Electric Power Systems Research 81 (2011) 13251333

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

High impedance fault detection methodology using wavelet transform and

articial neural networks
Ibrahem Baqui, Inmaculada Zamora , Javier Mazn, Garikoitz Buigues
Department of Electrical Engineering, Faculty of Engineering, Alameda de Urquijo s/n, 48013 Bilbao, Spain

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a new technique based on the combination of wavelet transform (WT) and articial
Received 11 February 2010 neural networks (ANNs) for addressing the problem of high impedance faults (HIFs) detection in electrical
Received in revised form distribution feeders. The change in phase current waveforms caused by faults and normal switching
15 December 2010
events has been used in this methodology. The discrete wavelet transform (DWT) used decomposes
Accepted 31 January 2011
the time domain current signals into different harmonics in time-frequency domain and extracts special
Available online 23 February 2011
features to train ANNs. This preprocessing reduces the number of inputs to ANN and improves the training
convergence. The ANN structure and learning algorithm used in this method is the multilayer perceptron
Keywords:
Articial neural networks
network and LevenbergMarquardt back-propagation algorithm, respectively.
Fault detection The signal data of several HIFs, low impedance faults (LIFs) and normal switching events have been
High impedance faults obtained by the simulation of a real distribution network, with ve feeders, under these different oper-
Wavelet transform ations conditions, using SimPowerSystem Blockset of MATLAB. The results obtained have validated the
effectiveness of the proposed methodology to detect HIFs and discriminate them from normal transient
operations.
2011 Elsevier B.V. All rights reserved.

1. Introduction The rst methods developed to detect high impedance arcing

faults were based on the direct measurement of the basic electric
One of the main problems the electrical utilities have to cope parameters, such as voltage and current, analyzing their varia-
with in power distribution systems is the failure of the HIF detec- tions or their harmonic components and using different methods or
tion, which may lead to a serious threat in electric shock or/and combinations of them [17]. However, these methods sometimes
re. These HIFs are caused by a downed energized conductor on mistrip due to similarity between frequency domain information
high impedance surfaces, such as dry ground, concrete, gravel, produced by HIFs and normal system switching events. Thus, it is
high resistively soils and asphalt road. Additionally, sometimes difcult to nd one or two critical harmonic components that can
HIFs appear when the energizing conductor makes contact with discriminate one disturbance from another.
grounded objects such as tree limbs. Besides, the high difference of Therefore, the HIF detection algorithms that are only based on
potential between the two contact points can produce an electric the power frequency or on some current harmonics (neglecting
arc reignition. other high frequency components that exist in the fault current)
In general, the current values that appear due to HIFs are lower are not reliable algorithms. For that reason, new methods have to
than the overcurrent thresholds, so the phase or residual overcur- be developed. Currently, there are multiple research works trying to
rent devices remain inactive. Moreover, the resulting current and achieve the best solution for this problem. Thus, additional methods
voltage waveforms are frequently characterized by random and used for the characterization of these faults can be based on the
intermittent data, caused by the arcing and/or nonlinearity of the injection of low frequency signal analysis [810], ANN [1114], WT
fault impedance, which makes impossible to nd single parame- [1519] and intelligent systems [2023].
ters. Hence, the detection of HIFs is really an important challenge The signal processing investigation on resulted signals from
from the public protection and operational reliability point of view. power systems, considering every power system condition, leads
Although many detection methods have been developed up to now, to develop algorithms based on time and frequency domain,
a perfect dependable and secure detection is still the objective of which improve the detection of HIFs in distribution systems. In
continuous research. spite of obtaining time and frequency information, the combined
information of high and low frequencies can be obtained by the
decomposition of the WT. Regarding to the methods based on using
Corresponding author. Tel.: +34 946014063; fax: +34 946014200. only WT to HIF detection, they have been often criticized because
E-mail address: inmaculada.zamora@ehu.es (I. Zamora). the data set of registered faults do not agree with other published

0378-7796/$ see front matter 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.epsr.2011.01.022
1326 I. Baqui et al. / Electric Power Systems Research 81 (2011) 13251333

Fig. 1. Schematic diagram of the methodology.

studies. On the other hand, the ANN algorithms have been tested by extraordinary computing environment in order to implement the
many research groups, but the perfect discrimination of capacitor WT in an efcient way onto a computer. This implementation can be
switching and high impedance faults is still an unsolved problem. done either by command line functions or by graphical interactive
Therefore, because of the drawbacks of using WT or ANN separately, tools.
the combination of both WT and ANN is presented as a better solu- More precisely, it has been considered the use of the DWT
tion to the problem of HIF detection in distribution power systems instead of the continuous wavelet transform (CWT), in order to
[24,25]. reduce the vast amount of computational work and data the latter
This paper proposes a new methodology by the application would require.
of WT and ANN to detect HIFs, LIFs and discriminate them from
normal transient switching operations. The data signals used in 2.1.1. Denoising process
the validation process are HIFs, LIFs and some normal transient The denoising process is used to eliminate the existing distortion
operations. These data have been obtained from simulations of a in the current signals, which may have been produced by several
real power distribution system, of 13.8 kV, using SimPowerSystem events: switching operations, events occurred in other feeders, etc.
Blockset of MATLAB. The main idea of this process is a DWT preprocessing, in order
to convert the three-phase current signals into one dimension hard
2. Proposed high impedance fault detecting methodology threshold denoising stage. For this purpose, a Daubechies order 5
(db5), level 4 wavelet decomposition is applied to the registered
The proposed methodology to detect HIFs and LIFs, as well as currents. The threshold rule selected is a heuristic variant of the
to discriminate them from normal transient switching operations, principle of Steins Unbiased Risk Estimate.
comprises two stages. The decomposition level and basis function have been selected
In a rst stage, the current signals of the feeders, using WT, are after a thorough analysis of different levels and many types of basis
analyzed to obtain the relevant data signals. Afterwards, in a second functions. The objective is to nd the optimum denoising for the
stage, the properly trained ANNs are used as a classier processing, signals in every simulated case of each feeder of the distribution
in order to classify the state of each feeder. The structure of the power system described in Section 3.
application of WT and ANN in the proposed method is presented in
Fig. 1, by means of a schematic block diagram, which describes the 2.1.2. Signal decomposition
different stages of the methodology. The information obtained from the time-domain signals regis-
Although some methods that use similar approaches may be tered under normal or fault situations, is usually not enough to
affected by topology changes in the distribution network, this detect the HIFs. Therefore, the DWT is applied to transform the
methodology is specially designed to be applied in the rural dis- denoised time signals to time-frequency domain signals, where the
tribution systems whose topology is almost invariable, scarcely different characteristics of each current signal may appear more
changeable. These kinds of distribution systems are very typical clearly. That is, by showing large coefcients in different frequency
in several countries. bands when the disturbance appears. This process is known as
decomposition process, in which the Daubechies basis of order four
2.1. Application of the wavelet transform (db4), in seven decomposition levels, has been applied to produce
different frequency bands of the signal. The features from all these
During the process of HIF detection, the signal data need to be frequency components are used to give good discrimination results.
analyzed to nd adequate information that can be useful for the When analyzing the characteristics of current signals by the use
fault detection, because it may not clearly appear in the original of DWT, the following parameters must be specied:
time signal. That is why the application of WT is divided in three
main groups: - Sampling frequency.
- Window length.
- Denoising process of the current signals. - Levels of decomposition.
- Signal decomposition. - Wavelet basis type.
- Feature extraction.
Usually, a low sampling rate leads to the reduction of the wavelet
Consequently, it has been considered the use of the Wavelet computation process. However, in some cases, such as in fault situa-
Toolbox from MATLAB [26], which provides useful functions and an tion and switching operations, the sampling rate needs to be high so
 I. Baqui et al. / Electric Power Systems Research 81 (2011) 13251333 1327

as to be able to capture the characteristic information of the signal. Table 1

Representative cases.
After testing many possibilities, among 128, 256 and 512 samples
per cycle, a sampling frequency of 512 samples per cycle (sampling Feeders Cases Train Test
rate of 2.56 kHz) has been adopted in the developed analysis. Feeder 1 8734 2612 6122
The sampling frequency, along with the Wavelet basis function Feeder 2 6121 1343 4778
selected, also inuences on the window length and on the decom- Feeder 3 5804 790 5014
position level number. In this methodology, the implementation of Feeder 4 5548 872 4676
Feeder 5 6267 964 5303
the DWT is based on the Multi Resolution Analysis (MRA) theory,
Total 32,474 6581 25,893
which requires ltering and downsampling. This decomposition
level (number of stages) is inversely related to the frequency com-
ponents of any level. Consequently, the higher decomposition level, Table 2
ANN output.
the lower frequency components are considered.
The characteristics of the Wavelet basis functions play an Feeder state Output 1 Output 2 Output 3
important role in their implementation too. Therefore, selecting Normal 1 0 0
an appropriate mother Wavelet to extract the useful information LIF 0 1 0
quickly and efciently is crucial to increase the performance of any HIF 0 0 1
particular application. The Wavelet basis selection usually depends
on the type of application or the similarity between the basis func-
tion and the signal to be analyzed. For example, Wavelets with small operation is based in three important stages that are described
number of lter coefcients or vanishing points are less localised below:
in frequency and more localised in time. Therefore, in applications
where the information needed is located at a specic instant of - ANN training
time, lower Wavelet basis with less number of coefcients are a - ANN verication
better choice than Wavelets with higher number of coefcients. - ANN operation stage
During the research developed for this methodology,
Daubechies Wavelets functions with different vanishing moments For this purpose, the Neural Network Toolbox from MATLAB [27]
have been tested, such as, db4, db6, db8. These mother functions provides different functions to design, initialize, simulate, train and
are orthogonal to any function obtained by stretching them by a show results of a neural network. This tool offers the possibility to
2j factor or shifting them right or left by the product of a constant select among many types of networks, such as perceptron, radial
to 2j. The orthogonal functions have been selected because of basis network, learning of vectorial quantization (LVQ), Elman net-
their speed in calculation. Besides, Daubechies waveforms have work, Hopeld network. It also permits the selection of several
some other useful properties to be considered, such as accurate transfer functions for each neuron, making possible different net-
detection of low amplitude of signals, short duration and fast work architectures by the combination of many units of neurons.
decay. The training parameters for different training algorithms can also
be selected.
2.1.3. Feature extraction
The main idea of making a feature extraction is to reduce the 2.2.1. Articial neural network training
amount of information, either from the original waveform or from This process is based on providing STD values of the decomposed
its transformation format. signals to the selected neural network. These values are extracted
In this work, the process of feature extraction consists on nding from the current signals of a representative number of cases, con-
the distinctive waveform parameter, with signicant information, sidering the different distribution system operations, some of them
that can represent the fundamental characteristics of the problem. used for training and the rest for testing (Table 1). As a result, the
The Standard Deviation (STD) of the coefcients, in each frequency ANN learns the problem characteristics and improves its general-
band, is chosen to be extracted and used as the input data vector to a ization property in the operation stage.
multilayer ANN. This feature has been selected after numerous tests The behavior of the selected ANN depends on numerous param-
and comparisons between the performance of neural networks eters, such as the number of hidden layers, the number of hidden
(generalization, simplicity, efciency and convergence speed) and neurons, transfer functions, initial weights and biases, training rule
other features like energy and RMS of each frequency bands (signals and training parameters. In some cases, choosing one of these
and coefcients). parameters wrongly may lead to overtting or other kind of prob-
The STD of the output signal is the square root of the data vector lems. In order to overcome this problem, a software tool called
variance, as it is shown in (1). This feature provides information SARENEUR [28] has been used. This tool is an automatic multifunc-
about the level of variation of the signal frequency distribution. tion program which allows nding, training and verifying optimal

2
 ANN architectures. After numerous tests and modications, it has
 
1  1
n n been found that an unidirectional multi layer neural network (mul-

STD =  xi xi (1) tilayer perceptron, MLP) provides a high degree of success for the
n 1 n
i=1 i=1 research developed in this methodology.
According to the feature extraction process, the input vector of
where x is the data vector and n the number of elements in that each ANN comprises 24 elements, 8 features in each phase current
data vector. obtained by the use of the DWT (D1, D2, D3, D4, D5, D6, D7, A7).
The output layer has three neurons, each one indicating one of the
2.2. Application of articial neural network feeder situations (Normal, LIF and HIF). Apart from that, a variable
number of neurons in hidden layers have been analyzed. The tan-
The ANNs have excellent characteristics, which make them a sigmoid transfer functions have been applied for both hidden layer
very powerful tool to be applied in pattern recognition and clas- neurons, whereas the log-sigmoid function has been selected for
sication. Therefore, the ANN is used in this method to learn output layer neurons. The combination of the neurons in the output
inputoutput relationship from input feature vectors. The ANN layer classies the feeder situation according to Table 2.
1328 I. Baqui et al. / Electric Power Systems Research 81 (2011) 13251333

During the learning stage, the supervised LevenbergMarquardt

algorithm has been utilised. The network stops learning when the
Mean Squared Error performance function (MSE) or the number of
iterations reached a predetermined value.
As it has been previously stated, in case of a distribution system
with very few topology changes, which is a common situation in the
distribution networks of several countries, there would be possible
to have an ANN trained for each foreseeable change. Consequently,
the number of ANNs would be small.

2.2.2. Articial neural network verication

When the training process is nished, the optimal network con-
gurations obtained are veried with the SARENEUR software tool,
in reference to their generalization ability under untrained situa-
tions.
After the training and testing process, SARENEUR found 47 dif-
ferent valid ANNs to operate. Among them, the neural network with
24-14-6-3 architecture was selected because it had the smallest
errors, according to the comparison between the network outputs Fig. 2. Diagram of the real distribution system.
and its corresponding targets over the test dataset, by means of (2).

target output calculated output allows engineers to build electrical models in Simulink environ-
%Error = 100 (2) ment and change their operation conditions.
target output

Taking as an example the dataset of feeder number one (Section 3.1. Distribution power system modelled
3), the solutions matrix obtained from (2) shows the percentage
comparison between the outputs and targets of 6122 testing cases, The modelled system, located in the Basque Country (Spain), is
using the previously mentioned ANN architecture. This matrix (3) a medium voltage network that consists of a substation and ve
is organized in such a way that the rst column corresponds to distribution feeders with radial conguration (Fig. 2).
the percentage output error (three outputs) under normal situa- The following system elements have been considered to be the
tions, while the second and the third columns correspond to the most relevant:
percentage output error of LIF and HIF cases, respectively.
 - A three column three-phase transformer, delta-wye connection
0.0030 0.0003 0.0003
with accessible neutral, voltage ratio of 30/13.8 kV and 12 MVA of
%Error = 0.0000 0.0000 0.0001 (3)
rated power.
0.0015 0.0001 0.0006
- Five feeders composed by 35.7 km of overhead lines and 8.9 km
of underground cables. The overhead lines consist of ACSR and
2.2.3. Articial neural network operation stage copper conductors. The underground cables are composed of alu-
Once the training and testing processes with SARENEUR have minum or copper conductors with EPR insulation.
been concluded, the resulting networks are ready to operate. - The system supplies a set of 98 loads. Because of the large number
Due to the use of the log-sigmoid transfer function in the output of loads, they have been concentrated considering the type of
layer, the ANN outputs vary from 0 to 1. Therefore, to reach the conductor. As it has been proved, this concentration of loads does
desired results of either 1 or 0, the outputs are processed with not affect the reliability of the results obtained in the study.
a round function, in order to round the output data vector to the
nearest integer value. The model of this real distribution system has been validated
Finally, the outputs of the proposed method give the state with two different software tools (ATP and Matlab/Simulink) and
(healthy or faulty) of a distribution feeder, when all stages of the with real data, provided by the Spanish electrical utility IBERDROLA.
method are followed. If the method gives indication of LIF or HIF in As an example, the complete conguration of the model of
12 consecutive iterations, the outputs of the method may be used feeder 5 can be observed in Fig. 3.
to take an appropriate control action. Nevertheless, when a feeder
is under a normal situation, the method turns back to take a new 3.2. Simulation of the power system operation
data window after 32 samples and then every step of this detection
methodology is repeated again. On the modelled distribution system, different operation con-
ditions have been simulated. In this section, the characteristics of
3. Modeling and simulation the different events considered are shown.

In order to verify the methodology of the proposed technique, 3.2.1. Normal situations
it is necessary to obtain the relevant data signals from the power Usually, the arcing and nonlinear characteristics of the currents
system, under different possible operating conditions. waves due to HIFs are similar to those caused by load changes and
It is well known that eld fault testing on real power systems is switching operations. Therefore, the modelled distribution system
difcult because of technical and economical reasons. In addition, has been loaded with both, linear (LL) and non-linear loads (NLL), to
the eld test data usually suffer from certain limitations. That is why reect a typical loading scenario with the purpose of investigating
a real electrical distribution system, under different conditions, the performance of the proposed algorithm under load switching
has been accurately modelled and simulated with SimPowerSys- operations.
tem Blockset of MATLAB. SimPowerSystem Blockset in MATLAB is The NLLs are usually dened as loads in which the load current
a complete and powerful electrical engineering software tool that waveform does not vary directly with the load voltage waveform.
 I. Baqui et al. / Electric Power Systems Research 81 (2011) 13251333 1329

Fig. 3. Conguration of feeder number 5.

Fig. 4. Nonlinear load model.

These loads are the main source of harmonics that distort voltage feeder currents. The large starting currents of a motor may alter the
and current waveforms in the power system. Contrary to LLs, in current characteristics of the feeder for a short time. Consequently,
which the voltage and current waves increase or decrease propor- an induction motor has been modelled using the Asynchronous
tionately. Typical examples of NLLs are air-conditioning, UPS, DC Machine block from SimPowerSystems.
motor drives and other power electronic devices [29]. Additionally, to make the distribution system stable and ef-
The model of the nonlinear load used in this analysis (Fig. 4) is an cient, reactive current must be supplied when and where it is
AC/DC/AC converter. This model consists of a rectier and an IGBT necessary. The practical and efcient way for the utility to pro-
inverter. The inverter is pulse width modulated (PWM) to produce vide this capacitive reactive current is through the installation of
a three-phase 50 Hz sinusoidal voltage to supply a 50 kW load. capacitor banks in the distribution systems.
Fig. 5 shows the current signal produced by the nonlinear load The transient phenomena produced by the switching of capac-
applied in feeder 4. itors appear in the three phases during a short period of time
Besides, in the study of distribution power systems protection, (0.53 cycles). Similar to motor starting, these transients have
it is also important to study the effect of motor operation on the high peak values with frequencies much greater than the power
system fundamental frequency. Besides, the transient phenom-
ena produced by capacitor switching may look similar to those
of HIFs in frequency domain, so it is necessary to test the reli-
ability of any HIF algorithm under this event. In the studied
distribution system, different possible instants of capacitor ener-
gisation have been taken into consideration. The importance of
transient instant application in this simulation is that the mag-
nitude and frequency of the transient current mainly depend on
the impedance of the actual circuit (system damping) and the
instant of energisation. The transients originated from the simula-
tion of this model are comparable with data previously published
[30].
Finally, in order to simulate normal load switching operation, a
three-phase time control circuit breaker has been used to connect
loads to the rest of the energised system at different instants. This
operation of connection may take several cycles to change the con-
tacts from open state to completely close state; hence, the transient
Fig. 5. Nonlinear load distorted current. waves may last up to two or three cycles.
1330 I. Baqui et al. / Electric Power Systems Research 81 (2011) 13251333

As an example, Fig. 7 shows a time domain representation of

the current signal and Fourier spectrum of a simulated HIF in the
feeder 4. This signal corresponds to a downed conductor to ground,
with a fault resistance of 200  and located 1.5 km away from the
substation busbar.

4. Results

In order to verify the developed methodology, a complete set

of cases have been tested. As an example, the following circuit
situations are presented in this section:

- High impedance fault: occurred in phase A of feeder 1, with a fault

resistance of 1.5 k. The fault location, arc voltage and incep-
tion time are considered to be 2 km from busbar, 4 kV and 0 s,
respectively.
- Low impedance fault: occurred in phase A of feeder 1, with a fault
resistance of 1 , located 2.7 km from the supplying busbar and
with 0 s of inception time.
- Capacitor switching: occurred in the busbar with 30 kVAr and
0.023 s of inception time.

Fig. 6. HIF model. The process starts by a de-noising process and the application of
DWT to the current signals. Then, the STD from all frequency levels
are calculated and used as the inputs to the trained ANN. Finally,
the state of a feeder is calculated according to the outputs of the
3.2.2. Low impedance fault situations
neural network.
The low impedance fault (LIF) model is quite simple and con-
The complete performance of the proposed technique has been
sists of one resistor connected to ground through a breaker. The
tested by its application to data under different conditions. The test
fault resistance values have been considered between 0 and 30 ,
set was formed by patterns from different situations compared to
while in the case of a high resistance fault, the corresponding val-
the training patterns. The test set patterns for normal events, LIFs
ues are higher than 200 . In this range of values, the overcurrent
and HIFs are 856, 10,139 and 14,898 patterns, respectively.
protection devices can detect the fault.

4.1. High impedance fault

3.2.3. High impedance fault situations
A new dynamic arc model has been considered, as a variant of After the denoising process, Fig. 8 shows the behavior of the
the model proposed in [4]. In this new model of HIF, a saw-tooth decomposed signals of the faulty phase current under a HIF situa-
wave shape with increasing and decreasing linearity has been used tion. The dominant Wavelet levels (high amplitude) are A7 and D7,
(Fig. 6). This model can represent the dynamic arc characteristics which represent the power frequency and sub-harmonic frequency
better than DC sources. It also allows the adjustment of the phase component. The high transient frequencies appear during the arc
difference between the applied voltage and fault current, by chang- period (arc re-ignition and extinction) which are clearly seen in the
ing the frequency of saw-tooth signal. Wavelet levels D1 to D5.

Fig. 7. HIF current spectrum.

 I. Baqui et al. / Electric Power Systems Research 81 (2011) 13251333 1331

Fig. 8. DWT of the faulty phase current signal of HIF.

Table 3 Table 5
Feature vector of a HIF case. Feeder state identication.

Levels Phase A Phase B Phase C Feeder ANN output Feeder state

T
D1 4.4e5 5.1e7 7.2e7 1 [0 1 0] LIF
D2 0.0003 4.1e6 7.9e6 2 [1 0 0]T Normal
D3 0.0010 2.0e5 3.7e5 3 [1 0 0]T Normal
D4 0.0050 6.2e5 9.1e5 4 [1 0 0]T Normal
D5 0.0100 0.0001 0.0002 5 [1 0 0]T Normal
D6 0.0500 0.0005 0.0007
D7 0.1000 0.0050 0.0050
A7 0.5000 0.0500 0.0500
4.2. Low impedance fault

Table 4 Following the same procedure, Fig. 9 shows the result of the
Feeder state identication. DWT decomposition applied. The peaks observed at the initial and
Feeder ANN output Feeder state nal edges in Wavelet levels, especially from D4 to D6, are caused
by discontinuities of data window, which have no effect on the
1 [0 0 1]T HIF
2 [1 0 0]T Normal developed method.
3 [1 0 0]T Normal Applying the STD in each decomposition level, the numerical
4 [1 0 0]T Normal values of patterns can be obtained from the analyzed signal. Using
5 [1 0 0]T Normal
these data as the input of the selected ANN, the result of the output
ANN simulation is shown in Table 5. It can be concluded that feeder
Applying the STD of each decomposition level, the numerical 1 is under LIF.
values of patterns can be obtained from the analyzed signal, as it is
shown in Table 3. 4.3. Normal situation
Applying these data matrix as the input to the selected ANN, the
result of the output ANN simulation is shown in Table 4. It can be Finally, Fig. 10 shows the behavior of DWT decomposition in
concluded that feeder 1 is under HIF. feeder 3 under a situation of capacitor switching (no fault).

Fig. 9. DWT of the faulty phase current signal of LIF.

1332 I. Baqui et al. / Electric Power Systems Research 81 (2011) 13251333

Fig. 10. DWT of capacitor switching signal.

Table 6 Nowadays, in order to test this methodology under real data,

Feeder state identication.
several contacts are being made with local utilities. This will be the
Feeder ANN output Feeder state next stage in the developing process of this methodology.
1 [1 0 0]T Normal
2 [1 0 0]T Normal
Acknowledgement
3 [1 0 0]T Normal
4 [1 0 0]T Normal
5 [1 0 0]T Normal The work developed in this paper has been partially supported
by the Basque Government (Ref. IT532-10).

The increasing or decreasing of the current values does not affect References
the results obtained by the method developed. This is due to the fact
[1] C.L. Huang, H.Y. Chu, M.T. Chen, Algorithm comparison for high impedance
that the high frequency part of the signal only appears during a very
fault detection based on staged fault test, IEEE Trans. Power Deliv. 3 (4) (1988)
short period of time, at the instant of capacitor switching. 14271435.
In this case, the proposed methodology provides the ANN output [2] A.A. Girgis, W. Chang, E.B. Makram, Analysis of high-impedance fault generated
data, shown in Table 6. It can be concluded that all feeders are under signals using a Kalman ltering approach, IEEE Trans. Power Deliv. 5 (4) (1990)
17141724.
normal operation. [3] B.D. Russel1, R.P. Chinchali, A digital signal processing algorithm for detecting
arcing faults on power distribution feeders, IEEE Trans. Power Deliv. 4 (1) (1989)
132140.
5. Conclusions [4] A.E. Emanuel, E.M. Gulachenski, High impedance fault arcing on sandy soil in
15 kV distribution feeders: contribution to the evaluation of the low frequency
spectrum, IEEE Trans. Power Deliv. 5 (2) (1990) 676686.
In this paper, a new HIF detection methodology based on the [5] A.F. Sultan, G.W. Swift, D.J. Fediechuk, Detection arcing downed wires using
combination of WT and ANN is presented. This methodology is an fault current icker and half cycle asymmetry, IEEE Trans. Power Deliv. 9 (1)
(1998) 461470.
alternative to conventional fault detection methods and its main
[6] B.D. Russell, R.P. Chinchali, C.J. Kim, Behaviour of low frequency spectra dur-
objective is to distinguish features of current waveforms produced ing arcing fault and switching events, IEEE Trans. Power Deliv. 3 (4) (1988)
by HIFs from those of normal switching events and LIFs. 14851492.
[7] B.D. Russell, K. Mehta, R.P. Chinchali, An arcing fault detection technique using
In this method, the input parameters are the sampled feeder
low frequency current components-performance evaluation using recorded
phase current under various transient conditions. Then, by DWT eld data, IEEE Trans. Power Deliv. 3 (4) (1988) 14931500.
analysis of the current waveforms and feature extraction, the time [8] T. Baldwin, F. Renovich, L.F. Saunders, Directional ground-fault indicator
and frequency domain information of these events are provided for high-resistance grounded system, IEEE Trans. Ind. Appl. 39 (2) (2003)
325332.
and used to categorize different faults. These data are sent to a [9] I. Zamora, A.J. Mazn, K.J. Sagastabeitia, J.J. Zamora, New method for detecting
previously dened ANN. The problem of selecting optimum net- low current faults in electrical distribution systems, IEEE Trans. Power Deliv.
work structures has been solved by using an automatic system 22 (4) (2007) 20722079.
[10] J.J. Zamora, I. Zamora, A.J. Mazon, K.J. Sagastabeitia, Optimal frequency value to
called SARENEUR. Since the topology changes are reduced in num- detect low current faults superposing voltage tones, IEEE Trans. Power Deliv.
ber where this methodology is expected to be applied, the amount 23 (4) (2008) 17731779.
of previously trained ANNs would be minimum. [11] A.F. Sultan, G.W. Swift, D.J. Fedirchuk, Detection of high impedance arcing
faults using a multi-layer perceptron, IEEE Trans. Power Deliv. 7 (4) (1992)
Additionally, the simulation of an accurate model for HIFs is dif- 18711877.
cult due to the different real variables involved. Therefore, a new [12] J.H. Ko, J.C. Shin, C.W. Ryu, C.G. Park, W.Y. Yim, Detection of high impedance
model has been proposed in which the parameters of the HIF can be faults using neural nets and chaotic degree, Intern. Conf. Energy Manage. Power
Deliv. 2 (1998) 399404.
changed to test the detection methodology under different surface
[13] L.A. Snider, Y.S. Yuen, The articial neural-networks-based relay algorithm for
conditions (level of dryness or wetness) and types of surfaces. the detection of stochastic high impedance faults, Neurocomputing 23 (1998)
To verify the performance of the methodology proposed, numer- 243254.
[14] M. Michalik, M. Ukowicz, W. Rebizant, L. Seung-Jae, K. Sang-Hee, New ANN-
ous cases have been analyzed. The results obtained have proved the
based algorithms for detecting HIFs in multigrounded MV networks, IEEE Trans.
correct operation (secure and reliable) of the developed method Power Deliv. 23 (1) (2008) 5866.
under various transient operation conditions in electrical dis- [15] S.J. Huang, C.T. Hsieh, High-impedance fault detection utilizing a morlet
tribution networks. This methodology gives the exact output wavelet transform approach, IEEE Trans. Power Deliv. 14 (4) (1999) 14011410.
[16] A. Lazkano, J. Ruiz, E. Aramendi, L.A. Leturiondo, Evaluation of a new proposal
corresponding to each HIF situation, and it is not confused by the for an arcing fault detection method based on wavelet packet analysis, Eur.
transients caused by LIFs and normal switching events. Trans. Electr. Power 14 (3) (2004) 161174.
 I. Baqui et al. / Electric Power Systems Research 81 (2011) 13251333 1333

[17] D. Chanda, N.K. Kishore, A.K. Sinha, Application of wavelet multiresolution anal- [23] A.H. Etemadi, M. Sanaye-Pasand, High-impedance fault detection using multi-
ysis for identication and classication of faults on transmission lines, Electr. resolution signal decomposition and adaptive neural fuzzy inference system,
Power Syst. Res. 73 (2005) 323333. IET Gener. Transm. Distrib. 2 (1) (2008) 110118.
[18] T.M. Lai, L.A. Snider, E. Lo, D. Sutanto, High-impedance fault detection using dis- [24] M. Michalik, W. Rebizant, M. Lukowicz, S.J. Lee, S.H. Kang, High-impedance fault
crete wavelet transform and frequency range and RMS conversion, IEEE Trans. detection in distribution networks with use of wavelet-based algorithm, IEEE
Power Deliv. 20 (1) (2005) 397407. Trans. Power Deliv. 21 (4) (2006) 17931802.
[19] N.I. Elkalashy, M. Lehtonen, H.A. Darwish, A-M.I. Taalab, M.A. Izzularab, [25] P.L. Mao, R.K. Aggarwal, A novel approach to the classication of the tran-
DWT-based detection and transient power direction-based location of high sient phenomena in power transformer using combined wavelet transform and
impedance faults due to learning trees in unearthed MV networks, IEEE Trans. neural network, IEEE Trans. Power Deliv. 16 (4) (2001) 654660.
Power Deliv. 23 (1) (2008) 94101. [26] Wavelet Toolbox Users Guide for Use with MATLAB, The Mathworks Inc., 2009.
[20] A.-R. Sedighi, M.-R. Haghifam, O.P. Malik, Soft computing applications in high [27] Neural Network Toolbox Users Guide for Use with MATLAB, The Mathworks
impedance fault detection in distribution systems, Electr. Power Syst. Res. 76 Inc., 2009.
(13) (2005) 136144. [28] A.J. Mazn, I. Zamora, K.J. Sagastabeitia, J.R. Saenz, Selecting ANN structures to
[21] P.R.S. Jota, F.G. Jota, Fuzzy detection of high impedance faults in nd transmission faults, IEEE Comput. Appl. Power 14 (3) (2001) 4448.
radial distribution feeders, Electr. Power Syst. Res. 49 (3) (1999) 169 [29] A.E. Emanuel, J. Janczak, Distribution feeders with nonlinear loads in the north-
174. east USA. I. Voltage distortion forecast, IEEE Trans. Power Deliv. 10 (1) (1995)
[22] Y. Sheng, S.M. Rovnyak, Decision tree-based methodology for high 340347.
impedance fault detection, IEEE Trans. Power Deliv. 19 (2) (2004) [30] D. Chan, T. Wai, X. Yibin, A novel technique for high impedance fault identi-
533536. cation, IEEE Trans. Power Deliv. 13 (3) (1998) 738744.