Professional Documents
Culture Documents
Estimation of Power System Harmonics Using Hybrid RLS-Adaline and KF-Adaline Algorithms
Estimation of Power System Harmonics Using Hybrid RLS-Adaline and KF-Adaline Algorithms
Abstract— This paper presents combined RLS-Adaline need for new algorithms that process the data, sample-by-
(Recursive Least Square and adaptive linear neural network) and sample, and not a window of data as in DFT and FFT, is of
KF-Adaline (Kalman Filter Adaline) approach for the estimation importance. Because these methods process data sample-by-
of harmonic components of a power system. The neural estimator sample, loosing one or more samples creates less leakage
is based on the use of an adaptive perceptron comprising a linear problems than DFT and FFT.
adaptive neuron called Adaline. Kalman Filter and Recursive
Least Square algorithms carry out the weight updating in
Kalman Filter [6, 10, 11, 14-15] is one of the robust
Adaline. The estimators’ track the signal corrupted with noise algorithms for estimating the magnitudes of sinusoids of
and decaying DC components very accurately. Adaptive tracking known frequencies embedded in an unknown measurement
of harmonic components of a power system can easily be done noise. But this algorithm cannot track abrupt or dynamic
using these algorithms. The proposed approaches are tested both changes of signal and its harmonics. Recently, many
for static and dynamic signal. Out of these two, the KF-Adaline researchers adopt Artificial Intelligence techniques [1] for
approach of tracking the fundamental and harmonic components harmonic estimation. Dash et al. [4] use an algorithm based on
is better. an adaptive neural network, which shows better tracking
Keywords-Harmonics Estimation; Adaptive Linear Neural capability of dynamic amplitude as to that of classic Kalman
Networks(Adaline); Discrete Fourier Transform(DFT); Fast
Fourier Transform(FFT);
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-
I. INTRODUCTION 13] use Genetic algorithm (GA), as a heuristic and stochastic
Due to the introduction of electronically controlled loads, the global searching algorithm for this purpose. This method
harmonic distortion in power system voltage and current provides excellent results but main disadvantage is that is
waveform is increased. Damped high-frequency transients are takes more time for convergence.
generated due to frequently switching on and off of power
semiconductors at points of voltage and current waveforms. II. PROPOSED ALGORITHM
So voltage and current waveforms of a distribution or a Let us assume the voltage or current waveforms of the known
transmission system are not pure sinusoids, but may consist of fundamental angular frequency Z as the sum of harmonics of
fundamental frequency, harmonics and high frequency unknown magnitudes and phases. The general form of the
transients. Also many of power system loads are dynamic in waveform is
nature which implies time varying amplitude of the current N
waveform
To provide the quality of the delivered power, it is
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
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)
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 )
(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 )
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
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
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
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 .
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 .
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
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 .
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
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.
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
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
Error (MSE) of signal using the two algorithms. From the 0.5
-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
{0.2 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
0 .5 3 4 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
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
TENCON 2009
6