IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO.

1, JANUARY 2005 397

High-Impedance Fault Detection Using Discrete

Wavelet Transform and Frequency Range and
RMS Conversion
T. M. Lai, L. A. Snider, Senior Member, IEEE, E. Lo, Member, IEEE, and D. Sutanto, Senior Member, IEEE

Abstract—High-impedance faults (HIFs) are faults which are tems [9], [10], and artificial neural networks (ANN) [11]–[13]
difficult to detect by overcurrent protection relays. Various pattern are used to classify the fault based on the extracted features. It
recognition techniques have been suggested, including the use has also been proposed [14], [15] that HIF earth fault protec-
of Wavelet Transform [1]. However this method cannot indicate
the physical properties of output coefficients using the wavelet tion is more reliable in a five-wire distribution system. Pattern
transform. In this paper we propose to use the Discrete Wavelet recognition methods aim at detecting characteristic voltage and
Transform (DWT) as well as frequency range and rms conversion current distortions caused by arcing faults.
to apply a pattern recognition based detection algorithm for It is generally recognized [16] that, while each of these tech-
electric distribution high impedance fault detection. The aim niques have had some success at HIF fault detection, there is
is to recognize the converted rms voltage and current values
caused by arcs usually associated with HIF. The analysis using still no single method to solve the problem of the HIF identifi-
Discrete Wavelet Transform (DWT) with the conversion yields cation completely. For example, various frequency ranges have
measurement voltages and currents which are fed to a classifier been proposed for feature extraction, however for the higher
for pattern recognition. The classifier is based on the algorithm frequency ranges the instrument transformers may introduce
using nearest neighbor rule approach. It is proposed that this errors. Furthermore, there is a trade-off between the criteria
method can function as a decision support software package for
HIF identification which could be installed in an alarm system. of achieving a high degree of dependability (where the HIF
detector consistently detects high impedance faults when they
Index Terms—High-impedance faults (HIFs), pattern recogni-
tion, wavelet transforms. occur) and security (ensuring that the HIF detector does not de-
tect an HIF when there is no HIF). The aim of the research pre-
sented in this paper, as with other papers dealing with this sub-
I. INTRODUCTION ject, is to make a meaningful improvement in HIF detection.
The method proposed in this paper uses digital signal pro-
H IGH-IMPEDANCE faults (HIFs) are difficult to detect.
When a conductor such as a distribution line makes con-
tact with a poor conductive surface or substance the resulting
cessing, where a novel application of the Discrete Wavelet
Transform (DWT) is used for extracting the features of the
level of fault current is usually lower than the nominal current distorted waveforms caused by HIFs, and the Nearest Neighbor
of the system at the fault location. Therefore, conventional pro- Rule (NNR) is used for fault classification. After capturing the
tection relay system will not be able to detect the HIFs and trip voltage and current waveforms from power system simulations,
the protection relay. The failure of HIF detection leads to poten- they are analyzed by DWT. The DWT output coefficients are
tial hazards to human beings and potential fire hazards [2]. converted to RMS values in various frequency ranges, which
HIFs on electrical transmission and distribution networks in- are then fed to a NNR pattern classifier to determine the fault or
volve arcing and/or nonlinear characteristics of fault impedance nonfault situations. The method incorporates the statistical na-
which cause cyclical pattern and distortion. Therefore, the ob- ture of HIFs and fault locations, and only low order harmonics
jective of most detection schemes is to identify special features of voltage and current are required. These schemes can also
in patterns of the voltages and currents associated with HIFs. potentially be applied for on-line training and customization
In general identification techniques comprises two basic steps: using actual field HIF data and can function as a decision
feature extraction and pattern recognition (classification). support software package for HIF identification which could be
Various feature extractors have been proposed by researchers installed in an alarm system.
and protection engineers, based on fractal techniques [3], digital The actual dependability of a proposed fault detection scheme
signal processing [4], crest factor [5], wavelet transform in high can only be determined through extensive staged fault testing,
frequency noise patterns [6], dominant harmonic vectors [7], and security can only be determined after an extended time in
[8]. The Pattern recognition methods which include expert sys- service. For the purposes of developing the methods, however,
simulation provides a practicable alternative. In our research,
three representative distribution systems were simulated in the
Manuscript received April 11, 2003; revised September 26, 2003. This work
was supported in part by Hong Kong Polytechnic University and in part by the Matlab environment, and several thousand fault and nonfault
Research Grant Council of Hong Kong under Project PolyU 5109/01E. Paper contingencies were stochastically applied. Dependability was
no. TPWRD-00172-2003. evaluated through determination of the success rate of recog-
The authors are with the The Hong Kong Polytechnic University, Hong Kong,
China. nizing fault cases, and the security was determined through de-
Digital Object Identifier 10.1109/TPWRD.2004.837836 termination of the immunity of the method to confounding from
0885-8977/$20.00 © 2005 IEEE
398 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 1, JANUARY 2005

Fig. 1. Schematic diagram of discrete wavelet transform (DWT) and its rms conversion.

contingencies such as capacitor switching as well as linear and where scale function and wavelet function is de-
nonlinear load switching. termined by the selection of a particular mother wavelet
and the following equations.

II. DISCRETE WAVELET TRANSFORM (DWT) MODEL AND

MULTI-RESOLUTION ANALYSIS (MRA) (3)

Wavelet theory and its applications are rapidly developing

The corresponding relationship of scaling function and wavelet
fields in applied mathematics and signal analysis. The wavelet
function between two consecutive resolution levels and
transform is a tool that divides up data, functions or operators
are defined in (4) using Mallet algorithm [20] and the low-
into different frequency components, and then evaluates each
pass filter and the highpass filter can be calculated using
component with a resolution matched to its scale [17]. Discrete
Matlab.
Wavelet Transform (DWT) resolves the input signal to time,
scales, scale coefficients and wavelet coefficients. Multi-reso-
lution analysis (MRA) [18], [19], based on the study of or-
thonormal, compactly supported wavelet bases, is an extension
of DWT and develops representations of a sophisticated signal (4)
in terms of wavelet and scaling functions in different fre-
quency resolution levels.
One-dimensional wavelet was applied to this power system ap-
The schematic diagram of MRA is shown in Fig. 1. The
plication. The scales and scale coefficients and wavelet coef-
scaling coefficients (approximation) can be computed by taking
ficients do not reflect the physical properties in the frequency
the inner products of the function with the scaling basis:
analysis. Consequently, it is difficult to understand and relates
the coefficients of the wavelet analysis to measurements. There-
fore, the following sections introduce mappings from scale to
frequency ranges and from scale coefficients and wavelet coef-
(1) ficients to the rms values of their scale coefficients and wavelet
coefficients in each DWT resolution level, respectively.
The wavelet coefficients (details) can be computed by taking the
inner products of the function with the wavelet basis: A. Mapping From Scales to Frequency Ranges
In discrete wavelet transform, a resolution scale is commonly
used to represent the degree of resolution. The structure of DWT
(2)
in MRA is shown in Fig. 1. Two definitions of resolution level
LAI et al.: HIGH-IMPEDANCE FAULT DETECTION USING DISCRETE WAVELET TRANSFORM 399

are commonly used: ascending order Level from the finest res-
olution level (1) to the coarsest resolution level and de-
scending order from the finest resolution level to the
coarsest resolution level (0) where is total resolution level.
The relationship between Level and is defined as follows:

Level (5)

and the resolution levels in terms of Level and are listed at

the bottom of Fig. 1. The resolution scale in each
resolution level is defined as:

scaleLevel (6) Fig. 2. Relationship of coefficients, signal energy, and rms values.

In each resolution level, the input signal in the upper reso-

lution level is split into the approximation by a lowpass filter
and the detail by the highpass filter in the lower res-
olution level. Both of output approximation and detail signal
are then decimated by 2. Based on the Nyquist theorem (which
states that the highest frequency which can be accurately rep-
resented is less than one-half of the sampling rate), the max-
imum frequency of original signal sampled at Hz
is Hz. The first approximation and first detail
(10)
in resolution level 1 are sampled at half of . There-
fore, the maximum frequencies of signals and in
each resolution level Level are given in (7): where
is total resolution level in this multi-resolution anal-
(7) ysis,
are the approximation in level 0 and the detail in
and level respectively,
Since the boundary of a lowpass and a highpass filter is half
of nyquist frequency, the upper boundary frequency of lowpass is signal energy of the approximation in level 0,
filter and the lower boundary frequency of highpass filter is signal energy of the detail in level and,
is the same as half of . Therefore, the lower boundary is signal energy of the scale coefficients and wavelet
frequency and upper boundary frequency of coefficients .
both lowpass filter and highpass filter in each resolution For the periodic mode of discrete wavelet transform, the rms
level Level are defined as: values of their scale coefficients and wavelet coefficients can
be calculated directly from their scale coefficients and wavelet
for lowpass filter (8) coefficients in their own resolution levels or through their signal
energies indirectly. The relationship map is shown in Fig. 2. The
conversion equations (11), (12), and (13) are shown as follows
and the proof is in Appendix.
for highpass filter
The rms values for the details or approximation in each
wavelet level can be represented in the periodic-padding (peri-
(9) odic extension at the edges) mode of discrete wavelet transform.
For the scale coefficient in level , the rms value
in level is
B. Calculation of RMS Values From Scale Coefficients and
Wavelet Coefficients
Scale coefficients and wavelet coefficients and (11)
their signal energy representing the distorted signal
at different resolution levels in multi-resolution analysis
(MRA) are defined as:
where
;
number of points in ;
number of points in scale coefficient .
400 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 1, JANUARY 2005

For the wavelet coefficients in each level, the rms values

in level are

(12)

where

number of points in wavelet coefficients and

Level is the resolution level in discrete wavelet transform.
Therefore, the rms vector calculated from scale co-
efficients and wavelet coefficients of the distorted signal at
different resolution levels is as follows:

(13a) Fig. 3. Simulation model procedures—Matlab application.

(13b)

(13c)

With the above equations, the rms values calculated from

scale coefficients and wavelet coefficients in multi-resolution
levels can be converted directly from their scale coefficients
and wavelet coefficients or from their signal energy
. It is clear from (13b) and (13c) of rms values
in terms of their signal energy , that the rms values
calculated from their scale coefficients and wavelet
coefficients are directly proportional to their signal energy in
their resolution levels with a constant . Fig. 4. Simplified two-diode fault model of high-impedance faults.

III. SIMULATION of the transformer were then simulated using discrete wavelet
A. Simulation Procedures transform to analyze the frequency characteristics in various fre-
The simulation comprises three parts: power system model, quency ranges. The rms values of the voltages and currents were
wavelet transform and pattern recognition model. Fig. 3 shows calculated using the voltage and current waveforms from the
the flowchart of the simulation. Matlab was used to generate discrete wavelet transform. A classifier was used to recognize
all the system parameters in the fault cases and nonfault cases the fault cases and nonfault cases using rms values of voltage
for the power system simulation according to random fault lo- and current in various frequency bands.
cation and fault situations. The system parameters in fault and
nonfault cases were then imported into the power system simu- B. Power System Model—High Impedance Faults (HIFs) and
lation and the Power System Toolbox was used to determine the Low Impedance Faults (LIFs) Model
targeted voltage and current waveforms of the circuit breaker A simplified 2-diode model [21] of HIFs, shown in Fig. 4,
in the faulted distribution line. These waveforms of voltages was used in the simulation to represent the low frequency phe-
and currents in the measurement points on the distribution side nomena typical of an arcing fault involving sandy soil. The
LAI et al.: HIGH-IMPEDANCE FAULT DETECTION USING DISCRETE WAVELET TRANSFORM 401

Fig. 6. Schematic diagram of the simulated 25 kV single branch power system

network.

Fig. 5. Typical current of high-impedance faults.

TABLE I
SCALE TO FREQUENCY CONVERSION BASED ON 50 Hz POWER FREQUENCY

Fig. 7. Schematic diagram of the simulated 25 kV radial power system

network.

model comprises two DC sources, and , which repre-

D. Pattern Recognition
sent the inception voltage of air in soil and/or between trees and
the distribution line. The two resistances, and , repre- Since normal operation and fault situations do not have spe-
sent the fault resistance: unequal values allow for asymmetric cific patterns for their respective distributions, a nonparametric
fault currents to be simulated. When the phase voltage is greater approach in supervised learning was used in this fault classifi-
than the positive DC voltage , the fault current flows toward cation problem. The nearest neighbor rule (NNR) method [22],
the ground. The fault current reverses when the line voltage is [23] is a typical recognition method of the nonparametric ap-
less than the negative DC voltage . For values of the phase proach. It is different from the black-box types of pattern recog-
voltage between and no fault current flows. The typ- nition methods, such as neural networks, since NNR can demon-
ical fault currents with the HIF model parameters are shown in strate the classification results and their decision boundaries be-
Fig. 5. tween two sorts of data by plotting two-dimensional contour
The low impedance fault (LIF) model comprises the appro- graphs.
priate fault resistance switched at the fault location in the distri- The root-mean-square values of voltages and cur-
bution line. rents in various frequency ranges from the DWT are classified
into a training set and a test set respectively. 95% of the rms
C. Implementation of Discrete Wavelet Transform (DWT) values were used as a training set to train the decision bound-
Model and Multi-Resolution Analysis (MRA) aries using NNR and the rest of the data acted as a test set to
validate their decision boundaries.
In this application, db4 in Daubechies family wavelets was
selected for the mother wavelet and the frequency bandwidths
IV. SYSTEM DESCRIPTION
of scale and wavelet coefficients of MRA were determined using
(8) and (9) shown in Table I after downsampling the input signal Three systems studied in this paper are parts of representative
to the sampling frequency of 9600 Hz in order to reduce the 25 kV power distribution networks: i) a distribution network
computational time. with a single branch of a nonlinear load, ii) a radial distribution
After the calculation of output voltage and current in various network with three branches of nonlinear loads and iii) a meshed
resolution levels from DWT, the rms values of volt- network with two sources and various nonlinear loads. Their
ages or currents are exported to a classifier. schematic diagrams are shown in Fig. 6, 7, and 8, respectively.
402 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 1, JANUARY 2005

Fig. 8. Schematic diagram of the simulated 25 kV meshed power system

network.

In the single branch distribution network, power is supplied

at 25 kV through a 20 km distribution line from a 5 MVA trans-
former. The network then distributes power to a 4 MVA linear
load with power factor 0.8 and a 1 MVA nonlinear load, where
a 6 pulse converter is used to represent the nonlinear load. In
the radial distribution network with three branches of nonlinear
loads, two additional branches, one with a 2 MVA 6-pulse con-
verter load and one with a 4.262 MVA 12-pulse converter load
are connected. The meshed network comprises a ring with two
3-phase voltage sources feeding two individual nonlinear loads,
one with a 1 MVA 6-pulse converter load and one with a 4.262 Fig. 9. Simulated waveforms and measured waveform containing
MVA 12-pulse converter load. The measurement points for the high-impedance faults. (a) Simulated waveforms containing high-impedance
voltage and current are located at the breaker as shown in the faults (prefault t < 0:013 and postfault t > 0:013 s). (b) Measured current
waveform containing high-impedance faults (prefault t < 0 and postfault
schematic diagrams. t > 0 s).
In this paper, 1000 nonfault cases and 1000 fault cases in
each system were simulated. The fault cases comprise both high
impedance fault model and low impedance fault models. In the other characteristics of the HIFs. The ratings of the linear loads
simulations, the significant parameters were stochastically se- and the nonlinear loads, operation schedule in the capacitors at
lected based on those in real operations. The fault currents were the sending end and receiving end in the distribution lines of
assumed to range from 0.25 to 2 p.u. of overcurrent device set- the simulated distribution networks were chosen randomly.
tings in the full load situation.
1) Case I—Non-Fault Cases (Normal Operation): In all V. SIMULATION RESULTS
distribution networks, shunt capacitors banks ranging up to 1
MVAr, linear loads up to their own maximum powers and non- A. Comparison of Simulation and Measurement of HIF
linear loads up to their own maximum ratings were switched The voltage and current waveforms at the sending end of the
at the sending end and receiving end of the distribution lines. transmission line of the network shown in Fig. 6 are shown in
The nonlinear loads comprise three phase 6 pulse converters Fig. 9(a). While field tests for this particular network are not
and 12-pulse converters and the three phase linear loads have available, it is instructive to make a qualitative comparison with
their own power factor in accordance with their system config- current waveforms corresponding to an HIF fault in a typical
urations. The ratings and breaker closing times of linear loads, distribution network. Such a waveform is shown in Fig. 9(b),
nonlinear loads and capacitors were stochastically selected. which was extracted form [4], and it compares quite well with
2) Case II—Fault Cases: The modeled phase-to-ground that of Fig. 9(a).
HIFs and LIF’s with various fault inception angles were simu-
lated at different positions along the faulted distribution lines
B. Simulation Results of DWT
from the circuit breaker at the source side to distribution loads.
Fault location, fault impedance and fault inception angles of The major simulation results of voltage and current in various
both HIFs and LIF’s were stochastically selected in the simula- frequency ranges in phase R at the circuit breaker are shown in
tion, representing the stochastic variation of soil properties and Figs. 10 to 14. Correlation scatter plots with decision boundaries
LAI et al.: HIGH-IMPEDANCE FAULT DETECTION USING DISCRETE WAVELET TRANSFORM 403

Fig. 10. Voltage (rms from d in Level 5: 150–300 Hz) against Voltage (rms Fig. 13. Voltage (rms from d in Level 5: 150–300 Hz) against Voltage (rms
from c in Level 6: 0–75 Hz). from d in Level 6: 75–150 Hz).

Fig. 11. Current (rms from d in Level 4: 300–600 Hz) against Voltage (rms
from d in Level 6: 75–150 Hz). Fig. 14. Current (rms from d in Level 5: 150–300 Hz) against voltage (rms
from d in Level 5: 150–300 Hz).

From the simulation results, the rms voltage from in res-

olution level 6 (dc to 75 Hz) under normal operation is close to
that of the phase voltage source. However, in the fault cases the
rms voltage from between 0 to 75 Hz has a range of approx-
imately zero to 13 kV, as shown in Fig. 10.
The range of voltage in the frequency range between 75 Hz
to 150 Hz from in resolution level 6 is around zero to 3.5
kV, as shown in Figs. 11, 12, and 13. The maximum currents
in the frequency range between 150 Hz to 300 Hz from in
resolution level 5 and between 300 Hz to 600 Hz from in
resolution level 4 are 8 A and 5 A respectively in the normal
operation and 60 A and 45 A respectively in the fault cases, as
shown in Figs. 11, 12, and 14. The maximum currents in the
above frequency ranges in the fault cases are 7 to 9 times the
magnitude of those in normal operation.
Fig. 12. Current (rms from d in Level 5: 150–300 Hz) against voltage (rms In the pattern recognition, the fault cases and nonfault cases
from d in Level 6: 75–150 Hz). are classified into the training set and test set in the simulation.
In the pattern recognitions lower probability errors mean higher
are used to present the correlation of each frequency band of successful rates to classify the two above cases. The ranges of
voltage and current components in accordance with Table I. errors in the training set in various correlated wavelet levels are
404 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 1, JANUARY 2005

the total error of the fault detection is minimized, or a trade-off

may be made between dependability and security, depending on
the specific application.

VI. CONCLUSION
This paper presented a study of the fault classification in 25
kV electrical distribution systems based on discrete wavelet
transform. The study involved computer simulation of power
systems, discrete wavelet transform and NNR classification.
The electrical faults including HIFs and common faults are
stochastic in nature, and depend on factors such as fault lo-
cation, fault impedance, fault inception angle, other electrical
loads, etc. A stochastic simulation was performed, and the
error probabilities of classification between the fault cases and
normal operation were determined. The classifier had a high
success rate at distinguishing between fault cases and nonfault
Fig. 15. Type I, II, and total error of different wavelet combinations. cases, including switching of capacitor banks and nonlinear
loads.
The method presented in this paper overcomes the difficulty
between 2.59% and 45.41%. The errors corresponding to rms
of using discrete wavelet transform that the output scale coef-
values of voltage wavelet coefficients in level 6 are be-
ficients and wavelet coefficients do not represent any physical
tween 2.59% and 16.83%.
properties. Using the relationship among scale coefficients and
For the errors in the test set, the ranges of errors in various cor- wavelet coefficients, signal energies and rms values, scale coef-
responding wavelet coefficients are between 1.83% and 45.33%. ficients and wavelet coefficients can be converted to rms values
The errors corresponding to rms values of voltage wavelet co- directly or through the calculation of signal energies. There-
efficients in level 6 are also between 1.83% and 16.50. fore, the clear distribution patterns among various characteristic
the average overall test error is approximately the same as the voltage and current rms values calculated from scale coefficients
overall training error. and wavelet coefficients are demonstrated.
The total errors show average errors of each combination of The NNR method performs best as a two-parameter input
wavelet coefficients. The errors corresponding to rms values of classifier to provide visual pictures of the analysis related to
voltage wavelet coefficients in level 6 are still approxi- the identification of HIFs. This is demonstrated by the decision
mately between 2.52% and 16.80%. The range of total errors is boundaries of the two-dimensional scatter plots. However,
from 2.52% and 45.4%. especially in the significant overlapping areas of fault cases
An important observation is that it is sufficient to consider and normal cases, recognition and analysis by experienced
only voltages in the frequency range from 0 to 300 Hz, and engineers is required. However, the identification algorithm of
currents in the frequency range of 0 to 600 Hz for the detec- DWT can be applied to produce a decision support software
tion of high impedance faults. Therefore, voltages and currents package for HIF identification which could be installed as an
with low frequency ranges are major factors to classify the high alarm system. For example, the package can be installed in the
impedance faults and the common faults. supervision control and data acquisition system as a plug-in
Dependability and security are the important issues of the HIF component for high impedance fault identification.
fault detection. A high level of dependability occurs when the
HIF detector correctly recognizes fault cases on its feeder. A APPENDIX
high level of security occurs when the HIF detector does not
misclassify the nonfault cases into the fault cases. The depend- A. Proof of (11) to 13
ability and security can be evaluated using type I and II errors,
This device of is based on periodic-padding (periodic exten-
where type I and II errors are the percent of incorrect classifica-
sion at the edges) mode of Discrete Wavelet Transform,
tion for fault and nonfault cases respectively. Fig. 15 summaries
type I, II, and total errors corresponding to different wavelet
combinations. For example, the type I, II errors and total error
of the combination of current and voltage are 2%,
3%, and 2.52%, respectively, and the scatter plot of this com-
bination shows in Fig. 13. Suppose that a lower error for fault
cases is required in order to achieve better dependability, the In multiresolution analysis, a set of nested subspaces and
decision boundary of the classifier can be adjusted. However, as are defined as
expected, this will cause an increase of the classification error
for the nonfault cases which reduces the security of the fault
detection. As a consequence, the HIF classifiers could be opti-
mized to balance between dependability and security such that
LAI et al.: HIGH-IMPEDANCE FAULT DETECTION USING DISCRETE WAVELET TRANSFORM 405

Therefore, a input signal can be decomposed into their The rms value for the approximation in the resolution
subset signals and in accordance with the subsets level 1 in terms of scale coefficient or its own signal
and respectively as follows [19]: energy is defined as

(A.1)

(A.2)
(A.3)

where . (A.4)
The rms value in level can be calculated
from the scale coefficient in level as follows: where

(A.1) number of points in and

number of points in scale coefficient .
Similarly, the rms value in level can be calculated from
the wavelet coefficients in level as follows:

(A.2)

By orthonormal basis of scaling functions with different trans-

lations

By orthonormal basis of wavelet functions with different scales

and translation
for all integer where

By orthonormal basis of scale functions

for all By orthonormal basis of wavelet functions

406 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 20, NO. 1, JANUARY 2005

The rms value for the detail in the resolution level in [8] D. I. Jeerings and J. R. Linders, “A practical protective relay for down-
terms of wavelet coefficients or their own signal energies is de- conductor faults,” IEEE Trans. Power Del., vol. 6, no. 2, pp. 565–574,
Apr. 1991.
fined as [9] C. J. Kim and B. D. Russell, “Classification of faults and switching
events by inductive reasoning and expert system methodology,” IEEE
Trans. Power Del., vol. 4, no. 3, pp. 1631–1637, Jul. 1989.
[10] B. D. Russell and C. L. Benner, “Arcing fault detection for distribu-
tion feeders: Security assessment in long term field trials,” IEEE Trans.
Power Del., vol. 10, no. 2, pp. 676–683, Apr. 1995.
[11] S. Ebron, D. L. Lubkeman, and M. White, “A neural network approach
(A.5) to the detection of incipient faults on power distribution feeders,” IEEE
Trans. Power Del., vol. 5, no. 2, pp. 905–914, Apr. 1990.
[12] L. A. Snider and Y. S. Yuen, “The artificial neural networks based relay
for the detection of stochastic high impedance faults,” Neurocomput.,
vol. 23, pp. 243–254, 1998.
[13] A. M. Sharaf, L. A. Snider, and K. Debnath, “A neural network based
relaying scheme for distribution system high impedance fault detection,”
(A.6) in Proc. 1st New Zealand Int. Two-Stream Conf. Artificial Neural Net-
works Expert Systems, 1993, pp. 321–324.
[14] T. A. Short, J. R. Stewart, D. R. Smith, J. O’Brien, and K. Hampton,
where “Five-wire distribution system demonstration project,” IEEE Trans.
Power Del., vol. 17, no. 2, pp. 649–654, Apr. 2002.
[15] D. J. Ward, J. F. Buch, T. M. Kulas, and W. J. Ros, “An analysis of the
five-wire distribution system,” IEEE Trans. Power Del., vol. 18, no. 1,
number of points in wavelet coefficients and pp. 295–299, Jan. 2003.
Level is the resolution level in discrete wavelet transform. By [16] J. Tengdin, R. Westfall, and K. Stephan, “High Impedance Fault Detec-
(A.3) to (A.6), the rms vector respected to scale coef- tion Technology,” PSRC Working Group Members, Rep. PSRC Working
Group D15.
ficients and wavelet coefficients of the distorted signal at [17] I. Daubechies, Ten Lectures on Wavelets. Philadelphia, PA: Soc. In-
different resolution levels is as follows: dustrial Appl. Math., 1992.
[18] G. Strang and T. Nguyen, Wavelets and Filter Banks. Cambridge, MA:
Wellesley-Cambridge, 1996.
[19] C. S. Burrus, R. A. Gopinath, and H. Guo, Introduction to Wavelets and
Wavelet Transform: A Primer. Englewood Cliffs, NJ: Prentice-Hall,
1998.
[20] S. G. Mallat, “A theory for multiresolution signal decomposition: The
wavelet representation,” IEEE Trans. Pattern Anal. Mach. Intell., vol.
11, no. 7, pp. 674–693, Jul. 1989.
[21] A. E. Emanuel, D. Cyganski, J. A. Orr, S. Shiller, and E. M. Gulachenski,
“High impedance fault arcing on sandy soil in 15 kV distribution feeders:
Contributions to the evaluation of the low frequency spectrum,” IEEE
Trans. Power Del., vol. 5, no. 2, pp. 676–686, Apr. 1990.
[22] M. Nadler and E. P. Smith, Pattern Recognition Engineering. New
York: Wiley, 1993.
[23] R. O. Duda, P. E. Hart, and D. G. Stoke, Pattern Classification. New
York: Wiley, 2001.

T. M. Lai received the B.Sc. degree in Engineering

from Hong Kong Polytechnic University in 1996 and
the Master’s degree in Philosophy from the Univer-
sity of Hong Kong, Hong Kong, China, in 2000. He
is currently pursuing the Ph.D. degree in the Depart-
ment of Electrical Engineering at Hong Kong Poly-
REFERENCES technic University.
His research areas include wavelet and artificial
[1] S. J. Huang and C. T. Hsieh, “High-impedance fault detection utilizing intelligence, harmonics, and transient in power sys-
a Morlet wavelet transform approach,” IEEE Trans. Power Del., vol. 14, tems.
no. 4, pp. 1401–1410, Oct. 1999.
[2] Detection of Downed Conductors on Utility Distribution Systems, 1989.
IEEE Tutorial Course Text, no. 90EH0310-3-PWR.
[3] B. D. Russell and C. L. Benner, “Analysis of high impedance faults
using fractal techniques,” IEEE Trans. Ind. Applicat., vol. 33, no. 3, pp. L. A. Snider (M’66–SM’82) was born in Montreal,
635–640, May/Jun. 1997. QC, Canada. He graduated from the Department
[4] A. A. Girgas, W. Chang, and E. B. Makram, “Analysis of high- of Electrical Engineering at McGill University,
impedance fault generated signals using a Kalman filtering approach,” Montreal, and received the M.Sc. and Ph.D. degrees
IEEE Trans. Power Del., vol. 5, no. 4, pp. 1714–1724, Oct. 1990. from the University of Birmingham, Birmingham,
[5] C. J. Kim and B. D. Russell, “Analysis of distribution disturbances and U.K.
Arcing faults using the crest factor,” Elect. Power Syst. Res., vol. 35, pp. Currently, he is a Senior Lecturer in the Elec-
141–148, 1995. trical Engineering Department of the Hong Kong
[6] D. T. W. Chan and X. Yibin, “A novel technique for high impedance fault Polytechnic University. He has an extensive back-
identification,” IEEE Trans. Power Del., vol. 13, no. 3, pp. 738–744, Jul. ground in electric utility research and was one of
1998. the founders of the simulation complex of the Hydro
[7] B. M. Aucoin and B. D. Russell, “Detection of distribution high Quebec Research Institute, QC, Canada. His interests include real-time power
impedance faults using burst noise signals near 60 Hz,” IEEE Trans. system simulation and high-voltage engineering, and computer applications in
Power Del., vol. PWRD-2, no. 2, pp. 342–348, Apr. 1987. teaching.
LAI et al.: HIGH-IMPEDANCE FAULT DETECTION USING DISCRETE WAVELET TRANSFORM 407

Edward Lo (M’83) received the B.Sc.(Eng.), Danny Sutanto (SM’89) received the B.Eng. and
M.Phil., and Ph.D. degrees in electrical engineering Ph.D. degrees from the University of Western
from the University of Hong Kong, Hong Kong, Australia, Perth, Australia, in 1978 and 1981,
China, in 1983, 1986, and 1996, respectively. respectively.
Currently, he is an Assistant Professor in the Currently, he is a Professor of electrical engi-
Electrical Engineering Department of the Hong neering with the Hong Kong Polytechnic University,
Kong Polytechnic University, Hong Kong, China. Kowloon, Hong Kong. He was also a Power System
He has more than 15 years of mixed experience in Analyst with GEC Projects, Sydney, Australia. His
teaching, research, industry, and consultancy. His main areas of research are power system analysis,
research interests include power quality, renewable power system economics, voltage stability, har-
energy, drives, and traction. monics, power electronics, and computer-aided
education.
Dr. Sutanto is the Executive Committee Member of the Joint Chapter of
PES, IAS, and PELS, IEEE HK Section and currently is the IEEE PES Region
10, Regional Representative. He was awarded the Noel Svennson’s Award for
Teaching Excellence in 1994 with the School of Electrical Engineering at the
University of New South Wales, Sydney, Australia.