Tutorial Wavelet t uansfo Yms

Wavelet transforms in power systems

Part 2 Examples of application to actual power system transients
This is the second in a series of two arid illustrates some practical applications o j the wavelet transform to power syslenzs: proteclionljault delection, delection o j power quality disturbances and analysis of the partial discharge phenomenon in GI3 (gas-insulaled substations). Emphasis is placed on U number of practical issues.

by Chul Hwan Kim and Raj Aggawal

art 1 of this t~itorial' gave a general introduction to the wavelet transform (\W).The WT represents a powerful signal processing with a U ide variety of applications: acoustics, communications, transient analysis, medicine etc. The main reason for this growing activity is the ability of thc WT not only to decompose a signal into its lrequency components, but also, unlike the Fourier translorin, to provide a non-uiiihrin division or the lrequency domain, whereby it rocuses on short tirne intervals lor the highfrequency coiiipoiieiits aiid long iiitervals for low frequencies. This attribute to tailor thc frquc'ncy resolution can greatly facilitate sigiial analysis and the detection o f signal features, which can be very uselul in characterising the SOLIKE of the transients and/or the state of the post-disturbance system. The WT normally uses both the analysis and synthesis wavelet pair. Synthesis is used for waveform recons1ructioii. The origiaal signal is decoinposed into its constituenl wavelet subhands or levels. Each of these levels represents that part of the original signal occurring at that particular time and in that particular frecqwncy band. These individual frequency bands are logarithinically spaced rather than unikormly spaced as in Fourier analysis. The clecorriposed signals possess a powerful time-frequency localisation property which is one of the major
POWER ENGINEERING JOURNAL AUGUST 2001

benefits provided by the WT. The resulting decoinposed signals can then be analysed in both the time and freciueiicy domains. Multiresolution analvsis refers to the proceclures to obvain low-pass approximations and band-pass details from original signals. An approximation is a low-resolution representation of the original signal, while a

1 Multi-resolution and wavelet decomposition. L = low-pass filter; H = high-pass filter; (i?' . . , = down-samplina operation; A1, A2, A3 etc. = approximation of the original signal at levels 1 , 2 , 3 etc. D1, D2, D3 etc. = WT transform coefficients at levels 1,
~ ~~~~~ ~~~~~ ~~~

2,3 etc.

level 1

level 2

level 3

193

Tutorial: Wavelet transforms

levels or decomposition are taken as an example for illustration. The details and approximations ol the original signal 5 are obtained by passing i t through a filter bank, which consists of low- and high-pass lilters. A low-pass filter removes the high-frequency components, while the high-pass filter picks out the high-lrequency contents in the signal heing analysed. With reference to Fig. 1, the multiresolution decomposi cion procedures are defined as:

partial discharge (9%) transformer

power measurement

\ _I_L__

power qualitv (45%)

p : E E m /
(23%)

7
condition monitoring
(2%)

power system transients (11%)

2 Application areas of

the wavelet transform in power systems

detail is Lhe dillerence between two successive low-resolution representations of the original signal. An approximation contains Ihe general trend of the original signal while a detail embodies the high-frequency contents of the original signal. ApproximaLions and details are obtained through a succession of convoluiion processes. The original signal is divided into dillerent scales of resolution, rather than different frequencies, as is the case in Fourier analysis. The algorithm of multiresolution decomposition is illustrated in Fig. 1, where three
Table 1 Software tools for wavelet transform

where I and h are low-pass and high-pass filter vectors, respectively, Di and A, are the detail and approximation at resolution j , ,j = 1, 2, ..., J , respectively, A,., is the approximation of the [Link] immediately above level j , h = l , 2, .._, K where K is the length of the filter vector. To have a non-redundant representation and unique reconstruction ol the original signal, orthogonal filter hanks are required. WT and multiresolution decomposition are closely related. Also, as shown in Fig. 1, wavelet decomposition is accomplished by including down-sampling operations into the multiresolution analysis.

name

URL

MSLib
(Multiscale Lihrary)
Wavelet Explorer

hup://wu~[Link]/urbail/[Link] tml
[Link]

cost fi-eeware commercial

(h4athematica application package)

XWPL

[Link] freeware

(The X Wavelet Packet Laboratory)

WLT

[Link]

freeware

(The Iinager Wavelet Library) The Rice Wavelet

Tools W-Matrix

[Link] (Rice Wavelet toolbox for MATLAB) MATLAB linplementatioil of W-Matrix Multiresolutionanalysis

freeware

WavBox and ToolSmiths WaveLab MATLAB toolbox

LIFTPACK WAIL1

h ttp://www. [Link]
(MATLAB Wavelet toolbox)

freeware commercial freeware

commercral

MlU1 Wavelet Workbench WaveLib

Uvi-Wave

[Link] [Link] [Link] html [Link] [Link] [Link] html [Link]

commercial freeware commercial shareware

freeware

freeware freeware

(MATLAB Toolhox) Macwavelets

[Link]

194

POWER ENGINEERING JOURNAL AUGUST 2001

Tutorial: W a d et tvunsfo m s
The maximum number of wavelet decomposition levels for WT is determined by the length of the original signal, the particular wavelel being selected, and the level of detail required. The low-pass and the higli-pass filters are determined by the scaling function and the wavelet function, respectively. Sigiial processing uses exclusively orthogonal wa\7elets. The non-redundant representation and perlect reconstruction ol the original signal can only be realised through coiiipactly supported orthogonal wavelets. The ones that are frequently used for signal processing are Daubechies, Morlets, Coillets and Synilets wavelets. These wavelets exhibit different attributes a n d performance criteria when applied to specific applications, such as detection of transients, signal coni~~ress'ion and denoising. Although there are no delinite criteria for the selection of wavelets, the best choice is a wavelel, which most strikingly exhibits the phenomena to be stuclied. In the past decade, there has been an upsurge in the number ol publications on applications of WT in power systems. Typical applications and the results reported so far are promising. Pig. 2 gives a breakdown of the application areas of the WT in power systems. When applying the WT, one can use either use a hasic programming language or a purpose-designed software tool. The C , C++ or, more recently, Java languages, are chosen by most developers of WT analysis. A programming language offers powerful flexibility and capability for iinpleinenling any WT. However, it is no mean task to write one's own software for specific applications. A software tool provides a coniplete environment for aiialysiiig a signal through a WT. The researcher is only responsible lor inputting the relevant data and lor analysing the output result. There are a number or WT software packages available o n the market, sonie o l which are listed in Tahle 1. Although there are inany potential applications or WT in power systems, this tutorial deals with only a few illustrative problems namely, power system protectiodfault location, detection of power quality disturbances and aiialysis ok partial discharge phenomenon in CIS. Particular emphasis is placed on important practical issues relating to WT analysis, including: selection o l the appropriate mother wavelet, selection of the most significant features, multi-resolution analysis and, finally, performance evaluation. We present and analyse clava obtained from power systems o l which we have had experience.
Power system protectiodfault detection As nientioned before, Fourier transforms are not very good for analysing signal transients ol a non-periodic nature. The WT, on the other hand, is useful in analysing transient phenomena such as that associated with transmission line faults and/or switching operations. Unlike Fourier analysis, it provides time information and is u d u l in realising nonstationary sigiials comprising both low- and high-frequency componmts through the use of a variable window length of a signal.
Typical waveforms through wavelet trunsjorni vulisnf ion -16 aid the tlevelopinent of the fault detection, the discrete Wavelet transform (DWT) (see Reference I) has been employed. This determines a set of coefficients for different detail levels (dl, d2 etc.,) using different
3 The system model and the associated fault current waveform simulated: the 154kV Korean transmission system; [b, a typical

current waveform at relaying point P

240 MVA

-1
end P

26 km

1end Q

180 MVA

a
10
8 6 4

a
c

2
0

5
L

-2
-4

-6 -8
-1 0

70

75

80

85

90

95

100

time, ms
b

POWER ENGINEERING JOURNAL AUGUST 2001

195

rut orial: Wavelet transforms

010

0.08
0.06
.-

E
0

0.04 0.02
0

E
r

E -0.02
-0.04

U
r

3 -

-0.02
-0.04

-006
-0.08

-0-06
-0.08

-0.10 65 70 75 80 85 90 time, ms 95 100 105

-0.10
65 70 75 80 85
time, ms

90

95

100

105

0.06

F
t
I

0.06

-0 08
-0 10

4
I
100

I
65 70

I 75

I '

I1

1 ,
105

-006

I
65

I
70

1
75

I
95

I
100

I
105

80

85 90 time, ms
C

95

80

85 90 time, ms d

4 The coefficients of dl under %'-earth HIF

using DWT [corresponding to the current waveform shown in Fig. 36): TI= fault inception angle; [a) db4 mother wavelet; (6) sym5 mother wavelet; [c) bior3.1 mother wavelet; [@I coif4 mother wavelet

mother wavelets lor an actual current waveform. We siniulated a fault on a 1SSltV Korean transmission circuit with 26km line length and terminated in two sources of 240MVA and 180MVA at ends P and Q, respectively, as shown in Fig. 3u. The nominal power frequency is 60Hz. Fig. 3b depicts the 'a'-phase current waveforiii (simulated at end P but scaled down as if through a current transformer); it is for an 'a'-phase-earth highimpedance fault (HIF) at lOltm from line end P and near voltage minimum of the 'a' phase. Also considered in this fault study is a high impedance 3O0Q non-linear fault arc resist" Compared to the prefault steadystale signal, i t is apparent from the waveforni that there is quite a signilicant element of distortion (both at low and high frequencies) and this can be largely attributed to the

complex, iioii-linear behaviour of the HIP arc. The mother wavelets we used were Daubechies (db4), Biorthogonal (bior 3.1), Coiflets (coif4) and Symlels (syni5). It is important LO note here that associated with each type of wavelet is a family of short and long mother wavelets. For example, in the case of Daubechies wavelet, dbl signilies the shortest mother wavelet, i.e. the associated digital filters h a x only two coefficients, db2 is longer than dhl by a factor o l 2 and so 011. The same is true for other families of wavelets. The choice or the length of a mother wavelet depends on the application. Fig. 4 typifies the behaviour of the coefficient of detail 1 (dl) (which also signifies the highesi Frequency in the input signal) for the different mother wavelets wiLh DWT realisation of the current waveform shown in Fig. 3b. As expected, all the

196

POWER ENGINEERING JOURNAL AUGUST 2001

Tutoria 2: Wave1et t ransforms

0.08 r

C phase (0")

10

15

20

25

10

15

20

25

fault distance, km
a

fault distance, km
b

"6 0-010

phase ( 0 )
\ +

2
E
X

010

1 Aphase (0 )
6 phase (90")

--_
c Phase
(go")

3 -

0.008 0.006
0.004

---

5
5
m
C phase (0")
.-

\,-- -"r C phase (90 )

__
I

i .

._ -. .
c

006
0 04

E 0.002 0 I
lA

B phase (04

10

15

20

25

10
d

15

'
20
25

fault distance, km
C

fault distance km

coefficients of LII 'increase on fault inception and there are sinal1 discernible differences in the DWT outputs for the four different mother wavelets considered. For brevity, realised wawforms through the DWT for. only the faulted 'a' phase are. presented. The performance o f the DWT realisation was ev al ~ iI ed a under different fault types , fau I t incepiion angles and fault locations and some of the results have been presented in this tutorial. Seleclioii @ molher wavelet for H I F detection Selecting the most suitable mother wavelet enhances the ability of the HIF detection technique to extract the uselul iriformation rapidly. One can accurately identify the faulted phase(s) and so improve speed of HIF detection.

I n order to select the optimal mother wavelet, the absolute values of the dl coefficient are summaied over a 1 -cycle window for each rauli position, at lkin intervals Irom end l? Consider, for example, the waveforms shown in Fig. 5, which illuslrales the summated value of dl coelficiems of the three-phase current signals (as measured at relaying point) for an 'a'-earth HIE From Fig. Sa, based on the db4 mother wavelet, i t is evident that the summated values of dl are significantly larger for the faulted 'n'-pliase than for the two healthy phases .h' and 'c~;this is also true when employing sym5 and bior3.1 m o h e r wavelets, although the values are much smaller in the case of syni5. However, when employing the coil4 inother wavelet: (i) the magnitudes of the summated quantities are consistently smaller in the case o P coif4, ailcl

5 Variation of the summated values of the d l coefficient for threephase current signals as a function of distance: (a)db4 mother wavelet; [b) syni5 mother wavelet; (c) bior3.1 mother wavelet; coif4 mother wavelet

[a

POWER ENGINEERING JOURNAL AUGUST 2001

197

Tutorial: Wavelet transforms

(ii) the differences in summated values between the faulted dphase and the two healthy phases are significantly reduced in the case of the latter; this is true lor all lault positions. We have the selection of the inother wavelet as the magnitudes of the summated coefficients dl and the differences in magnitudes between the faulted and healthy phases. In this respect, after a seriesof studies einploying the foregoing dl cmefficient distribution approach for a variety of practically encountered different systein and fault conditions, both db4 and sytn5 mother wavelets seemed most effective to detect HIF on transmission lines. The db4 inother wavelet will be discussed further. useful information through DWT realisation under HIF exists. SUM-dl is the absolute sum value of the detailed output (dl coinpoiient) for one cycle period, %Cis a preset pick up level, which needs to be exceeded by the absolute value of dl before the counter FI can be incremented, and D is an integer which signifies a trip level, This decision logic has been specifically devised to discriminate accurately between HIF and non-fault transient events such as capacitor and line switching, arc furnace loads etc. As can be seen, when SUM-dl is greater than or equal to FC, the value ol FI is incrernented and as soon as it attains the preset level D, this indicates an internal fault and a trip signal is initiated. SUM-dl is based on summating the dl coefficieni over a one-cycle window and the sampling rate employed is 3840Hz, i.e. 64 samples/cycle at hOHz. The whole process is based on a moving window approach whereby

Fault detection algorithm Fig. 6 shows the fault detection procedure of the proposed technique, where FI is a counter that indicates the sample number for which

nth sample data

-..-

..

__
FI = FI + 1

- . ... --

.-

I
o

[ 6 Block diagram of the fault detection technique

A trip signal T

198

POWER ENGINEERING JOURNAL AUGUST 2001

Tutorial: Wavelet transforms

the one-cycle window is moved continuously by one sample. It is apparent from the foregoing decision logic that the criterion for the protection relay to initiate a frip signal is such that SUM-dl must stay above the threshold level FC continuously for D samples after fault inception. Extensive studies have revealed that, to maintain relay stability for external faults and also restrain under no-Iault conditions, the optimal settings lor FC and D are 0. I and 128, respectively The parameter 128 correspoiids LO a two-cycle 1x1-iodat power frequency The criterion Cor an internal fault is that the value of FC needs to he 2 0.1 for at least two cycles. Fig. 7 shows the behaviour of absolute sumniated values of dl for one particular fault position, lOkm from sending cnd P, as a function ot time; the graphs also depict the moving-window regime adopted. The graphs shown in Figs. 7a and b are for an .a-phase HIF (arcing), and are. for fault inception angles of V,,ur: and V,,,?, rcspectivcly These are the two extreme fault cases, the first giving rise to maximum high-frequency lraiisirnls in voltage signals ancl negligible DC ollsets in current signals, and the second case producing little voltage transients hut large DC offsets in currents. As expected, the magnitudes of the faulted dphase are much higher than the healthy phases and this is true for both fault inception angles considered. In both cases, the hulled a-phase stays above the set threshold level (FC=O.l) for approximately three cycles; this corresponds to a sample number of 192 and is well above the set level of D=l28; the protective relay will thus initiate a trip signal. Fur comparison purposes, the healthy phase currents b and c are also shown and, as expected, their magnitudes are well below the threshold level of FC=O.l. It should be mentioned that, for this condition, the i:aulL current is approximately five times the normal load current. It is apparent from this example that, apart from accurate fault detection, the DWT-based technique also has an inherent attrihute of being able to select correctly the faulted phase. This is of course an important feature in transmission systems employing single-pole auto-reclosure schemes.
Detection of power quality disturbances There are many forms of disturbance that can occur in power systems. In this next example,

0.1 4 0.12 0.10

9
E,

0.08

0.06

0.04

0.02

20

40

60

80

100

120

140

time, ms

::::
0.09

0.12

1
t

A phase

B phase

20

40

60

EO time, ms

100

120

140

we consider a power quality disturhaiice with deviations o l voltage and/or current from their ideal waverorms. Typical examples of power quality disturbances are: voltage sag (also known as voltage clip), voltage swell, outage (due, for example, to a fault), very short duration fault impulse, harmonic distortion, capacitor energisation transients and arc Iuriiace dislortion. To minimise their adverse impact on both the quality o f power supply and customer equipment. new techniques are needed. The WT provides a powerful tool in achieving this goal and, as iiidicated in Fig. 2, its usage as a research tool in striviiig 10 solve power quality problems is widespread. The current stale o l the art with respect t o

7 Variation of summated values of the coefficient dl for a-earth HIF as a function of time: (a) Fault near voltage maximum of %-phase; (b) Fault near voltage zero of a-phase

POWER ENGINEERING JOURNAL AUGUST 2001

199

Tu torial: Wav e k t t Tans forms

8 Voltage sag analysis using the wavelet transform: (a) Sample transmission-level voltage sag; (b] Scale 1 wavelet transform coefficients

400

200

d 9
9

o
-200

-400

I
0

I
0.1 time, s

I
0.1 5

I
0.20

1
0,25

0-05

6oooo

4 0 000 0

I
0.05

I
0.1

I
0.15

I
0,20

I
0.25

time, s

clelecting power quality disturbance is based on point-to-point comparison of adjacent cycles. The drawback of this approach is that it [ails to detect disturbances that appear periodically, such as phase controlled load waveshape disturbances. Another approach to detecting and identifying disturbaiices is based on neural networks. This approach can be elfective in detecting a i d identifying one particular type of disturbance, but in general it is not suited for detecting a variety of disturbances. We now illustrate the application o f the WT in detecting voltage sag.

ioltage sug disturbance detection Because o f . the increased use -of sensitive electronic equipment, vollage sag is having a more extensive impact on customers than before. As a consequence, monitoring and

assessing the system performance at both the transmission and distribution voltage levels is becoming increasingly important. Traditionally the sag RMS voltage and duration have characterised voltage sags. For some types ol equipment, though, more information is needed to assess the impact of sag. Parameters such as point-on-wave of fault inception and recovery, unbalance, distortion and phase-angle shirt, may be needed in addition to the stanclard depth and duration. Accurate time localisation or the sag disturbances is an iniportant first step in ascertaining some of these additional paramelers. Monitoring the transmission system is likely to become more important il transmission companies are contractually required to provide distribution companies, or other large customers, with a minimum level of power

200

POWER ENGINEERING JOURNAL AUGUST 2001

Tu toyi d : Wave k t transforms

-0-2

I
O

1
1.O

I
1 -5

I
2-0

I
2.5

0-5

time, us

level 1

level 2

05

i o

1.5

20

2.5

time, ~5
b

02

I
U

z g
U a ,
c

9 Realisation of a noise-ridden PD signal through the wavelet transform: (a) Noise-ridden PD signal in a 1 lOkV GIs;

c m

02
0

05

10

15

20

25

time, ys
C

(b) Four-level realisation of signal through a WT; (c) Reconstructed PD signal after de-noising

POWER ENGINEERING JOURNAL AUGUST 2001

20 1

Tutori d: Wavek t trunsfo rms

quality, quantified by some type of performance index. Disturbance characteristics such as fault start/stop times and sag magnitudes are likely to be included in the calculation of such indices. As mentioned before, in the application of the DWT in waveform realisation, different WT coefficients are attained at dirlerent decomposition [Link]. Scale 1 (dl) contains the highest-lrequency information present in the input signal, Scale 2 (d2) the next highest, and so on. Fig. 8a typifies a voltage sag disturbaiice simulated in a high-voltage transniission system and Fig, 8b shows its corresponding dl wavelet decomposition. It is apparent from the latter that the DWT is insensitive to the steadystate voltage but that it picks up the higher frequency components presenl at the sag start and end. Since the locations of the peaks in Fig. 8 h correspond to the locations o l the sag start/stop, the latter times can be found by devising a relatively simple search technique for automatically searching lor peaks in the DWT data. It should be noted that, in the foregoing example, the choice of the appropriate mother wavelet is again crucial. In this respect, an extensive series ol studies relating to the voltage sag problem have revealed that Daubechies db6 (which is classed as a medium wavelet) is ideally suited for solving the foregoing volvage sag problem.
Analysis of partial discharge phenomenon in GIS Gas insulated substations (GIS) are used worldwide, and much attention has to be paid to keeping them reliable and efficient. On-line condition moniloring of such equipment through, for example, acoustic detection and ultra high frequency (UHF) detection ol electromagnetic transient signal bursts, associated with PDs is an important advance in this area; the technique described here is based on the latter. The signal from PL) activity in a GIs is intermittent and weak and can be easily lost amongst many dilkerent types of extraneous noise such as corona discharge, thermal agitation of conductors/cables, etc., found in the substation environment. The key is to be able to distinguish and extract useful information from the iiieasured signals, i.e. denoise the noise-ridden PD signal. .Ihe efkctivenrss of a technique to do this depends

on the method of measuring the PU signals in the first place. Amongst several PD measurement systems available, the one based on ultra high frequency (UHI;) detection has a number of advantages. This technique employs a coupling sensor connected on the outside ol a GIS to detect and measure the UHF signals that are excited within (he GIS by short bursts of PD current emanating from a defect. Typically, the bandwidth of a signal generated ranges from about 25 to 400 MHz. A GIS contains many insulating spacers through which a UHF signal associated with the PD phenomenon propagates to the outside world. It is thus common practice to place the sensor close to a spacer in order to maximise the magnitude of the PD signal captured. Fig. 9a typifies a LJHF signal detected by a sensor in a IlOkV (;IS containing both useful PD information and extraneous noise. Fig. 9b shows its WT realisation over four levels of detail by the Db4 inother wavelet. Through applying a series of certain thresholds to [Link] the corresponding maxiina o l tlw WT, the effect of extraneous noise can be eliminated, and the genuine PD signal, as shown in Fig, 9c can he recoiwrructed. This example demonstrates the ekctiveness of the WT in extracling the noise elements that cannot be otherwise easily recognised by employing conventional techniques. In practice, a series of patterns can be built up by recording a number of tests over a length of time; these can then be used to ascertain, through artificial intelligence based pattern recognition techniques, whether there is a potential prohlem within a GIS.
Further reading
1 KIM, C . H., md AGGARWAL, R.: Wavelet translorm in power systems Part l, Power higinerring Journal, April 2000, 14, (21, pp. 81-87 2 STRANG, G . , a n d NCXJYEN, T.: Wavelets a n d filter hanks. (Welleslcy-Cambridge Press, 1996, ISBN 0 96140 887 1) 3 BOLLEN, M. H. J.: Understancling power quality problems (LEEt lress, 1 W Y ) 4 IEEE Guide for gas-insulated suhslations (IFFE Press, 1994)

0 IEE: 2001
Chul Hwan Kiln is a Professor in the School of Electrical and Coinputer Engineering, Suiigkyiinkwan University, South Korea. Raj Aggarwal is Professor in the Power ancl Enel-gy Systems Group, Department or Pleccronic C;r Electrical Engineering, University o l Bath, Bath BA2 7AY, UK. Prof. Kim is a meinber of Korean IEE and Prof. Aggarwal is an IEE Fellow

202

POWER ENGINEERING JOURNAL AUGUST 2001