian M, Taleghani Asl SMA, Aggarwal RK (2004) Accu-
19. Kalogirou SA (1999) Applications of artificial neural networks in rate fault locator for EHV transmission lines based on radial basis
energy systems: a review. Energy Convers Manag 40:1073–1087 function neural networks. Electr Power Syst Res 71:195–202
20. Ilunga M, Stephenson D (2005) Infilling stream flow data using 26. Temurtas H (2004) The joint based trajectory control with neural
feed-forward back-propagation (BP) artificial neural networks: generalized predictive control for a three joint robotic manipulator
Application of standard BP and pseudo Mac Laurin power series PhD. thesis, Sakarya University, Turkey
BP techniques. Water SA 31(2):171–176
21. Haykin S (1994) Neural networks, a comprehensive foundation.
Prentice-Hall, New Jersey

123