Sei sulla pagina 1di 5

Modal Estimates from Normal, Operation of Power Systems.

Gerard Ledwich arid Edward Palmer


Queensland University of Technology EESE GPO Box 2434 Brisbane 4001
Email p.ledwich@out.edu.au Ph 61-7-38642864 Fax 61-7-38641516

Abstract - Oscillation frequency and dattiping of noise. This is often referred to as a random walk or a
interconnected power systems have traditionally been Brownian noise process. From this hypothesis the
determined from line switcliing or generator tripping. In a composite load would be the integral of white noise.
cottipetitive environment scheduling generation and limiting To examine this hypothesis consider that angle
transfers to ensure the system remains stable bring cost impacts
for the test. Another factor infruencing the test is that the
measurements made over the Queensland system at a lOHz
resulting oscillation must be significantly above &lienoise level sample rate. The Queensland system in Fig I shows two
frotti customer load variations. Examination of normal main modes central against Northern and central against
operation angle measurements across 12OOktn in Qiieetisland the south. This system is not planned for connection with
and power ttieasurettients in Victoria show that frequencies and other networks until late 2000.
dattiping of tlie modes can be determined. This paper presents a
method for extraction of ittipuke response parattieters with
particular ettiphasis on separation of tlie rtiodes close in
frequency.
Keywords: power system oscillations, modal identification,
Prony analysis

1. INTRODUCTION

The confirmation of the small signal damping performance


for interarea oscillations traditionally involves line
switching tests or special signal injection. This invasive
procedure will impact on the safe operating range for the
lines and may require compensation to market participants.
Figure I Measurements on tlie Queensland system
This paper examines the use of background perturbations
The angle measurements across the two links in Figure 2
present in the system to help identify the modal
show some evidence of oscillations with the lower curve
frequencies and damping. For Prony analysis a single
being less damped.
signal or event must occur and the response stay above the
..7 I
“noise level” sufficiently long to identify the desired
model order. The higher the model order the longer the .4 I

signal must remain significantly larger than the “noise”. In ..I I

this paper the “noise” is used to identify the system


.am
response.
Both angle and power measurements from the Australian .4I 5

system are used to confirm that the extracted signals do .5 0

correspond to known modes, and confirm that models can


.IO I
be extracted.
I I I I I
10 ao a ao 4 a0 a 10 m 11
on. 014.1.
2. IDENTIFICATION FROM NOISE nII””I.

Every change of customer power can act as a disturbance Figure 2 Time measurements of angles
on the inter area power transfer. The daily load curve is a The first stage is to observe that the spectral characteristics
well-known feature of the repeatability of customer of the angle on the interconnectors behave as a random
demand. Other loads such as draglines have a repetition walk process exciting the electromechanical modes of the
period of 1 to 2 minutes. There are other loads with short- power system. Figure 3 shows the presence of resonant
term repetition but these are often very small compared bumps in the average spectra of the angle signals. Note
with the total load. In periods up to 10 seconds there is that the low frequency energy is rising sharply.
evidence that the changes in composite customer load are
unpredictable. The hypothesis is that in thc frequency
range of electromechanical oscillations the composite
customer load changes are uncorrelated or pure white

0-7803-5935-6/00/$10.00 (c) 2000 IEEE 1527


0 Q5 1 1.5 2 25 3
I+
Figure 3 Averaged spectra of angles between Queensland nodes
0 1 2 3 4
Figure 5 Autocorrehtion offiltered Nth Qrcl data
To confirm the hypothesis of the angle being the
oscillation mode driven by a random walk signal examine The power data was sampled at 50Hz in Victoria at a
the Log-log plot of figure 4. The trend is linear apart from substation on the link North to Snowy and NSW. The
the resonant bumps which helps confirm the expected l/f power data was not quantised for transmission like the
spectral characteristic. angle data and could be whitened by one step differencing.
[n Figure 6 we see the autocorrelation from one line at the
Dederang substation showing some high and some low
frequency terms. Figure 7 shows the analysis of the
summation of power on all lines at Dederang. Note that the
'2Hzlocal machine oscillation has been almost completely
Ieliminated.
AubmlIelahOndedersnp
I

10' .....................
. .
.........
.....................
I I * I , . ,
0.8

0.6

0.4

02

1O0
-1o'= 10" 1oa 0
Frequency l-tz

d20 05 1 15 2 2.5 3 3.5 4 4.5 5


Figure 4 Log Log plot of angle spectra SEC
Application of simple filtering tools reduces the problem Fig 6 Autocorrehtwn of Power from Dederang- Wadanga line
to that of a system excited by white noise. For the angle
data the filter consisted of a high pass filter whose gain
rose linearly with frequency over the band of interest. The l :
.. .. ... ... ... ... ... ... ...
: : : ! ! : : :
.. .. .. .. ..
0.8 _ ........i .................. .................... ......... .........:..........:.......... ........
autocorrelation of the resulting signal is shown in Figure 5 . i
. .;
. ; ; 4 i
.. . . . . . . . .
I

showing the expected frequency and damping .... .... .... .... .... .... .... ....
characteristic. Note that autocorrelation of a system 06 - .. .. . .. .. : : : : -
.............................................................................................
. . . . .
... ...
.
...
.
...
.
...
.
...
.
... .
. ... ...
.
excited by white noise shows the impulse response of that ........f... ....+. ........:............:...........;......... ..L. .........:...........:..........:
..........
.
... ... ...
... ... ... ...... ...... ...... ...... ...... ......
0.4
system. ([3] pp65). The lag zero value of the ..
autocorrelation is always 1. .. ..j .. i.. Aj.......... ;i........... ;j.......... 1i..........
.................................................... .
3. ANGLE VS POWER !
,,./:
. ;. .......i...@.;. ., i
.
.......;
>>FA.A+i*\4 .
.....+*F...i .........i.. .....A# 4.b
' ' '

The power transferred across a link also contains


.. ..
;&.? ..
:
..
;... ...:. :... .:.. :...
, ,-yr-L*,*xA:

information about the interarea modes. One must be . 0 2 " * ' ' ' ' ~ ~

careful that the measurement of power contains data from


all parallel paths so the focus is on the inter area terms
only.
'The difference between power and angle signals as a tool
for extraction of modal parameters should be noted. The

0-7803-5935-6/00/$10.00 ( c ) 2000 IEEE 1528


power signal has a significant component of signal energy at 2Hz corresponding to a local mode. The basic Prony
that is not associated with electromechanical oscillations, analysis responded strongly to the 2Hz mode and did not
this power variation which is not a result of angle changes identify the O.SHz mode. Once again it is seen that the
in the system acts as noise in the measurement. The Prony approach works well only when the one-step ahead
sutocorrelation of any sequence of lag zero is always 1 as prediction is sufficiently sensitive to any error in the
seen in Fig 5,6,7. For the power signal in Fig 6,7 the particular mode.
oscillatory response has magnitude 0.2. In other words
80% of the signal at any sample is not correlated with
.
adjacent samples. In contrast the angle signal is a direct
1.6
measurement of the modes and is not subject to the same
level of noise; the mode magnitude is of size 1. 1.4

1.2

1
4. EXTRACTION OF MODAL PARAMETERS
In recent times the extraction of modal parameters has 0.8
been done using Prony’s Method, [l]. This method 0.6
attempts to fit a one step ahead predictor model to the
0.4
signal x[n] as described by (1)
0.2
P 0
s [ n ]= - C a [ k ] x [ n - k] -0.2
i=l
-0.4
0 1 2 3 4 5
Details of this technique may be found in [I], and [2]. pp.
224-225 where the sensitivity of the technique to
measurement noise is described.
Figure 9 Idenfijication at Dederang (Wadonga)doniinafedby
5. LIMITATIONS OF DIRECT APPLICATION OF local mode
PRONY 6. INTERLEAVED PRONY ANALYSIS
There are two frequencies present in the Heywood data but Investigations by the authors have found that even a
they are nominally at 0.48 and 0.52Hz.The beat frequency minute amount of noise is sufficient to give widely
is thus at 0.04Hzwhich then gives a period of 25sec. Thus incorrect results. This is particularly true when the signal
predictions would have to be over a quarter cycle of this is sampled at a high rate.
beat frequency (G sec) for the prediction to be strongly Under such conditions where there is very little difference
influenced by the not separating the modes. This is between one sample and the next a small amount of noise
illustrated by the simple model fitted to Figure 8 where the is sufficient to cause gross inaccuracy in the model
model did not separate the modes and the error is not estimate.
strong for several seconds. In order to overcome this deficiency it is proposed that
instead of using a one step ahead predictor model a k-step
ahead predictor be used. This will ensure sufficient
difference between each successive pair of samples used in
the algorithm thus enhancing the robustness of the
algorithm to noise. At the same time, there is no loss in
data as every sample may still be used. For example in a
4* order model 3 step ahead predictor algorithm, the O* ,
3d, 6*, and 9* samples would be used to predict the 12’
sample, while the I”, 4*, 7*, and 10* samples would be
used to predict the 13* sample etc. We will refer to this as
3 step Interleaved Prony analysis.
Overall the system parameters are able to be identified
from single and multi-site measurements and the noise
sensitivity reduced by adjusting the Prony algorithm of the
estimation process.
The advantages of the interleaved Prony technique may be
Figure 8 Fitted model compared with autocorrelatwn for illustrate by the following examples. Firstly consider a
Heywood.
signal samples at 50Hz containing two modal components
Figure 9 shows the identification from one of the line at at 0.48Hz and 0.52 Hz. Damping of each mode is assumed
Dederang. Here the power signal had a strong component to be negligible which would be expected to enhance
identification.

0-7803-5935-6/00/$10.00 ( c ) 2000 IEEE 1529


Table 1 below gives estimated frequencies for various
Signal-to-Noise ratios (SNR) using standard Prony
analysis.
-
Tabie 1 Standard Prony Analysis
IEstimated I SNR 1 SNR I SNR 1 SNR 1
I Freq -
1-00 1 =150dB I =130dB I =llOdB I
fl 0.4799 0.4802 0.4841 0.5020
f2 0.5201 0.5199 0.5146 0
As can be seen standard Prony analysis fails for a quite
high SNR of 1lOdB, being only able to identify one of the
two frequencies present. This would be further
exacerbated by and signal damping being present.

It is seen that at the sampling rate of 50Hz a minute


amount of noise present in the signal will cause the one
step ahead predictor algorithm to give widely erroneous
results. One way of overcoming this problem would be to 60 50 40 30 20 10 0
Signal-to-Noise Raio - dB
re-sample the signal at a lower rate, however this is seen as
discarding information which may be important as well as
requiring the design of digital filters to overcome potential Figure 10 Comparkon of Perforinance as function ojinterleave
step k
aliasing problems.
lit is seen that using a significant data point separation will
The alternative that is proposed is interleaved Prony vastly improve the robustness of Prony's method to noise
analysis whereby the one step ahead predictor implicit in down to quite a low SNR. This is seen as a major
standard Prony is replaced by a k-step ahead predictor. If
improvement on standard Prony's method which is highly
say, a two step ahead predictor is used in the above
susceptible to noise, particularly at high sampling rates.
example the performance is greatly improved over a wider
range of SNR values giving frequency estimates of
0.4848Hzand 0.5 162Hz at an S N R of 1IWB, and failing
at 90dB. Increasing the number of steps of the predictor
'1. MULTISITE IDENTIFICATION
allows the SNR to be lowered further. The upper limit of
With estimation of modes each of the sites could contain
the interleave step is based on aliasing. With 50Hz
some information about all of the modes. Some Prony
sampling to identify a 0.5Hz oscillation mode the
;inalysis formed the estimates of the major modes from all
corresponding Nyquist rate is 1Hz or 50 step interleaving. sites and averaged the inferred frequency and damping
For this example a 20 step interleave was the practical ratios. As seen in Figure 8 there is a strong correlation of
limit for improvement in estimation.
Ihe oscillation modes between sites.
The upper limit on the model order of the predictor is set CrosxocreUcn DederanpHeywood
by the length of the data set. It is not recommended that 0 1 . 5 .! .
! ! . ! . ! . ! ..! .
! ! .
this upper limited be used as there would be no ... .... .
...
.
...
.
...
.
... ...
.
...
.
...
redundancy in the data set and any noise in the
measurements would be reflected in the model.

Now it is useful to examine the performance of interleaved


Prony in identifying pole locations for varying signal to
noise ratios. In order to measure performance of the
algorithm, a distance between the estimated pole location
and the true pole location is introduced as . , ...
.. .. .. ... ...
-...................,......... ....................
I . .

-0.1
... ,.. ... ....
. .. ..............................
... .. .. .. .-
. . . .
e
I..
. ,
... .
...
n-..- c < i t t i ~ ~ i 1 ~
-10 -8 4 -4 .Z 0 2 4 6 8 10
sec
where p i refers to the true pole i" location and t i i s the
Fig 10 Cross Correlation of Power difference at separate sites.
corresponding estimate and M is the number of poles.
One of the aims of the analysis is whether the joint
Plots of this distance measure vs SNR for various predictor processing of power signals can help increase signal
orders or lags are shown bclow. confidence. In this case the dederang dam is cross
correlated with the Heywood power signal as shown if
Figure 10.. There is no evidence of the 2Ht signal which is

0-7803-5935-6/00/$10.00( c ) 2000 IEEE 1530


not visible at Heywood. The peak of the correlation 'way modal estimates may be optimised, taking note of the
indicates that largely when the heywood power is fact that some locations may give better estimates of some
increasing the dederang power is decreasing. Note that the modes than others.
noise term is highly correlated at a delay of zero indicating
the data is well time aligned. In addition the peak of the '9. DISCUSSION
negative response is delayed by about 0.3 sec The identification of system modes from background data
corresponding to a phase shift of about 54" between the :is based on two assumptions. We assume that there is
two points. This high degree of alignment points the way :sufficientvariation in consumer load to make a measurable
to not only extracting the modes but also to extract the change in the angles of generator groups. The second
mode shape from the multisite data. Thus the strenth of ;issumption is that there is little correlation to the load
involvement in the major modes for each of the generator changes in the frequency band of the electromechanical
groups can be extracted. oscillations. The evidence from angle and power
measurements clearly indicates the strength of the change.
8. ALGORITHM FOR MULTISITE ESTIMATION 'There can be no definitive proof of the lack of correlation
Consider two data sets obtained from different locations but in the cases studied the extracted modes do relate well
being processed to obtain the parameters of an AR 1.0 the expected frequencies.
(AutoRegressive) model for a signal from which the modal .4ngle signals appear to have greater potential for accurate
parameters of the system may be extracted. Let y , , y.,be extraction of the modal signal than power. The power
signal however does contain the modal data and can yield
the output vectors for the two locations, X , , X 2 be good models but at a higher noise level.
matrices of output values arranged such that the product of 'The extraction of the impulse response from the
its tith row with the AR parameter vector 6' gives the nth autocorrelation offers no significant difficulty. For the
numerical extraction of damping careful extraction of the
output value. Let 0,,8,be AR parameter vector
parameters at a time difference well suited to the expected
estimates obtained from data obtained at the two locations. modes is strongly indicated.
'There are distinct advantages for interleaved Prony
analysis but it still presents difficulties for extraction of
In the development below 8 denotes the parameter vector damping for closely spaced modes. For tight
obtained from combining the two data sets and will be 'LO.CONCLUSIONS
expressed in terms of the parameter vectors 8,,6,, 'The background power or angle signal variations are able
obtained by considering each data set on its own. I O provide estimates of the major power system modes.
Numerical procedures for accurate extraction of the
damping are recommended which focus on the degree of
= x,e y , = x,e
forward prediction for the estimator.
Define square error of each data set to be 'The case for improved estimates from time aligned
measurements at multiple sites appears quite compelling
XI4
E, = ( Y , - X , Q ) r ( Y , -
and a numerical procedure for the data combination is
E? = ( Y 2 - x,Q)'(Y* - x 2 4 presented as a logical extension to standard least squares
Total error estimation.
E = E , t E~ = ( y , - x,e)'(y, - x,e) :11.REFERENCES
111 J. F. Hauer, 'Application of Prony analysis to the
+(YZ - X2Q>7(Y2 - X28) determination of modal content and equivalent models
aE
-
ae = o = - 2 x : y l t 2x:x,e+ -2x:yz + 2 x : x 2 e
for measured power system response', IEEE
Transactions on Power Systems, Vol. 6, No. 3, August
:.e = ( x r x , t X : X ~ ) - I ( X +: ~x ;, y 2 ) 1991, pp. 1062-1068

121 S. M. Kay, 'Modern Spectral Estimation' , Prentice-


If estimates 8, & e2are computed separately then Hall Inc. 1988
(x:x1)% = X T Y I 131 A. Oppenheim and R. Schafer "Discrete-Time Signal
(xIx,)B, = x : y 2 Processing" Prentice Hall 1989
.*.(x:x,)B,+ ( x f x 2 ) 3 2 = x : y , + X,TYZ
= (x:x, t x:x,)B '12. ACKNOWLEDGEMENTS
The authors acknowledge the support provided by
:.e = ( x : x , + x : x , ) - ' [ ( x : x , b , + ( X ; X , ) ~ ~ ] I?owerlink in obtaining the angle measurements and from
Hence it is possible to combine the two data sets and so 'Vencorp in supplying the power data.
obtain the best estimate for the parameter vector. In this

0-7803-5935-6/00/$10.00( c ) 2000 IEEE 1531

Potrebbero piacerti anche