Estimation of Power System Harmonics Using Hybrid

RLS-Adaline and KF-Adaline Algorithms

B.Subudhi and P.K.Ray
Department of Electrical Engineering
National Institute of Technology,
Rourkela, India
bibyadharnitrkl@gmail.com

Abstract— This paper presents combined RLS-Adaline
(Recursive Least Square and adaptive linear neural network) and
KF-Adaline (Kalman Filter Adaline) approach for the estimation
of harmonic components of a power system. The neural estimator
is based on the use of an adaptive perceptron comprising a linear
adaptive neuron called Adaline. Kalman Filter and Recursive
Least Square algorithms carry out the weight updating in
Adaline. The estimators’ track the signal corrupted with noise
and decaying DC components very accurately. Adaptive tracking
of harmonic components of a power system can easily be done
using these algorithms. The proposed approaches are tested both
for static and dynamic signal. Out of these two, the KF-Adaline
approach of tracking the fundamental and harmonic components
is better.
Keywords-Harmonics Estimation; Adaptive Linear Neural
Networks(Adaline); Discrete Fourier Transform(DFT); Fast
Fourier Transform(FFT);
I. INTRODUCTION
Due to the introduction of electronically controlled loads, the
harmonic distortion in power system voltage and current
waveform is increased. Damped high-frequency transients are
generated due to frequently switching on and off of power
semiconductors at points of voltage and current waveforms.
So voltage and current waveforms of a distribution or a
transmission system are not pure sinusoids, but may consist of
fundamental frequency, harmonics and high frequency
transients. Also many of power system loads are dynamic in
nature which implies time varying amplitude of the current
waveform
To provide the quality of the delivered power, it is
need for new algorithms that process the data, sample-by-
sample, and not a window of data as in DFT and FFT, is of
importance. Because these methods process data sample-by-
sample, loosing one or more samples creates less leakage
problems than DFT and FFT.
Kalman Filter [6, 10, 11, 14-15] is one of the robust
algorithms for estimating the magnitudes of sinusoids of
known frequencies embedded in an unknown measurement
noise. But this algorithm cannot track abrupt or dynamic
changes of signal and its harmonics. Recently, many
researchers adopt Artificial Intelligence techniques [1] for
harmonic estimation. Dash et al. [4] use an algorithm based on
an adaptive neural network, which shows better tracking
capability of dynamic amplitude as to that of classic Kalman
filtering approach. In that algorithm, the weights are updated
using Widrow-Hoff delta rule. As estimation of harmonic
parameters is a nonlinear problem, Qidwai and Betayeb [12-
13] use Genetic algorithm (GA), as a heuristic and stochastic
global searching algorithm for this purpose. This method
provides excellent results but main disadvantage is that is
takes more time for convergence.
II. PROPOSED ALGORITHM
Let us assume the voltage or current waveforms of the known
fundamental angular frequency Z as the sum of harmonics of
unknown magnitudes and phases. The general form of the
waveform is
N
y(t) ¦A sin(Z t I )  A
n1
n n n dc exp(Ddct)  P(t) (1)

important to know the harmonics parameters such as Where, N is the number of harmonics. Z n n2Sf 0 ; f 0 is the
amplitude and phase. This is essential for designing filter to
eliminate or reduce the effects of harmonics in a power system fundamental frequency; H (t ) is the additive noise;
[4]. So many algorithms have been proposed to evaluate the Adc exp(D dc t ) is the probable decaying term.
harmonics [1-2, 4-9, 11-15]. In order to get the voltage and
current frequency spectrum from discrete time samples, most The discrete time version of Eq. (1) can be represented as
N

¦A sin(Z kT I )  A
frequency domain harmonic analysis algorithms are based on
the discrete Fourier Transform (DFT) or on the fast Fourier y(k) n n s n dc exp(DdckTs )  P(k) (2)
n1
transform (FFT). However, these two methods suffer leakage
effect [3]. Although other methods, including the proposed Where, Ts is sampling period.
algorithms in this paper, suffer that problem, and this is due to Approximating decaying term using first two terms of Taylor
existing high-frequency components in the measured signal, series as
but truncation of the sequence of sampled data, when only a
fraction of the sequence of a cycle exists in the analyzed
y dc Adc  AdcD dc kTs (3)
waveform, boost leakage problem of DFT method. So the

978-1-4244-4547-9/09/$26.00 ©2009 IEEE TENCON 2009

1
 K(k 1) P(k)x(k 1)[1I(k 1)T P(k)x(k 1)]1 (10)
Now Eq. (2) becomes
N The updated covariance of parameter vector using matrix
y(k) ¦A sin(Z kT I )  A
n n s n dc  AdcDdckTs  P(k) (4) inversion lemma
n1 P(k  1) [ I  K (k  1) x(k  1)T ]P(k ) (11)
The nonlinearity arises in the model is due to phase of the Taking some initial values for the estimate at instant t
sinusoids. Bettayeb and Qidwai used GA for estimating phases initializes these equations. As the choice of initial covariance
of the harmonics. Although using GA for estimating phases matrix is large it is taken P=I where  is a large number and I
alleviated the mentioned problems but due to introduction of is a square identity matrix.
GA, the algorithm becomes slow and cannot be used as an The weight vector of the Adaline is updated using Kalman
online estimator. In this paper a fast neural network called Filter Algorithm as
Adaline is used. The weight of the adaline is updated using
Recursive Least Square and Kalman Filter algorithm.
G(k) P(k / k 1) x(k)T (x(k) P(k / k 1)x(k)T  Q)1 (12)
For estimation amplitudes and phases Eq.(4) can be rewritten G is the Kalman gain, x is the observation vector, P is the
as
covariance matrix, Q is the noise covariance of the signal.
N [An sin(ZnkTs ) cosIn  An cos(ZnkTs ) sinIn]
y(k) ¦ A
n1  AdcDdckTs  P(k)
(5) So the covariance matrix is related with Kalman gain with the
following equation.
dc

W1 P (k / k ) P (k / k  1)  G (k ) x(k ) P (k / k  1) (13)
x1 y(k) So the updated estimated state is related with previous state
with the following equation.
W2 y^(k)
š š š
x2 W(k / k) W(k / k 1) G(k)(y(k) x(k)W(k / k 1)) (14)
After the updating of weight vector using RLS or KF
W3 algorithm, amplitudes, phases of the fundamental and nth
x3 harmonic parameters and dc decaying parameters are derived
e as
e(k)
(W22n  W22n 1 )
.
An (15)
. Wn
.
§W ·
xn In tan 1 ¨¨ 2 n ¸¸ (16)
Weight © W2 n 1 ¹
updation Adc W2 n 1 (17)
§ W2 n  2 ·
D dc ¨¨ ¸¸ (18)
© W2 n 1 ¹
Fig.1 Block diagram of Adaline Because
W [ A1 cos(I1 ) A1 sin(I1 ) ... An cos(In ) An sin(In ) Adc AdcDdc ]T (19)
This equation gives an idea of using an adaptive linear
combiner comprising a neural network called ‘Adaline’ to III. SIMULATION RESULTS AND DISCUSSIONS
estimate the phases of the harmonics. A. Static signal corrupted with random noise and decaying DC
The block diagram of the Adaline is shown in fig.1 component.
The input to the Adaline is The power system signal used for the estimation,
x(t) [sin(Z1t) cos(Z1t) ... sin(ZNt) cos(ZNt) 1  t]T (6) besides the fundamental frequency, contains higher harmonics
The weight vector of the Adaline of the 3rd, 5th, 7th, 11th and a slowly decaying DC component
[2]. This kind of signal is typical in industrial load comprising
W(k) [W1(k) W2(k) ... W2N1(k) W2N(k) W2N1(k) W2N2(k)]T (7) power electronic converters and arc furnaces [2].
is updated using Recursive Least Square Algorithm as y (t ) 1 .5 sin( Z t  80 0 )  0 .5 sin( 3Z t  60 0 )
š š (20)
W (k  1) W (k )  K (k  1)H (k  1) (8)  0 .2 sin( 5Z t  45 0 )  0 .15 sin( 7Z t  36 0 )
Error in measurement is  0 .1 sin(11Z t  30 0 )  0 .5 exp(  5t )  P (t )
š The signal is corrupted by random noise
H (k  1) y(k  1)  x(k  1)T W(k) (9) P (t ) 0.05rand (t ) having normal distribution with zero
The gain K is related with covariance of parameter vector mean and unity variance. Fig. 2 and 3 show actual vs.

TENCON 2009
2
estimated signal using rls-adaline and kf-adaline method 1.52
respectively. It is seen from the fig. that actual and estimated rls -a d a lin e
k f-a d a lin e
signal almost matches with each other in both the cases. Fig. 4 1.51

shows the comparative estimation of fundamental amplitude

1.5
of signal using both rls-adaline and kf-adaline method. In case

Amplitude in p.u.
of rls-adaline, estimated amplitude oscillates between 1.47 p.u. 1.49
to 1.51p.u. but kf-adaline estimates fundamental amplitude as
1.5 p.u. Fig. 5 provides a comparative estimation of third 1.48

harmonic component of signal using the above two methods.

1.47
In the estimation using rls-adaline, during first few initial
times period estimation is 0.535 p.u. after that it settles around 1.46
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
0.515 p.u. KF-Adaline estimates third harmonics amplitude as T im e in s e c .

0.502 p.u. Fig.6 shows the estimated result of 5th harmonic Fig.4 Estimation of amplitude of fundamental component of signal
amplitude using both the algorithms. RLS-adaline estimates it 0.5 45
between 0.19 to 0.22 p.u. having more oscillations during 0 .54
r ls - a d a li n e
k f- a d a li n e
initial time period. But the estimated value of fifth harmonic 0.5 35
component using KF-Adaline is 0.202 p.u. Fig. 7 gives the
0 .53
estimated result of 7th harmonic component of signal using the

Amplitude in p.u.
0.5 25
two algorithms. RLS-Adaline estimates it around 0.155 p.u.
0 .52
but oscillation varies from 0.145 to 0.16 p.u. KF-Adaline
estimates 7th harmonic component as 0.149 p.u. which is more 0.5 15

accurate. Fig. 8 provides a comparative estimation of 11th 0 .51

harmonic component of signal using both the algorithms. 0.5 05

RLS-Adaline estimates it between 0.09 to 0.13 p.u. with it’s 0.5

0 0.01 0 .02 0.0 3 0.04 0.0 5 0.0 6 0.0 7 0 .08 0.0 9 0.1
Tim e in s e c .
oscillations around 0.11 p.u. but KF-Adaline estimates it as
Fig. 5 Estimation of amplitude of third harmonic component of signal
0.101p.u.
3
a c tu a l
0.225
2 .5 e s t im a t e d
rls -a d a lin e
2 k f-a d a lin e
0.22
1 .5
Amplitude in p.u.

1 0.215
Amplitude in p.u.

0 .5
0.21
0

-0 . 5 0.205

-1
0.2
-1 . 5

-2
0 10 20 30 40 50 60 70 80 90 100 0.195
S a m p le n o .

Fig.2 Actual and Estimated value of signal using rls-adaline 0.19

0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
T im e in s e c .
3

2.5
ac tual Fig. 6 Estimation of amplitude of fifth harmonic component of signal
e s t im a t e d

2 0.165
rls -a d a lin e
1.5 k f-a d a lin e
Amplitude in p.u.

1 0.16

0.5
Amplitude in p.u.

0 0.155

-0 . 5

-1
0.15
-1 . 5

-2
0 10 20 30 40 50 60 70 80 90 100 0.145
S a m p le n o .

Fig.3 Actual and Estimated value of signal using kf-adaline

0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
T im e in s e c .

Fig. 7 Estimation of amplitude of seventh harmonic component of signal

TENCON 2009
3
 61
0.135
rls -a d a lin e
0.13 k f-a d a lin e 60

0.125 59

0.12

Phase in deg.
58
Amplitude in p.u.

0.115
57
0.11
56
0.105

0.1 55
rls -a d a lin e
k f- a d a l i n e
0.095 54
0 10 20 30 40 50 60 70 80 90 100
S a m p le n o .
0.09

0.085
Fig.12 Estimation of phase of third harmonic component of signal
0 0.0 1 0.0 2 0.0 3 0.0 4 0.0 5 0.06 0.07 0.08 0.09 0.1
Tim e in s e c . 46

Fig. 8 Estimation of amplitude of eleventh harmonic component of signal 44

0.5 35 42

r ls - a d a li n e

Phase in deg.
0 .53 k f- a d a li n e 40

0.5 25 38

0 .52 36
Amplitude in p.u.

0.5 15 34
rls -a d a lin e
k f- a d a li n e
0 .51 32
0 10 20 30 40 50 60 70 80 90 100
S a m p le n o .
0.5 05
Fig.13 Estimation of phase of fifth harmonic component of signal
0.5
42
0.4 95 r ls - a d a li n e
k f- a d a li n e
0 .49 40
0 10 20 30 40 50 60 70 80 90 100
S a m p le n o .
38
Fig. 9 Estimation of amplitude of dc component of signal
Phase in deg.

36

5 .4 34
rls -a d a lin e
5 .2 k f- a d a l i n e
32

5
30
4 .8
Amplitude in p.u.

28
4 .6 0 10 20 30 40 50 60 70 80 90 100
S a m p le n o .
4 .4
Fig.14 Estimation of phase of seventh harmonic component of signal
4 .2
55
rls -a d a lin e
4
k f- a d a l i n e
50

3 .8
0 10 20 30 40 50 60 70 80 90 100
45
T im e in s e c .

Fig. 10 Estimation of amplitude of alphadc of signal

Phase in deg.

40

35
Fig. 9 shows the comparative estimation of dc component of
30
signal using both the algorithms. Estimated value of signal
using KF-Adaline is 0.496 p.u. but it’s value using RLS- 25

Adaline varies from 0.492 to 0.525 p.u. and it settles at 0.495 20

0 10 20 30 40 50 60 70 80 90 100
p.u. Fig.10 gives the estimated result of alphadc using both the S a m p le n o .

algorithms and it shows that estimation using KF-Adaline is Fig.15 Estimation of phase of eleventh harmonic component of signal
more accurate. -4
x 10
3 .5
8 0 .5 rls -a d a lin e
k f- a d a l i n e
3
80

2 .5
7 9 .5

2
Phase in deg.

MSE

79

1 .5
7 8 .5
1
78
0 .5

7 7 .5
rls -a d a lin e
0
k f- a d a l i n e 0 0 .0 1 0 .0 2 0 .0 3 0 .0 4 0 .0 5 0 .0 6 0 .0 7 0 .0 8 0 .0 9 0 .1
77 T im e in s e c .
0 10 20 30 40 50 60 70 80 90 100
S a m p le n o .
Fig.16 Estimation of MSE of static signal
Fig.11 Estimation of phase of fundamental component of signal

TENCON 2009
4
 0 .6 2 performance. Fig 23 and 24 show the tracking of fundamental
0 .6
and 3rd harmonic component of phases of a dynamic signal
0 .5 8
using both the methods. Both the methods estimates correctly
but KF-Adaline performance is better having negligible
0 .5 6
Amplitude in p.u.

oscillations. Fig.25 provides a comparative performance of

0 .5 4
Mean Square Error (MSE) of signal. KF-Adaline performance
0 .5 2
is slightly better in MSE analysis.
0 .5
2.5
ac tual
0 .4 8
rls -a d a lin e 2 e s t im a t e d
k f- a d a l i n e
0 .4 6 1.5
0 0 .0 1 0 .0 2 0 .0 3 0 .0 4 0 .0 5 0 .0 6 0 .0 7 0 .0 8 0 .0 9 0 .1
T im e in s e c .
1

Fig. 17 Estimation of 3rd harmonic component of signal during amplitude

Amplitude in p.u.
0.5
change 0
Fig. 11-15 show the tracking of the fundamental, 3rd, 5th, 7th -0 . 5
and 11th harmonic component of signal in presence of random -1

noise and decaying DC component using RLS-Adaline and -1 . 5

KF-Adaline methods. In the above estimation process, KF- -2

Adaline is tuned optimally by properly choosing the -2 . 5

0 10 20 30 40 50 60 70 80 90 100
covariance and noise covariance matrices. The time required S a m p le n o .

for trapping the fundamental and harmonic is approximately Fig.18 Actual and Estimated waveforms of dynamic signal using rls-adaline
0.02 sec. (20 samples) for RLS-Adaline method but KF- 2.5
ac tual
Adaline traps the fundamental and harmonic components 2 e s t im a t e d

initially with more correct estimation. 1.5

Fig.16 shows the comparative estimation of Mean Square 1

Amplitude in p.u.

Error (MSE) of signal using the two algorithms. From the 0.5

figure, it is found that, MSE performance in case of KF- 0

Adaline is comparatively better than RLS-Adaline method. -0 . 5

Fig.17 shows the tracking of 3rd harmonic component of -1

-1 . 5
signal, when it’s amplitude suddenly changes from 0.5 p.u. to
-2
0.6 p.u. at 0.05 sec. From the fig., it is seen that both the
methods track the 3rd harmonic component but tracking by
-2 . 5
0 10 20 30 40 50 60 70 80 90 100
S a m p le n o .
KF-Adaline is comparatively better. Fig.19 Actual and Estimated waveforms of dynamic signal using kf-adaline
B. Harmonic Estimation of a Dynamic Signal
1 .5 5 9 5
To examine the performance of RLS-Adaline rls -a d a lin e
1 .5 5 9 k f- a d a l i n e
algorithm in tracking harmonics and its robustness in rejecting
noise, a time-varying signal of the form 1 .5 5 8 5

y (t ) {1.5  a1 (t )} sin(Z 0 t  80 0 )
1 .5 5 8
Amplitude in p.u.

1 .5 5 7 5

 {0.5  a3 (t )}sin(3Z 0 t  60 0 ) (21) 1 .5 5 7

 {0.2  a5 (t )} sin(5Z 0 t  45 )  P (t )
1 .5 5 6 5
0
1 .5 5 6

is used where: 1 .5 5 5 5

a1 0.15 sin 2Sf 1t  0.05 sin 2Sf 5 t

0 0 .0 1 0 .0 2 0 .0 3 0 .0 4 0 .0 5 0 .0 6 0 .0 7 0 .0 8 0 .0 9 0 .1
(22) T im e in s e c .

Fig.20 Estimation of amplitude of fundamental component of dynamic signal

a3 0.05 sin 2Sf 3 t  0.02 sin 2Sf 5 t (23) 0 .5 3 5 5

a5 0.025 sin 2Sf 1t  0.005 sin 2Sf 5 t (24) 0 .5 3 5

f1 1.0 Hz f 3 3.0 Hz. f 5 6.0 Hz.

Amplitude in p.u.

0 .5 3 4 5

and random noise P (t ) is same as in the case of static rls -a d a lin e

k f- a d a l i n e
0 .5 3 4
signal.Fig.18 and 19 show the actual vs. estimated value of
signal using RLS-Adaline and KF-Adaline respectively. In 0 .5 3 3 5

both the cases actual and estimated value closely matches with
each other. Fig. 20-22 show the tracking of fundamental, 3rd 0 .5 3 3
0 0 .0 1 0 .0 2 0 .0 3 0 .0 4 0 .0 5 0 .0 6
T im e in s e c .
0 .0 7 0 .0 8 0 .0 9 0 .1

and 5th harmonic component of amplitude of a dynamic signal

using both RLS-Adaline and KF-Adaline methods. In all the Fig.21 Estimation of amplitude of third harmonic component of dynamic
three cases RLS-Adaline provides oscillatory estimation but signal
KF-Adaline provides more accurate and consistent

TENCON 2009
5
 0 .2 1 1 6 out. Both the algorithms track the fundamental and harmonic
0 .2 1 1 4 signals very well for both static and dynamic signal but the
0 .2 1 1 2 performance of tracking using KF-Adaline is better than RLS-
0 .2 1 1 Adaline. Both the static and dynamic signals are generated in
Amplitude in p.u.

0 .2 1 0 8
rls -a d a lin e
MATLAB platform. The used PC had a 1.46 GHz CPU and
0 .2 1 0 6
k f- a d a l i n e
1GB RAM. The same algorithms can be applied to other areas
0 .2 1 0 4 such as communication channels, telephones and other
0 .2 1 0 2
encrypted signals.
0 .2 1
IV. CONCLUSIONS
0 .2 0 9 8
0 0 .0 1 0 .0 2 0 .0 3 0 .0 4 0 .0 5 0 .0 6 0 .0 7 0 .0 8 0 .0 9 0 .1 This paper presents two new approaches for adaptive
T im e in s e c .
estimation of amplitudes and phase angles of harmonics in a
Fig.22 Estimation of amplitude of fifth harmonic component of dynamic
signal power system. The approaches are based on weight vector
8 0 .0 4
estimation of an Adaline using Recursive Least Square and
8 0 .0 3
rls -a d a lin e
k f- a d a l i n e
Kalman Filter algorithms. Comparing with RLS-Adaline,
simulation results reveal improvement in the performance of
8 0 .0 2
proposed KF-Adaline in tracking harmonic parameters even in
8 0 .0 1
presence of white noise and decaying dc components.
Phase in deg.

80

7 9 .9 9
REFERENCES
7 9 .9 8 [1] H. Mori, An artificial neural network based method for power
7 9 .9 7 system voltage harmonics, IEEE Trans. PD-7 (1) 402–409, (1992).
[2] E.A.Abu, Al-Feilat, I.El.Amin, and M.Bettayeb, “Power system
0 10 20 30 40 50
S a m p le n o .
60 70 80 90 100 harmonic estimation a comparative study,” Electric Power System
Research, vol-29, pp.91-97, 1994.
Fig.23 Estimation of phase of fundamental component of dynamic signal [3] Y.N.Chang, Y.C.Hsieh, C.S.Moo, “Truncation effects of FFT on
6 0 .1 estimation of dynamic harmonics in a power system” Power
rls -a d a lin e
6 0 .0 9 k f- a d a l i n e System Technology, 2000, Proceedings, International Power
6 0 .0 8 Conference 2000. Vol.3, pp. 1155-1160, 4-7 Dec., 2000.
6 0 .0 7 [4] P.K.Dash, D.P.Swain, A.Routray and A.C.Liew, “Harmonic
6 0 .0 6 estimation in a power system using adaptive perceptrons,” IEE
Phase in deg.

6 0 .0 5
Proceedings Generations, Transmission and Distribution, vol.143,
6 0 .0 4
no.6, pp.565-574, Nov.1996.
6 0 .0 3
[5] T.Lobos,T.Kozina & H.J.Koglin, “Power system harmonic
6 0 .0 2
estimation using Linear Least Square method & SVD” IEE
6 0 .0 1
Proceedings Generations, Transmission and
Distribution,vol.148,no.6,pp.567-572, Nov.2001.
60
0 10 20 30 40 50
S a m p le n o .
60 70 80 90 100 [6] Steven Liu, “An adaptive Kalman filter for dynamic estimation of
harmonic signals” in: 8th International Conference On Harmonics
Fig.24 Estimation of phase of third harmonic component of dynamic signal and Quality of Power, ICHQP’98, Athens, Greece, October 14–16,
-4
1998.
x 10
1.5 [7] P.K.Dash and A.Sharaf, “A Kalman filtering approach for
rls -a d a lin e estimation of power system harmonics,” in Proc. 3rd Int. Conf.
k f-a d a lin e
Harmonics Power System, Nashville, TN, 1998.
[8] N.R.Watson, J.Arrillaga, “Review Harmonics in large systems”
Electric Power System Research, vol-66, pp.15-29, 2003.
1
[9] S.Mishra, “A Hybrid Least Square-Fuzzy Bacterial Foraging
Strategy for Harmonic Estimation” IEEE Trans. On Evolutionary
MSE

Computation, vol.9, no.1, February 2005.

[10] Karren Kennedy, Gordon Lightbody & Robert Yacamini, “Power
0.5 system harmonic analysis using the kalman filter” IEEE Power
Engineering Society General Meeting-2003.
[11] Ashwini Kumar, Biswarup Das, Jaydev Sharma, “Robust dynamic
state estimation of power system harmonics” Electrical Power and
0 Energy Systems, vol-28, pp.65-74, 2006.
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Tim e in s e c . [12] U. Qidwai, M. Bettayeb, “GA based nonlinear harmonic
estimation” IEEE Trans.Power Delivery, December 1998.
Fig.25 Estimation of MSE of dynamic signal
[13] M. Bettayeb,U.Qidwai, “A hybrid least squares-GA-based
The performances of the proposed hybrid algorithms algorithm for harmonic estimation”, IEEE Trans. Power Delivery,
are very dependent on the initial choice of weight vector W 18 , 2, April, 2003.
[14] Simon Haykin, Thomas Kailath, “Adaptive Filter Theory” 4th
and Covariance matrix P . By using an optimal choice of Edition, Pearson Education, 2007.
weight vector, faster convergence to the true value of signal [15] L. Jung, “System Identification-Theory for the User”, 2nd Edition,
parameter can be achieved. After the optimization of the PTR Prentice Hall, Upper Saddle River, NJ, 1999.
weight vector, online tracking of the changes in amplitudes
and phases of the fundamental and harmonic components in
presence of noise and decaying dc components can be carried

TENCON 2009
6