Design and Training of Artificial Neural

Networks for Locating Low Current Faults in

Distribution Systems
J. Coser, D. T. do Vale, Student Member, IEEE, and J. G. Rolim, Member, IEEE

application of support tools to these systems is also welcome.

Abstract—Artificial Neural Networks constitute a suitable When applied to the fault location problem, ANN can be
approach for pointing out a fault location in radial distribution trained with a historical database and after this phase be used
feeders, even when the fault current has a small value, near the to provide the distance to the cause of the problem when a
normal load of the system. Some publications have described
new fault occurs. However, in most cases there are not enough
successful application of artificial neural networks to the fault
location problem, but there are still some difficulties that may historical cases to train an ANN, so the faults are simulated
limit their applicability to a real system, mainly the complexity of [2]. The input variables may be the fundamental frequency
the problem when lateral derivations are included as possible phasors [3] or harmonic components of currents and voltages
fault locations. There are some inherent aspects in distribution during the short circuit [4].
networks that prevent the straightforward application of In [5] the authors apply an ANN to diagnose cases when the
transmission network methodologies to distribution systems. This
fault resistance is lower than 20 Ω in a 300 km transmission
paper describes a new approach to the use of Artificial Neural
Networks for the solution of the fault location problem in energy system with monitoring devices available only at one terminal.
distribution systems. The objective is to obtain accurate results The input data are the complex three-phase voltage and
and to optimize the training stage, all using only the current, including pre-fault information. The training set is
fundamental frequency component of the currents monitored at composed by 2515 patterns. According to its description, the
the substation method could be extended to deal with high impedance faults.
The large amount of patterns necessary to train the network
Index Terms-- Fault location, Neural network application,
Power distribution faults.
requires a considerable effort to obtain and organize them.
Some fault location methods are based on characteristics of
I. INTRODUCTION transmission lines, missing some important and particular
aspects of radial distribution feeders and consequently
A RTIFICIAL Neural Networks (ANN) applications are
common in several tasks in electrical power systems
planning and operation [1], for instance: load forecasting,
providing unsatisfactory results. Besides, there is the difficulty
of training a single ANN for the complete main feeder, even
when the method despises the existence of lateral derivations.
security analysis, digital protection and fault location. They In [3] the authors apply a method based on ANN for fault
are especially suitable to the solution of problems that are location in distribution feeders. The input data are voltage and
difficult to model and solve through traditional algorithms. In current signals monitored at the substation. The test system
these cases, ANN may provide a low cost solution with has a three section feeder connecting the substation to the
reasonable accuracy. load. 1075 patterns, obtained through simulations of several
Permanent outages in power systems may lead to the kinds of faults, were necessary for training the ANN.
interruption of the energy supply, causing losses for An approach combining two types of ANN and aimed to
consumers and utilities. Particularly in aerial distribution fault diagnosis in distribution systems with limited monitoring
systems, the network is exposed to several factors that may is analyzed in [6]. In order to avoid measurement errors in
cause short circuits, as contact with trees and animals, voltage signals provided by potential transformers and to
insulators breakdown, storms and car accidents involving reduce implementation costs, the authors in [7] only use
posts. Although they are less accident prone, underground current signals as input information. .
distribution cables are very difficult to inspect, so the The discussions about the application of artificial neural
networks to the problem of diagnosing faults in power systems
This research work was supported by CNPq (Conselho Nacional de should include a critical evaluation of the following aspects:
Desenvolvimento Científico e Tecnológico).
Joni Coser is with CEFET (Centro Federal de Educação Tecnológica), • Generally, the relevant variables for fault location
Chapecó, Santa Catarina, Brazil (phone: +55 49 3331-4651, e-mail: are measured only at the substation;
[email protected]). • Possibility of limited or incoherent information in the
Jacqueline G. Rolim and Daniel T. do Vale are with the Power Systems
Group at Federal University of Santa Catarina, Florianópolis, Brazil. (phone: fault occurrence database with;
+55 48 3721-9593, e-mail: [email protected] and [email protected]). • The required training set may become weighty;
 • Lateral derivations of the main feeder and their A. Faults in Radial Feeders
influence in the training and validation stages. Once the protection system detects a fault and disconnects
It is essential to employ a suitable model of the feeder and a faulty circuit, the restoration process must be efficient and
an suitable type of ANN for this particular problem, as well. quick, as power companies are subject to penalties for long
Appropriate strategies for reduction of the shortcomings in restoration times. Besides, the public image of the company is
the training process of the ANNs increase their applicability in directly affected by an inefficient restoration process.
a real environment, bearing in mind the required precision in Fault current magnitude in a feeder circuit is dependent
the localization of faults for a given system. upon a number of factors, such as the fault cause (or the fault
This article describes a modular analysis of the feeder for impedance), distance of the fault from the substation, pre-fault
the training of ANNs used for fault location in a distribution voltage and loading, and feeder impedance.
system, focused on to the cases of reduced fault current levels. Commonly, researchers of fault location, classify faults in
The ANNs are configured independently for the main feeder high and low impedance types. Example of low impedance
and its more significant derivations. The input data are faults is an arc between phases. A high impedance fault is
fundamental frequency phasors of sampled currents measured exemplified by a fallen conductor touching a dry pavement.
only at the substation. The outputs are possible fault distance Fig. 2 illustrates a single phase to ground fault model. High
values. impedance faults between phases only are rare events. In
general there is an earth path of high impedance in such type
II. IMPORTANT ISSUES ON DISTRIBUTION FEEDERS AND of fault. [8].
FAULTS
Distribution feeders constitute the major link between
electric energy systems and utilities’ final customers. The
current flow is commonly unidirectional, flowing from the
substation, in a three phase main circuit, serving loads at
several locations. Load unbalance between phases is quite
common. Aerial lines are predominant, but underground
cables are also very usual. Some of these aspects can be
related in to Fig. 1.

Fig. 2. Single Phase to Ground Fault Representation.

It is remarkable that, in very long feeders, faults located far

away from the substation also can produce very low fault
current levels. Hence, in this paper, the faults are classified
according to the fault current magnitude, in order to designate
events with fault current levels very close to the normal
loading current of the system.
B. Fault Location in Distribution Networks
Visual inspection is a pioneer method for fault location [9].
Obviously this method is limited, so several procedures for
automatic fault location functions have been developed over
the recent years. Different approaches are characterized mostly
by the input data used to estimate the fault location.
Fig. 1. Typical Distribution Feeder Elements. The measurement of the impedance seen from the
substation during a fault [10], which characterizes the well-
Characteristic impedance of the conductors, tower design known reactive component method, is the classical
and phasing configuration may change along the feeder, from formulation still employed in digital fault locators and
one segment to another. Lateral derivations are originated protection devices. However, this method is limited by the fact
from the main feeder to supply one, two, or three phase loads that load currents are not taken into account, so the errors are
at remote locations. very high for when the fault current is low.
Utilities frequently use a recloser at the substation and fuses Another approach for dealing with the fault location
for the lateral derivations. Enhanced schemes are employed problem is the use of fault generated traveling waves [11].
depending on the available resources and utilities’ practices. Accurate estimation is provided by this method for faults in
Digital relays used to protect distribution feeders are often the main circuit, but a number of shortcomings arise for
provided with fault location functions, but for the most part, locating faults in lateral derivations. The high sampling rates
they are not efficient for low current fault location. and communication requirements demand more expensive and
complex devices, in comparison to the fundamental frequency
approach. IV. PROPOSED METHOD FOR LOCATING LOW CURRENT
The use of artificial intelligence methods, particularly FAULTS USING ANNS
ANNs, is adopted as an useful option in cases where the
analytical methods, such as the reactive component, do not A. Input Parameters
provide satisfactory results (such as low current faults). This Once a feeder model for a given operating condition is
approach has a low cost of implementation when compared to built, fundamental frequency current phasors (magnitude and
the use of traveling waves (considering that fundamental phase angle) derived from waveforms at the substation during
frequency variables can be employed as input data for feature a simulated fault are employed as input data for the proposed
extraction). ANNs. Pre-processing of current samples is necessary for
ANNs are adequate for the problem of locating low current calculating the fundamental frequency phasors. Afterwards, it
faults in distribution feeders, when no analytical method is necessary to detect and to point out the faulted phases, since
provide reasonable responses, as long as appropriate data the distance to fault evaluation stage will use this information.
selection and training strategies are adopted. B. Fault Identification
C. Multiple Estimations for Locating a Single Fault in a The fault type is identified by comparing the magnitude of
Distribution System the phase and zero sequence currents during the fault to
Faults in different locations of a distribution feeder may threshold values defined for each feeder
lead to similar values of voltages and currents measured at the As the occurrence of low current faults in distribution
substation, giving raise to multiple estimations for the location feeders are usually related to faults involving the ground, the
of a fault. Bearing this in mind, it is necessary to filter out analysis presented in the studied cases will focus on such fault
false locations using additional information to the voltage and types.
current measurements at the substation. For the location of high current faults between phases, it is
useful to employ analytical methods which perform such task
III. TRAINING AND CONFIGURING ANNS TO LOCATE with good performance. [8].
DISTRIBUTION FEEDER FAULTS
C. Proposed Neural Model
Fault location via ANN is performed by supervised
The proposed neural network model employs Feedforward
training, where the input training set is composed voltages and
Multilayer perceptrons, trained with the Levenberg-Marquardt
current signals, either fundamental frequency [3] or harmonic
algorithm. Some tests using radial basis ANNs have also been
content of the waveforms during the fault [4]. A number of
performed [12], but the first mentioned neural architecture
variations of these two basic schemes are reported in the
provided better results, so it has been adopted in further stages
literature.
of the method’s development.
In order to achieve a good performance, the ANN must be
Two intermediate layers of neurons are used with a
trained with a set of representative fault cases. There should be
sigmoidal activation function. The output layer uses a linear
some extra cases to evaluate the ANN and assure it has good
activation function, which provides output values that are
generalization capacity.
proportional to the distance to fault (Figure 3).
The choice of the number of intermediate layers of neurons
is defined by the complexity of the function determined by the
training samples. A neural network based tool design is
problem dependent and demands accurate analysis for correct
parameter adjustment.
When the application is fault location, the usual approaches
fit well to the transmission line model, whose topology with
no lateral derivations allows to obtain satisfactory accuracy
using a single training set.
When distribution feeders are considered, superposition of
training patterns may occur due to faults which lead to similar
values at the substation, but are placed in different locations
(main feeder or laterals). This can reduce the practical interest
of applying ANNs to this specific problem. Fig. 3. Topology of the Neural Network.
From this point of view, it is important to derive particular
procedures for ANNs application in distribution feeder faults, The stop criterion for training is the mean square error.
in order to avoid superposition, alleviate simulation effort in Information on other parameter adjustments is presented in
the composition of the training/validation/and test sets and Table I.
obtaining accurate estimates in a real application starting from D. Specific Design for Distribution Feeder Faults
a simulated environment.
The training strategy described bellow is intended to work
with a relatively low number of training cases, but obtaining
adequate generalization capacity.
 The fundamental principle of the method is related to the segment are assigned in Fig. 5 as points to simulate faults for
situation in Fig. 4, where faults located in the points indicated training.
may lead to similar values of current magnitude and phase Regarding the fault impedance, the training set included
angle when seen from the substation faults simulated using the following values: 100, 150, 200,
250, 300, 350, 400, 450, 500 and 550 Ω. By using such values,
TABLE I the fault current is in the range between two and four times the
ANNS ADOPTED PARAMETER ADJUSTMENT.
nominal operating current, for the IEEE 34 Bus Test-System
Input Layer Neurons 2 x Faulted Phasees
Intermediate Layers 02 employed in the results analysis (to be described in Section
Intermediate Layer Neurons 05 V).
Output Layer Neurons 01 The output of the neural network varies from 0.1 for faults
Activation Function of the Intermediate Layer Neurons Sigmoid at the beginning of the segment, to 0.9, for faults at the end of
Activation Function of the Output Layer Neurons Linear
the segment. It is also necessary to perform the simulations
Maximun Mean Square Error 1e-6
Learning Rate 0.01
with different load periods.

Fig. 5. Points of Main Feeder where Faults are Simulated to Compose the
Training Set

E. Dealing with Multiple Solutions

Many approaches may be applied to reduce the number of
possible locations of a fault in a distribution feeder. For
instance, the trouble phone calls after an interruption, in
Fig. 4. Multiple Faults which May Lead to Identical Current Phasors at the together with information about the system topology, may
Substation.
define the feeder segment where the fault actually happened.
By taking a glance to the situation of Fig. 4, it is possible to In some distribution systems special devices installed in
understand that a ANN trained with fault cases simulated on derivation points to detect the presence of current, facilitate
the main feeder can also estimate the distance to fault for an the task of identifying the correct location among multiple
event on a lateral. Hence, the number of required simulations solutions provided by the automatic tool.
can be reduced, once the neural network trained for the main The variation of the feeder load after the protection trip
feeder can be used to other faults in laterals. may also help in this task. [2].
Additional information is obviously necessary after
obtaining the possible fault locations, in order to know if the V. IMPLEMENTATION AND RESULTS
fault is in the main feeder or in a lateral, but the distance to
fault estimation can be executed using data from the main A. IEEE 34 Node Test System
feeder in a first evaluation stage. The IEEE 34 Bus Test System, presented in Fig. 6, was
Regarding the faulted phases, the proposed tool uses a used to simulate the necessary training cases and to evaluate
specific neural network for each fault type. the performance of the trained ANN. The platform used to
For lateral derivations whose impedance is significantly simulate the fault cases was ATP (Alternative Transient
different from the main feeder, the indirect estimation Program). The feeder was represented by its π model with
mentioned above may not work. This fact has been analyzed concentrated parameters and loads modeled by constant
in a number of simulation studies. In such cases, specific impedances concentrated at the end of each branch. The rated
neural networks must be trained using faults simulated on line voltage is 24.9 kV and the total installed load is 2.054
these laterals, in order to point out a possible fault location in MVA. The main feeder (between nodes 800 and 840) is 57.68
such segments. km long. This feeder is long, so the task of locating low
The location of the faults simulated to compose the training current faults using information obtained only at the substation
set has been uniformly distributed along a feeder segment, in becomes more difficult.
order to obtain an ANN with good generalization ability.
Notice that five locations from the substation to the end of the
 Some results with faults simulated between phase a-earth
in several nodes of the main feeder are presented in Table III.
It is possible to observe in this table that the location system
provided good results.

TABLE III
FAULT LOCATION FOR SHORT CIRCUITS ON THE MAIN FEEDER
Node Rf (Ω) Location (km) Error (%)
150 11.28 1.26
250 11.39 2.24
808
(11.14 km) 350 11.26 1.08
450 11.26 1.08
550 11.35 1.88
150 35.02 0.52
250 34.99 0.43
824
Fig. 6. IEEE – 34 Nodes Test System (34.84 km) 350 35.19 1.00
450 35.24 1.15
550 35.28 1.26
B. Training ANNs for the Test System 150 41.53 0.12
The ANNs trained for locating faults in the main feeder 250 41.49 0.02
854
350 41.47 0.02
were also able to provide satisfactory results for faults in the (41.48 km)
450 41.46 0.05
lateral derivations, except for those that begin in nodes 816 550 41.53 0.12
and 854. 150 53.96 0.44
Consequently, two specific neural network modules 858
250 54.17 0.06
(ANNs path2 and ANNs path3 presented in figure 7) were (54.20 km) 350 54.07 0.24
450 54.08 0.22
trained for these two laterals 550 54.04 0.30
The total number of ANNs trained for faults between a 150 56.37 0.41
single phase and earth is presented in Table II, which also 860
250 56.49 0.19
gives information on the amount of simulations necessary for (56.60 km) 350 56.36 0.42
450 56.31 0.51
composing the training set, considering three different loading
550 56.20 0.71
profiles.

TABLE II
The same ANN provides results with good accuracy for
AMOUNT OF SIMULATIONS NECESSARY FOR TRAINING THE ANNS faults in some laterals, as for instance when located at node
Trained 834. Figure 8 illustrates how an indirect estimate of a fault
Feeder Section Input Vectors
ANNS located in a lateral may be obtained. The results presented in
Main Feeder 09 450 Table IV show the efficiency of the ANN trained for the main
Lateral 816-822 03 90 feeder for locating faults in laterals, except for some cases of
Lateral 854-856 03 90 faults in lateral 816-822. The errors for faults in these laterals
Total 15 630
indicate the necessity of an ANN specific for locating faults
on this branch.
The results provided by the ANN trained for path 2 (fig. 7)
are presented in Table V. The accuracy of the results in all
cases used in the validation phase is high, especially
considering that this tool is dealing with low current faults.

Fig. 7. ANNs Modules Trained for the IEEE 34 Bus Test System

C. Results
The ANNs were evaluated applying cases simulated with Fig. 8. Estimative of the Distance to a Fault in a Lateral using the output of the
the maximum load and high impedance faults (low current). ANN trained for the main feeder.
 The real location of the fault when the distance to a fault VII. REFERENCES
indicate more than one possible location may be obtained [1] R. Aggwrwal and Y. Song, “Tutorial – Artificial Neural Networks in
using the techniques described in section IV.E. Power Systems, Part 3: Examples of Applications in Power Systems.” in
Proc. 1998 Power Engineering Journal, vol. 12, n. 6 (Dec.), pp. 279-
TABLE IV 287.
FAULT LOCATIONS FOR SHORT CIRCUITS ON LATERALS USING THE ANN [2] J. Zhu, D. L. Lubkeman and A. A. Girgis, “Automated Fault Location
TRAINED FOR THE MAIN FEEDER and Diagnosis on Electric Power Distribution Feeders”, in Proc. 1997
IEEE Transactions on Power Delivery, vol. 12, n. 2 (April), pp. 801-
Node Rf (Ω) Location (km) Error (%) 809.
150 55.77 0.52 [3] E. A. Mohamed and N. D. Rao, “Artificial Neural Network Based Fault
250 55.90 0.28 Diagnostic System for Electric Power Distribution Feeders”, in Proc.
842
(56.06 km) 350 55.75 0.55 1995 Electric Power System Research, n. 35, pp. 1-10.
450 55.68 0.68 [4] Y. Assef, O. Chaari and M. Meunier, “Classification Of Power
550 55.56 0.89 Distribution System Fault Currents Using Wavelets Associated To
150 57.48 0.19 Artificial Neural Networks”, in Proc. 1996 International Symposium On
250 57.58 0.02 Time-Frequency And Time-Scale Analysis, pp. 421-424.
846
350 57.50 0.16 [5] Z. Chen and J. C. Maun, "Artificial Neural Network Approach To
(57.59 km)
450 57.51 0.14 Single-Ended Fault Locator For Transmission Lines", IEEE
550 57.45 0.24 Transactions On Power Systems, vol. 15, Issue 1, pp 370-375, Feb.
150 32.71 1.46 2000.
250 32.80 1.74 [6] D. Thukaram, U. J. Shenoy and H. Ashageetha, “Neural Network
818 Approach for Fault Location in Unbalanced Distribution Networks with
(32.24 km) 350 33.17 2.88
Limited Measurements”, in Proc. 2006 IEEE Power India Conference, 8
450 33.35 3.44
pp..
550 33.51 3.94
[7] L. Souza Martins, J. F. Martins, C. M. Alegria and V. Fernão Píeres, “A
150 50.20 6.99
Network Distribution Power System Fault Location Based on Neural
250 51.11 8.93 Eigenvalue Algorithm”, in Proc. 2003 IEEE Bologna Power Tech
820
(46.92 km) 350 52.20 11.25 Conference, vol. 2 (June), pp. 23-26.
450 53.57 14.17 [8] R. Das, “Determining the Locations of Faults in Distribution Systems”
550 54.97 17.16 Doctoral Thesys, Dept. of Electrical Engineering, Univ. of
Saskatchewan, Saskatoon, Canada, 1998.
TABLE V [9] T. W. Stringfield, D. J. Marihart and R. F. Stevens, “Fault Location
FAULT LOCATIONS FOR SHORT CIRCUITS ON LATERALS USING THE ANN Methods for Overhead Lines”, in Proc. 1957 Transactions of the AIEE,
TRAINED FOR PATH 2 part III - Power Apparatus and Systems, vol. 76 (Aug.), pp. 518-539.
[10] K. Srinivasan and A. St-Jacques, “A New Fault Location Algorithm for
Node Rf (Ω) Location (km) Error (%)
Radial Transmission Lines With Loads”, in Proc. 1989 IEEE
150 32.23 0.03 Transactions on Power Delivery, vol. 04, n. 03 (July), pp. 1676-1682.
250 32.23 0.03 [11] D. W. P. Thomas, R. J. Carvalho and E. T. Pereira, “Fault Location In
818
(32.24 km) 350 32.25 0.03 Distribution Systems Based On Traveling Waves”, in Proc. 2003 IEEE
450 32.25 0.03 Bologna Power Tech Conference, vol. 2.
550 32.27 0.09 [12] J. Coser, D. T. do Vale and J. G. Rolim, “Artificial Neural Network
150 47.10 0.38 Based Method for Fault Location in Distribution Systems”, in Proc.
250 46.95 0.06 2005 13th International Conference on Intelligent Systems Application
820
(46.92 km) 350 46.97 0.11 to Power Systems, pp. 157-162.
450 46.90 0.04
550 46.96 0.09 VIII. BIOGRAPHIES
Joni Coser received his degree in Electrical Engineering from Federal
University of Santa Catarina (UFSC), Brazil, in 2000, and the M.Sc. and
VI. CONCLUSIONS Ph.D. degrees in Electrical Engineering also from UFSC, in 2002 and 2006,
respectively. He is currently with CEFET (Centro Federal de Educação
The tool for fault location described in this paper was Tecnológica), Chapecó, Santa Catarina, Brazil. His research interests and
specifically designed for distribution networks, being able to recent professional experience include power transmission and distribution
locate low current faults not only in the main feeder, but also system planning, modeling and analysis; fault location methods and artificial
intelligence applications in power systems.
in lateral derivations.
This tool employs a modular structure of artificial neural Daniel T. do Vale is an undergraduate student of Electrical Engineering at the
networks, and efforts were dedicated to reduce the number of Federal University of Santa Catarina (UFSC), Brazil. He is currently working
cases necessary to train the ANN, without loss of on research projects at the Power System Group. His main areas of interest are
power systems protection and artificial intelligence applications in power
performance. systems.
The input data are the magnitude and the angle of the
fundamental frequency component of the current monitored at Jacqueline Gisèle Rolim received her BSEE and M.Sc in Power Systems
from the Federal University of Santa Catarina (UFSC), Brazil, in 1982 and
the substation, thus the approach is simple to implement with a 1988, respectively. From 1985 to 1991 she worked in HV and EHV substation
low cost, as it does not require any special device for projects. She was in a split PhD program at UFSC and Brunel University
monitoring the feeder variables. (England) and received her degree in 1995. Since 1996 she is an associate
On the other hand, in some cases the distance indicated in professor at UFSC and her main areas of interest are power systems
protection, diagnosis and the application of artificial intelligence techniques.
the output of the ANN may lead to more than one possible
location of the fault and extra information would be necessary
to identify the correct alternative.