Electric Power Systems Research 53 (2000) 23 – 30

www.elsevier.com/locate/epsr

Wavelet analysis based scheme for fault detection and

classification in underground power cable systems
W. Zhao a, Y.H. Song a,*, Y. Min b
a
Department of Electrical Engineering and Electronics, Brunel Uni6ersity, Uxbridge UB8 3PH, UK
b
Department of Electrical Engineering, Tsinghua Uni6ersity, Bejing 100084 , People’s Republic of China

Received 6 August 1998; accepted 22 December 1998

Abstract

This paper presents a new method for detecting and classifying fault transients in underground cable systems based on the use
of discrete wavelet transform. A 400 kV underground cable system is simulated by ATP/EMTP (electro-magnetic transients
program) under various system and fault conditions. Daubechies eight wavelet transform is employed to analyze fault transients
for the development of a novel fault detection and classification scheme. Simulation results show that this scheme is effective and
robust. © 2000 Elsevier Science S.A. All rights reserved.

Keywords: Wavelet analysis; Fault detection; Underground power cable systems

1. Introduction it does not provide multiple resolution in time and

frequency because the window is fixed. As the fre-
For various reasons, there is an increasing tendency quency is increased, there are more and more cycles
towards the use of underground cables at all voltage within the windows. Thus, individual frequency compo-
levels. This inevitably presents additional problems to nents are not treated in the same way. Wavelets, how-
power transmission protection since the underground ever, can be chosen with very desirable frequency and
cables have different electrical properties from those of time characteristic as compared to Fourier techniques.
the overhead lines. No matter how well the cables are In contrast to the STFT, which uses a single analysis
designed and manufactured, underground cable systems window, the wavelet transform uses short windows at
are prone to numerous hazards which may lead to their high frequencies and long windows at low frequencies.
failure. In most of the power system relay schemes, the Thus the windowing of wavelet transform is adjusted
first step always involves fault detection and classifica- automatically for low or high frequencies and each
tion. The information provided by this step is essential frequency component gets treated in the same manner
for the fault location scheme. without any reinterpretation of the results.
It is widely known that faults on underground cable The wavelet analysis procedure is to adopt a wavelet
systems produce a wide range of frequency compo- prototype function, called a ‘mother wavelet’. Temporal
nents. Previous studies on distorted waveform gener- analysis is performed with a contracted, high frequency
ated by cable faults mainly use Fourier Transform [1]. version of the prototype wavelet, while frequency anal-
The basis functions used in Fourier analysis, sine waves ysis is performed with a dilated, low frequency version
and cosine waves, are precisely located in frequency, of the prototype wavelet. Wavelet has been successfully
but exist for all time. Although the short-time Fourier applied in several fields, such as image and signal
Transform (STFT) which windows the input signal analysis, radar, acoustics, and mechanical vibrations.
overcomes the time location problem to a large extent, Recently, several papers have been presented proposing
the use of wavelets for power system analysis [2–8].
* Corresponding author. Fax: +44-1895-258728. The aim of this paper is to investigate the application
E-mail address: [email protected] (Y.H. Song) of wavelets in cable fault transients analysis and thus to

0378-7796/00/$ - see front matter © 2000 Elsevier Science S.A. All rights reserved.
PII: S 0 3 7 8 - 7 7 9 6 ( 9 9 ) 0 0 0 3 3 - 4
24 W. Zhao et al. / Electric Power Systems Research 53 (2000) 23–30

provide a new concept on cable faults detection and simply choose a0 = 2 and b0 = 1, a dyadic-orthonor-
classification, based on some initial results reported in mal wavelet transform is obtained. With this choice,
ref. [8]. Section 2 contains a brief introduction to there exists an elegant algorithm, the multiresolution
wavelet and the discrete wavelet transform adopted in signal decomposition (MSD) technique [11], which
the paper. The results presented in this paper relate can decompose a signal into levels with different time
to a typical 400 kV underground cable system, the and frequency resolution. At each level j, approxima-
faulted responses of which are attained using ATP. In tion and detail signals Aj, Dj can be built. The words
Section 3, fault transients are analyzed by using ‘approximation’ and ‘detail’ are justified by the fact
wavelet transform. Then a novel fault detection and that Aj is an approximation of Aj − 1 taking into ac-
classification scheme based on discrete wavelet trans- count the ‘low frequency’ of Aj − 1, whereas the detail
form analysis, together with some test results, is pre- Dj corresponds to the ‘high frequency’ correction. The
sented in Section 4. original signal can be considered as the approxima-
tion at level 0.
The coefficients C ( j,k) generated by the DWT are
2. Theory of wavelet analysis something like the ‘resemblance indexes’ between the
signal and the wavelet. If the index is large, the re-
Wavelets are functions that satisfy certain require- semblance is strong, otherwise it is slight. The signal
ments. The very name wa6elet comes from the re- then can be represented by its DWT coefficients as
quirement that they should integrate to zero, ‘waving
’ above and below the x-axis. The diminutive conno- s(n)= % % C( j, k)gj, k (n) (3)
jN kZ
tation of wa6elet suggests the function has to be well
localized. Other requirements are technical and When fix j and sum on k, a detail Dj is defined as
needed mostly to insure quick and easy calculation of
the direct and inverse wavelet transform. Compared Dj (n)= % C( j, k)gj,k (n) (4)
with traditional Fourier method, there are some im- kZ
portant differences between them. First Fourier basis
functions are localized in frequency but not in time Then sum on j,the signal is the sum of all the details
while wavelets are localized in both frequency (via
dilation) and time (via translation). Moreover, s(n)= % Dj (n) (5)
jN
wavelets can provide multiple resolution in time
and frequency. Second, many classes of functions Take a reference level called J, there are two sorts
can be represented by wavelets in more compact of details. Those associated with indices j5 J corre-
way. For example, functions with discontinuities spond to the scales 2 j 5 2J, which are the fine details.
and functions with sharp spikes usually take substan- The others, which correspond to j \J, are the coarser
tially fewer wavelet basis functions than sine-cosine details. If these latter details are grouped into
basis functions to achieve a comparable approxima-
tion. Aj = % Dj (6)
There are many types of wavelets [9,10], such as j\J

Harr, Daubechies 4, Daubechies 8, Coiflet 3, Symmlet which defines an approximation of the signals. Con-
8 and so on. One can choose between them depend- nect the details and an approximation, the equality
ing on a particular application. As with the discrete
Fourier transform, the wavelet transform has a digi-
tally implementable counterpart, the discrete wavelet s= AJ + % Dj (7)
j5J
transform (DWT). If the ‘discrete’ analysis is pursuing
on the discrete time, the DWT is defined as which signifies that s is the sum of its approximation
AJ and of its fine details. The coefficients produced
C( j, k)= % s(n)gj, k (n) (j  N, k  Z) (1) by DWT, therefore, can be divided into two cate-
nZ gories: one is detail coefficient, the other is approxi-
mation coefficient. To obtain them, MSD provides an
where, s(n) is the signal to be analyzed and gj,k (n) is efficient algorithm known as a two channel sub-band
discrete wavelet function, which is defined by coder using quadrature mirror filters [12]. Then the
gj, k (n)= a − j/2
g(a − j detail part is still represented by wavelets, which can
0 0 n −kb0) (2)
be regarded as series of band-pass filters, whereas the
Select a0 and b0 carefully, the family of scaled and approximation is represented by the dilation and
shifted mother wavelets constitute an orthonormal ba- translation of a scaling function, which can be re-
sis of l 2(Z) (set of signals of finite energy). When garded as a low-pass filter.
 W. Zhao et al. / Electric Power Systems Research 53 (2000) 23–30 25

Fig. 3. Section connection of a crossbonded cable.

Fig. 1. Cross-section. tion and identification. Since Daubechies 8 is localized,

i.e. compactly supported, in time, it is good for short
3. Wavelet analysis of fault transients and fast transients analysis. After examinations of sev-
eral types of wavelet, Daubechies 8 is chosen in this
3.1. Cable system simulation scheme.
Fig. 4 illustrates the detailed wavelet analysis results
Single-conductor coaxial cable (Fig. 1) is generally of faulty phase-a current for a phase-a to ground fault,
used for EHV applications because they are the way of which happened at the left 200 m of section 2 (D=
carrying the necessary load current and maintaining the 0.125 pu) with the fault resistance (R) of 5 V. The three
insulation requirements. The underground cables mod- phase current waveforms are original signals of Bus1
eled in this paper are based on 400 kV oil-filled cables. simulated by ATP. Each original signal has 2048 sam-
Fig. 2 illustrates a model system. A 400 kV source ple points, which means the total simulation time dura-
with the frequency of 50 Hz is applied to the system. tion is 51.2 ms with a sampling time of 0.025 ms. Since
Three phase balanced loads are presented in the system. 2048= 211, when they are decomposed into Daubechies
The cables are buried in a horizontal (flat) configura- 8 wavelet components, there are 11 levels. From the
tion 1 m below the surface. The 4.8 km long cable line results of an extensive series of studies using wavelet
is split into four of 1.2 km major sections, each of analysis, the time and frequency localization character-
which consists of three minor sections. At the end of istics of wavelet decomposition of fault transients asso-
each minor section, the core of phase ‘a’ is connected to ciated with different fault conditions have been
the core of phase ‘b’, the core of phase ‘b’ is connected obtained. It is well known [13] that when a fault occurs
to the core of phase ‘c’, and core of phase ‘c’ is on an underground cable, different frequency compo-
connected to the core of phase ‘a’(Fig. 3). The cables nents are produced. The wavelet levels to be selected
are modeled using the cable constants routine of the must best reflect the fault characteristics under various
ATP and the parameter is simulated using J. Marti, system and fault conditions. In this respect, according
frequency dependent traveling wave model. Based on to the analyses of different wavelet levels of current
this cable system, transients of four types of fault waveform as shown in Fig. 4, the level 4 (D4) and level
(single-phase to ground, two- phase fault, two-phase to 9 (D9) details, are utilized to extract some useful fea-
ground and three-phase fault) under different fault tures. This is because the level 4 details generally reflect
conditions are simulated by ATP. the dominant non-frequency transient generated by
faults. Since level 9 details contain most of the funda-
3.2. Transients analysis mental harmonic, which is of 50 Hz in this system, the
sum of three phase of them (D9a, D9b, and D9c) have
By using wavelets analysis, sub-band information can similar characteristics of zero component which can be
be extracted from the simulated transients, which con-
used to differ phase-to-ground fault and phase-phase
tain useful fault features. By analyzing these features of
fault, two-phase to ground fault and three-phase fault.
the detail signals, different types of fault can be de-
These two level details of three phase currents for an
tected and classified. As mentioned earlier, the choice of
a-phase-to-ground fault are presented in Fig. 5. Also
analyzing wavelets plays a significant role in fault detec-
shown is the sum of three phase level 9 details, D9.
Figs. 6–8 present wavelet analysis results of other types
of fault, which are two-phase fault, two phase to
ground fault and three phase fault, respectively. It can
be seen that the level 4 detail signals in Figs. 5–8 are
very small during the normal operation time. The
Fig. 2. Configuration of a 400 kV cable system. spikes which are seen at the initial and final edges are
26 W. Zhao et al. / Electric Power Systems Research 53 (2000) 23–30

caused by discontinuities or edge effects inherent in the

analysis process that can be easily eliminated by using
proper window scheme. When a fault happened, several
spikes can be seen in the level 4 signals, which is due to

Fig. 5. Wavelet analysis of three phase currents for an A –G fault:

D =0.125 pu, R =5 V.

the high harmonics generated in the fault conditions.

By detecting the fault generated spikes, fault can be
detected. Figs. 5–8 show some differences among these
four types of fault as well. It is apparent that these
faults have their own characteristics in the level 4
details of three phase (D4a, D4b, D4c) and the sum of
three phase level 9 (D9) details, which can be success-
fully used to classify different types of fault, as will be
seen in next section.

4. Fault detection and classification scheme

Fig. 4. Wavelet analysis of phase-a current for an A–G fault: After analyzing numerous fault transients simulated
D =0.125 pu, R= 5 V. by ATP using wavelet transform, several fault related
 W. Zhao et al. / Electric Power Systems Research 53 (2000) 23–30 27

features have been identified. Based on these informa-

tion, a novel fault detection and classification scheme as
shown in Fig. 9 is presented, which has been summa-
rized as the following 20 rules.
1. Variable initialization.
2. Calculate three phase level 4 details:
D4p, p(a, b, c)
3. For any one of the three phases, if there are three
consecutive absolute values (excluding the spikes in
the initial and final edges) of level-4 details are
greater than t1, claim a fault. Otherwise move
analysis window and go to step (1).
4. Calculate the maximum absolute value (excluding
the spikes in the initial and final edges) of level 4
details in each phase:
xp = max(D4p(i)), p= {a, b, c}

Fig. 7. Wavelet analysis of three phase currents for an A – B– G fault:

D =0.125 pu, R =5 V.

Then, obtain the maximum one of the three: x=

max (xa, xb, xc).
5. Check which phase x belongs to. That is, if x=xp,
let k= ip, ip is the index of xp in vector D4p, p
stands for phase a, b or c.
6. Let

Ep(n)= D4p(k+n− 1)

where p = {a, b, c}, n= 1, 2, 3. Calculate the sum

of Ep (1), Ep (2), and Ep (3):
5
Fig. 6. Wavelet analysis of three phase currents for an A–B fault: ep = % Ep(n)
D =0.125 pu, R= 5 V. n=1
28 W. Zhao et al. / Electric Power Systems Research 53 (2000) 23–30

p (a, b, c)
7. Let e=max (ea, eb, ec). Calculate the ratio of each
phase: rp = ep/e, where p={a, b, c}.
8. Calculate the sum of three phase level 9 details:
D9 = D9a + D9b + D9c, let
md9= max(D9(i )).
9. If
ra − rb− rb − rc  \t2, ra −rc −rb −rc \ t2
and md9\t3, it’s an A – G fault.
10. If
rb −rc− ra − rc \t2, rb −ra −ra −rc \t2

Fig. 9. Procedure of fault detection and classification scheme.

and md9\ t3, it’s a B–G fault.

11. If

rc − ra− rb − ra\ t2, rc − rb− rb − ra\ t2

and md9\ t3, it is a C–G fault.

12. If ra \ t4, rb \ t4 and md9B t3, it is an A–B fault.
13. If rc \ t4, rd \ t4 and md9Bt3, it is a C–B fault.
14. If ra \ t4, rc \ t4 and md9B t3, it is an A–C fault.
15. If md9B t3, it is a three-phase fault.
16. Let r= min (ra, rb, rc).
Fig. 8. Wavelet analysis of three phase currents for a three phase 17. If r=ra, it is a BC–G fault.
fault: D= 0.125 pu, R= 5 V. 18. If r= rb, it is an AC–G fault.
Table 1
Ratios in different types of fault

W. Zhao et al. / Electric Power Systems Research 53 (2000) 23–30

Fault conditions a=90° a =45° a =0°

Type D (pu) R (V) ra rb Rc md9 (A) ra Rb rc md9 (A) ra Rb rc md9 (A)

A–G 0.125 5 1.00 0.000 0.000 3.5×104 1.00 0.007 0.007 3.8×104 1.00 0.002 0.002 3.8×104
0.125 500 1.00 0.001 0.002 5.2×102 1.00 0.003 0.001 5.4×102 1.00 0.013 0.007 5.2×102
0.875 5 1.00 0.000 0.000 3.4×104 1.00 0.002 0.002 3.6×104 1.00 0.000 0.001 3.7×104
0.875 500 1.00 0.001 0.001 5.1×102 1.00 0.000 0.001 5.4×102 1.00 0.012 0.007 5.2×102

A–B 0.125 5 .998 1.00 0.002 2.235 .998 1.00 0.005 2.211 0.997 1.00 0.001 2.202
0.125 500 1.00 1.00 0.001 1.186 .999 1.00 0.001 1.189 1.00 0.997 0.001 1.184
0.875 5 1.00 0.999 0.000 2.232 1.00 0.997 0.001 2.333 0.998 1.00 0.001 2.397
0.875 500 1.00 0.999 0.000 1.173 1.00 0.999 0.000 1.172 1.00 0.998 0.001 1.172

A–B–G 0.125 5 1.00 0.602 0.001 2.7×104 .677 1.00 0.005 2.7×104 0.238 1.00 0.012 2.8×104
0.125 500 1.00 0.983 0.002 3.7×102 .973 1.00 0.006 3.7×102 1.00 0.949 0.002 3.8×102
0.875 5 1.00 0.600 0.000 2.6×104 .687 1.00 0.001 2.6×104 0.202 1.00 0.000 2.7×104
0.875 500 1.00 0.982 0.000 3.7×102 .980 1.00 0.001 2.6×104 1.00 0.941 0.001 3.8×102

Three-phase 0.125 5 1.00 0.528 0.470 20.65 .771 1.00 0.236 21.11 0.156 0.848 1.00 20.99
0.125 500 1.00 0.528 0.471 2.641 .768 1.00 0.234 2.647 0.156 0.848 1.00 2.592
0.875 5 1.00 0.530 0.469 16.27 .768 1.00 0.237 16.16 0.150 0.850 1.00 15.48
0.875 500 1.00 0.530 0.469 2.640 .766 1.00 0.235 1.705 0.151 0.851 1.00 2.737

29
30 W. Zhao et al. / Electric Power Systems Research 53 (2000) 23–30

19. If r = rc, it is an AB – G fault. md9 is less than t3, it is a three-phase fault, otherwise it
20. End. is a two-phase to ground fault. The healthy phase in
t1, t2 t3 and t4 are thresholds which are selected accord- two-phase to ground fault can be found easily because
ing to the detail values in normal and fault operations the ratio of it is the smallest one among the three.
in different cable systems. In this paper, t1 = 1, t2
= 0.9, t3 =100 and t4 =0.99.
When the proposed technique is used as part of a 5. Conclusions
protective relay, an appropriate data window length
must be adopted. Two important aspects must be con- This paper proposes a novel fault detection and
sidered: the window contains enough information to classification scheme based on discrete wavelet analysis
obtain the required wavelet levels and it is short enough for power cables. The property of multiresolution in
to produce the expected speed. Initial studies suggest time and frequency provided by wavelets allows an
that the proposed technique, based on the two wavelet accurate time location of fault transients while simulta-
levels, should be able to have acceptable performance in neously retaining information about the fundamental
terms of speed and accuracy. However, detailed studies frequency and its high-order harmonics, which is pow-
are required to fully assess the speed and security erful to extract characteristics of different types of fault
requirements of a protective relay. This is out of the in underground cable systems. Numerical test results of
scope of the paper. the scheme are very encouraging. It must be pointed
Since the fault resistance, fault location and fault out that it is necessary to fully evaluate the proposed
inception angle have important effects on analysis re- technique as part of protection relays or fault locators
sults, they have been considered in this scheme. Some under a wide range of system and fault conditions.
ratios associated with different fault conditions are These are currently under investigation.
presented in Table 1 which clearly show that this
scheme can be used to identify different types of fault
effectively in underground cable systems. The parame- References
ters ‘D’ and ‘R’ in the table stand for fault location and
resistance, respectively, ra, rb, rc and md9 have already [1] Z.Q. Bo, Q.Y. Xuan, R.K. Aggarwal, A.T. Johns, H.Y. Li, P.J.
been defined above in the rules. Moore, A neural network based protection and fault location
scheme for power distribution cables, APSOM’95, Hong Kong,
Suppose an A–B fault happened, which is in left 200 (1995).
m of section 11 (D =0.875pu) of the cable line and the [2] W.A. Wilkinson, M.D. Cox, Discrete wavelet analysis of power
fault resistance of it is 5 V, the parameters are obtained system transients, IEEE Trans. Power Syst. 11 (4) (1996) 2038–
from Table 1 as follows: ra =1.000, rb =0.999, rc = 2044.
0.000 and md9= 2.232. They are obviously not satisfac- [3] P. Pillay, A. Bhattacharjee, Application of wavelets to model
short-time power system disturbances, IEEE Trans. Power Syst.
tory to the rule (9), (10) and (11) which are used to 11 (4) (1996) 2031 – 2037.
identify single-phase fault in which two of the three [4] D.C. Robertson, O.I. Camps, Wavelets and electromagnetic power
ratios (ra, rb, rc) need to be very small and be almost system transients, IEEE Trans. Power Deliv. 11 (2) (1996) 1050–
equal while another ratio is always equal to one. The 1056.
md9 of single-phase fault is always very large compared [5] J. Liang, S. Elangovan, J.B. Devotta, A wavelet multiresolution
analysis approach to fault detection and classification in transmis-
with two-phase fault and three-phase fault. The sion lines, Elec. Power Energy Syst. 20 (5) (1998) 327 –332.
threshold t3(100) is set to define the ‘large’ and the [6] S. Santoso, E.J. Power, W.M. Grady, P. Hofmann, Power quality
‘small’ of md9. Since two ratios associated with two assessment via wavelet transform analysis, IEEE Trans. Power
faulty phases in two-phase fault are almost equal to Deliv. 11 (5) (1996) 924 – 930.
[7] I.K. Yu, Y.H. Song, Development of novel adaptive single-pole
1.000, seen from the example presented above (ra =
autoreclosure schemes for extra high voltage transmission system
1.000, rb =0.999), threshold t4(0.99) is set to distinguish using wavelet transform analysis, Elec. Power Syst. Res. 47 (1)
this type of fault from others to some degree. When (1998) 11 – 20.
two of the three ratios are both greater than t4 and the [8] W. Zhao, Y.H. Song, I.K. Yu, Wavelet analysis of fault generated
value of md9 (2.232 in the example above) is less than transients in power cables, Proceedings of UPEC, UK, 1998.
[9] A. Graps, An introduction to wavelets, IEEE Comp. Sci. Eng.,
t3, it is a two-phase fault. The ratios of two-phase to
2 (2) (1995) 50 – 61.
ground fault and three-phase fault change greatly under [10] A.W. Galli, G.T. Heydt, P.F. Ribeiro, Exploring the power of
different fault conditions which bring difficulties to wavelet analysis, IEEE Comp. Appl. Power (1996) October 37–41.
identify these two types of fault. Fortunately, they have [11] S.G. Mallat, A theory for multiresolution signal decomposition:
different characteristics in the sum of level 9 details, the wavelet representation, IEEE Trans. Pattern Anal. Machine
Intelligence 11 (7) (1989) 674 – 693.
md9 can be used to distinguish them from each other. If
[12] M. Vetterli, C. Herley, Wavelets and filter banks: theory and
three ratios of a fault are not satisfactory to rules design, IEEE Trans. Signal Proc. 40 (9) (1992) 2207 – 2232.
(9) – (14), that is to say, it is not a single-phase fault nor [13] G.W. Swift, The spectra of fault-induced transients, IEEE Trans
a two-phase fault, then check the value of md9. When PAS 98 (3) (1979) 940 – 947.