APPLICATION OF LINEAR PREDICTOR AND WAVELET TRANSFORM

IN PARTIAL DISCHARGE LOCATION IN POWER CABLES

Babak Badrzadeh and S.M.Shahrtash
Department of electrical engineering
Iran University of Science and Technology
Tehran Iran
IRAN

Abstract: Partial discharges are caused by various defects such as voids, shield protrusions, contaminants, advanced
stages of water tree, etc and known the main cause of breakdown in insulation materials. Partial discharge (PD) location
is a powerful and useful tool for the maintenance and operation of electrical power distribution cables. This paper
describes the application of two digital signal-processing techniques (Linear prediction and Discrete Wavelet Transform)
to partial discharge site location. These techniques have been used for both on-line and off-line PD location. Results
show the effectiveness of these algorithms and by applying them to the acquired signal one obtain almost the true shape
of PD signal under noisy conditions.

Key words:- Wavelet transform- linear prediction-detection and location-condition monitoring-partial discharge-cable

1. INTRODUCTION This, of course, means outage cost, a cost, which is

The distribution of power in industrial complexes often unnecessary if the results of health check, was fine.
relies in medium voltage shielded power cable systems. Electricity companies would like to be able to do the
Power outages due to failure of cables or their health check on-line, thus removing the outage cost.
accessories during operation could cause the forced While evaluating the location of PD, one encounters two
interruption of critical processes, but in the current major problems; first, the significant reduction of the of
climate of electricity de-regulation; it is not acceptable reflected pulse, as it travels for a long segment of the
anymore due to outage cost. It is therefore important to cable and broadens (As shown later, for PD location, one
know the state of health of the cable network (as one of should identify amplitude first and second reflection of
the asset) in an electricity supply system. PD pulses form far end of the cable)[5]. The reduction in
Owing to faulty manufacturing processes or, to aging the amplitude caused by attenuation often makes it very
after several years of service, shielded high voltage difficult to trace the reflected pulse. The distortion of the
power cables may develop various defects such as pulse worsens with the time, causing a shift of the pulse
electrical tree or water tree within their insulation layer maximum. All these lead to a decline in the location
[1]. If the electric stress applied is high enough, these accuracy [5]. The second problem is the external
defects may cause PD. Locating PD sites in underground disturbances such as radio and TV station broadcasting,
power cables is a powerful and useful tool for evaluating corona, PD produced by other high voltage equipment
their conditions [2-3]. The incipient defects, which can and… that occur during the on-site measurement [6].
lead to failure of the cable in service, are thus located. It The first problem is not very serious and can be solved by
is also possible to identify high-risk sections of circuits in appropriate selection of sensor’s bandwidth [5], but the
advance and to replace just the few defective meters [4]. second is very important and one must suppress the noise
Thus one can prioritize different problem feeders and for accurate PD Location. As yet, various methods of
replace or repair them in according to asset management noise reduction have been developed and proven
programs [4]. As a result, The reliability of a plant can be successful in on-line and off-line PD measurements such
substantially improved if potential cable failure sites as correlation[1], adaptive filtering [2-3], etc. However,
could be identified and corrected after their installation. in these techniques, it is necessary to measure the noise
On the other hand, if a cable were being monitored, an prior to a measurement (Closed-loop noise reduction)
outage had to plan because current technology requires thus computational time increasing dramatically [6].
the cable to be taken out of circuit (Off-line testing). For implementing on-line measurement it is necessary to
keep computational time as low as possible. For this based but this technique has a serious drawback. In
purpose, Open-loop noise reduction techniques such as transforming to the frequency domain, time information
wavelets has been proposed [6] .It is no need to is lost [7]. This drawback is very important for
determine the reference noise in theses techniques, thus nonstationary signals such as PD. In an effort to correct
reducing the computational time to lowest values. this deficiency Gabor approved the Short-Time Fourier
In this paper we focus on two digital filters: Linear Transform (STFT) analysis [7]. The STFT compromise
Prediction (LPC) and Discrete Wavelet Transform between time and frequency information, but its
(DWT). drawback is that once you choose a particular size for the
time window, the window is the same for all frequencies.
2 NOISE REDUCTION Many signals require variable window size to determine
more accurately either time or frequency.
2.1 LINEAR PREDICTION To overcome the limitation of Fourier methods, the
Linear prediction is a powerful technique for predicting a Wavelet analysis is a variable windowing technique and
signal, which is, in this study, corrupted by noise. This allows the use of long time intervals where we want more
algorithm predicts the following sample with a certain precise low-frequency information and shorter regions
probability by valuing the previous samples.Fig.1 shows where we want high frequency components. For digital
the schematic diagram of LPC for PD detection and/or implementation of wavelets we use from DWT that is
location. For a signal or a class of signal (s(n)) the given by [7]:
estimated value by linear predictor of order N is[2]:
DWT(m,n)=∑ (a0m)-0.5f(k)[ψ((n-ka0m)/(a0m))] (4)
sp(n)=∑Ni=αis(n-i) (1)
Where ψ(k) and f(k) are wavelet function(mother
The optimum predictor is the one that minimize the mean wavelet) and original signal, a0m and ka0m are scaling and
square prediction error [2]: translation constants, respectively and k,m are integer
variables.
E=∑e2(n)=? [(s(n)-? Ni=1αis(n-i))]2=min (2) Wavelets are formed using high- and low-pass FIR filters
(Fig.2)[7]. In this figure the arrows denote down-
The coefficients of the predictor αi are to be determined sampling by 2.
from degree N so that the expected value of the Outputs of these high- and low-pass-filters can
prediction fault signal E goes to zero [2]: mathematically be expressed by [7]:

∂E/∂αi=0 (3) yHP[k]=∑ u[m].g[2k-m]

(5)
This equation leads to calculation of auto-correlation yLP [k]=∑u[m].h [2k-m]
function of the signal, but description of calculation
procedure would get beyond the scope of this paper. Where u denotes original signal and g, h denote high- and
low-pass FIR filters, respectively.
DWT permits decomposition into the sum of
approximation (high-scale, low frequency components)
and details (low-scale, high frequency components)
elements. In other words, detail and approximation
elements relate to high- and low-pass filters, respectively.
Fig.1 schematic diagram of PD detection by LPC The Inverse DWT (IDWT) is then used to reconstruct the
signal to the original signal with no loss of information
(Fig.3)[7]. In this figure the arrows denote up-sampling
2.2 WAVELET TRANSFORM by 2.Outputs of high- and low-pass filters in
reconstruction stage can be expressed by [7]:
2.2.1 WAVELET THEORY
Fourier analysis is a mathematical tool for transforming u[m]=∑{yHP(k).g[2k-m]+yLP(k).h[2k-m] (6)
our view of the signal from time-based to frequency-
 If |x(t)|>λ; then δλH=x(t) ; else δλH=0 (7)

And the soft threshold signal is [8]:

If |x(t)|>λ then; δλS=SGN(x(t)(x(t)-λ));else δλS=0 (8)

Fig.2 Dyadic wavelet decomposition Where SGN denotes sign function. The hard procedure
creates discontinuities at x=± λ , while the soft procedure
dose not, as a result, it is advisable to use soft
thresholding for de-noising.
In this method the key is the step two: numerical choice
of the threshold λ. The choice is critical, if the threshold
is to small or to large. Either the result of de-noising is
dissatisfied or the large distortion can be caused. There
are four selection rules that are as follows [10]:
Fig.3 Dyadic wavelet reconstruction • rigsure : adaptive threshold selection using
principle of Stein’s Unbiased Risk Estimate (SURE)
The type of wavelet used is determined by the • heursure: heuristic variant of the first option
decomposition filters. Any type of mother wavelet can be
• sqtwolog: universal threshold
used in the wavelet analysis, but we have experimentally
• minimaxi: minimax threshold
chosen to use db9 in all the transformations.Fig.4 shows
but by using these rules,the noise is not completely
the db9 mother wavelets.
removed and for best performance, it would better to
select the thresholds manually.

3. PD LOCATION PROCEDURE
Energize the cable with an appropriate excitation voltage
and increase the voltage level until a PD is detected or
the maximum pre-established is reached. Reduce the
voltage to zero after a very short dwell time. reader is
referred to [11] for further information.
Fig.4 db9 mother wavelet
3.1 OFF-LINE PD LOCATION
2.2.2 WAVELET DE-NOISING This is explained by means of Fig.5. Assume that PD
The general de-noising procedure involves three steps. occurs at a point C along a cable AB. The PD pulse splits
The basic version of the procedure follows the steps into two identical pulses that travel in opposite directions.
described below: The pulse traveling left reaches the near end, A, and is
1.decompose recorded as the first pulse on the figure. The pulse
Choose a wavelet, choose a level N (2-10), and compute travelling in the opposite direction is entirely reflected at
the wavelet decomposition of the signal at level N. point B and travels back to point A where it is recorded
2.Threshold detail coefficients as the second pulse on the figure. The pulse that had
For each level from 1 to N, select a threshold and apply reached point A first is reflected at A, then at B and
soft thresholding to the detail coefficients. returns back to A after having traveled one round trip
3.Reconstruct along the cable. It is recorded on the figure as the third
Compute wavelet reconstruction using the original pulse. The unknown distance AC can be given by the
approximation coefficients of the level N and the following equation [11].
modified (after thresholding) detail coefficients of levels
from 1 to N.Two points must be addressed: how to AC=(t1/t2) AB (9)
choose threshold, and how to perform thresholding.Let λ
denote threshold. The hard threshold signal is [8]:
 are first decomposed into eight levels using db9 wavelet
(Figs 11,12). In these figures a8 ,d1-d8 denote
approximate and detail coefficients at different levels.
The first five scales are omitted as they contain mostly
noise, thus, wavelets coefficients are set to zero for these
scales. Soft thresholding are performed on scales 6,7 and
8 since the PD pulses are prominent in those scales (We
don’t use from a8 in threshoding).
After soft threshing process, the signals are reconstructed
using IDWT with the results as illustrated in Figs. 13,14.
Fig.5 Principle of PD location by reflectometry In these cases calculation show that the PDs are located
165.8 and 107.6 meters from near end of the cables.
3.2 ON-LINE PD LOCATION
The principle of Reflectometry can be readily applied if 4.2 ON-LINE PD LOCATION
both cable ends are disconnected from the system and if In this case, data are recorded from both ends of the cable
the cable section has no branching. The arrival time PD (say end A, B) during in-service PD measurements of a
location method illustrated in Fig.6 can be applied even 1.4-km, 11 kV cable [12]. The velocity of propagation
when the cable ends are not disconnected from the and sampling rate are 168 m/µsec and 50
system and whenever the cable is branched. sample/µsec.The real PD site is located 1110 meters from
Capacitate or inductive sensors, installed at both ends, near end of the cable. Fig. 15 shows the signal that
detect the arrival times, t1and t2, of the PD pulse. recorded from end A of cable.
Knowing the total cable length, L, and the propagation Wavelet de-noising procedure is similar to previous and
velocity, v, the location, x, of the PD site can be we don’t repeat it. Figs. 16,17 show the de-noised signals
determined as follows [11]: of end A and end B of the cable, respectively.
By applying (10) to these signals,PD location has been
x=1/2[v(t1-t2)+L] (10) determined 113.08 meters from near end of the cable.

Fig.6 Principle of PD location by arrival time method

4. SIMULATION RESULTS
Fig.7 PD signal from 10 kV cable Without LPC filtering
4.1. OFF-LINE PD LOCATION
In this case, PD signals are recorded from 6 and 10 kV
cables [1]. Figs.7, 8 show these signals respectively. Both
cables have the length of 600 meters. The real PDs are
located 160 and 110 meters from near end of the 10 and
6kV cables, respectively. We first apply LPC technique
to extract actual PD signals from these raw signals.
Figs.9, 10 show de-noised signals. For the best
performance of LPC, the prediction order (N) is set
experimentally to 34. According to Figs.9, 10 and by
using (9), PD locations can be determined 165.3 and
107.3 meters from near end of 10,6 kV cables,
respectively.In wavelet de-noising method, the signals Fig.8 PD signal from 6 kVcable Without LPC filtering
Fig.9 PD signal from 10 kV cable With LPC filtering Fig.13 PD signal from 10 kV cable
With wavelet de-noising

Fig.10 PD signal from 6 kVcable With LPC filtering

Fig.14 PD signal from 6 kV cable
With wavelet de-noising

Fig.11 Decomposition of PD signal of

Fig.15. PD signal of end A of a 11 kV cable
Fig.7 into Eight levels
Without wavelet de-noising

Fig.12 Decomposition of PD signal of

Fig.16. PD signal of end A of a 11 kV cable
Fig.8 into Eight levels
With wavelet de-noising
 dielectric power cable and implications thereof for
PD location ‘‘, IEEE Electrical Insulation
Magazine, Vol.12, No.1, pp.9-16, Jan/Feb.1996.

[6] I.Shim, J.J.Soraghan, W.H.Siew,’’ Digital signal

processing applied to the detection of partial
discharge: An Overview’’, IEEE Electrical
Insulation Magazine, Vol.16, No.3, pp.6-12,
May/June.2000.
Fig.17. PD signal of end B of a 11 kV cable
[7] ’’ Wavelet toolbox for using with MATLAB’’,
With wavelet de-noising
MATWORKS inc., 2000.

6. CONCLUSION [8] I.Shim, J.J.Soraghan, W.H.Siew,’’Detection of PD

utilizing digital signal processing methods, part3:
This paper has presented two methods for PD location in open loop noise reduction’’ IEEE Electrical
cable networks. In off-line PD location, PDs has located Insulation Magazine, Vol.17, No.1, pp6-13,
by TDR principle (One end measurement) and noise Jan/Feb.2001.
suppressed by LPC and DWT. Results show the
effectiveness of both algorithms, while in on-line PD [9] I.Shim, J.J.Soraghan, W.H.Siew, F.McPherson,
location, PDs located by arrival time (Two end K.Sludden,’’ Robust partial discharge measurement
measurement) method and results shoe the effectiveness in MV cable networks using discrete wavelet
of DWT noise suppression (LPC isn’t an effective transform’’ IEEE Power Engineering Society
technique in on-line PD location). Winter meeting, pp.718-723, 2000.

7. REFERENCES [10] W.Xiaorang, G.Zungjun, S.Yong, Y.Zhang,’’

Extraction of partial discharge pulse via wavelet
[1] R.Villefrance, ‘‘Mobile system for assessment of shrinkage’’, IEEE 6th International Conference on
AC high voltage cables ’’, Ph.D. thesis, Technical Properties and Applications of Dielectric Materials,
University of Denmark, Feb.2000. China, pp.685-688, June.2000.

[2] H.Borsi,’’ Digital location of partial discharges in [11] M.S.Mashikian,’’ Partial discharge site location,
HV cables’’, IEEE Trans. On Electrical Insulation, selected topics’’, ICC Meeting, Fall.2000.
Vol.27, No.1, pp.28-36, Feb.1992.
[12] I.Shim, J.J.Soraghan, W.H.Siew,’’Partial discharge
[3] H.N.Bidhendi, Q.Su,’’Partial discharge location on location on high voltage cables’’, IEE High Voltage
power cables using linear prediction’’ IEEE 5th Engineering Symposium, August.1999.
International Conference on Properties and
Applications of Dielectric Materials, Seoul, Korea,
May.1997.

[4] C.M.Walton,’’ Detecting and locating MV failure

before it occurs, Experience with live line partial
discharge detection on underground paper insulated
11kv cables in London’’, IEE CIRED.2001
Conference, June.2001.

[5] S.Boggs, A.Pathak, P.Walker, ’’ Partial discharge

XXII: High frequency attenuation in shielded solid