80 IEEE Transactions on Power Systems,Vol. 5, No.

1, February 1990

INITIAL RESULTS IN P R O W ANALYSIS OF POWER SYSTEM RESPONSE SIGNALS

J.F. Hauer C.J. Demeure L.L. Scharf

Senior Member Member Fellow

BONNEVILLE POWER ADMINISTRATION TEE UNIVERSITY OF COLORADO

PORTLAND, OREGON BOULDER, COLORADO

Abstract BPA has since made substantial changes to

DTRANSCIENT. The objectives are to evaluate the
Prony analysis is an emerging methodology that method, to revise the code for utility applications,
extends Fourier analysis by directly estimating the and to fortify both for use with large models.
frequency, damping, strength, and relative phase of Polynomial rooting, a critical and numerically
modal components present in a given signal. The demanding task, is now accomplished by a routine that
ability to extract such information from transient was extracted from the NASA program SAMSAN 1121 and
stability program simulations and from large-scale converted to quadruple precision. The revised
system tests or disturbances would be quite valuable DTRANSCIENT has been combined with comprehensive
to power system engineers. This paper reports early driving logic to form SIGPAKZ. This program.close1y
results in the application of this method to parallels the BPA Fourier analysis programs SIGPAK
stability program output. It also includes [131 and SIGPAKR [141. All three readily accept
benchmarks against known models and a brief signal data from the BPA TSP, BPA's Power System
mathematical sumary. Disturbance Monitor amd variety of special sources.

Keywords: Prony, signal analysis, modal analysis, This paper reports early results in the application
dynamics, stability, eigenvalue, singular of SIGPAKZ to power system problems. It also
value includes benchmarks against known models and a brief
mathematical description of Prony analysis.

I. INTRODUCTION
11. A BRIEF DESCRIPTION OF PRONY'S METHOD
Prony analysis 11-11] is an emerging methodology that
extends Fourier analysis by directly estimating the Like Fourier analysis, what we shall call Prony
frequency, damping, strength, and relative phase of analysis originated in an earlier century [l]. Its
modal components present in a given signal. The practical use has awaited the digital computer, and
ability to extract such information from transient means for dealing with some inherently ill-conditioned
stability program (TSP) simulations and from mathematics.
large-scale system tests or disturbances would be
very useful to power system engineers. Suppose that a linear, time-invariant dynamic system
is brought to an "initial" state x(to)=x, at time
Such a tool would be particularly valuable for TSP to. by means of some test input or disturbance.
output analysis, where it could provide Then, if the input is removed and there are no
subsequent inputs or disturbances to the system, it
o parametric surrmaries for damping studies (data will "ring down" according to a differential equation
compression) of form
o quantified information for adjusting remedial = A x (1)
controls (sensitivity analysis and performance
evaluation) where x is the state of the system and n is the
number of components in x (i.e., the order of the
o insight into modal interaction mechanisms (modal system). Let Xi,pi,q$. be respectively the
analysis ) eigenvalues, right elgenvectors, and left
eigenvectors of (n x n) matrix A. (See 1151 or [161
o reduced simulation times for damping evaluation for details). Then the solution to (1) can be
(prediction) expressed as

These considerations lead BPA to contract with

Dr. Louis Scharf for the writing of a FORTRAN program
based upon his research into Prony methods. The
effort was lead by Dr. Cedric Demeure, who produced
the interactive FORTRAN program DTRANSCIENT.

where Ri=pi$ is an (n x n) residue matrix.

89 SM 702-2 PWRS A paper ret-nded and approved Note that qlx, in (2A) is scalar (that is, a
by the IEEE Power System Engineering Caomittee of the simple constant). This implies that, though ~0
IEEE Power Engineering Society for presentation at the (with qT) determines the stimulus to the mode
IBEEIPES 1989 Summer Meeting, Long Beach, Califoda, associated with eigenvalue Xi, the distribution
-
July 9 14, 1989. Hanuscript submitted of modal response among the components of x is
January 26, 1989; made available for printing entirely determined by the corresponding right
May 23, 1989.
eigenvector pi. Consequently information about
pi can be extracted by an appropriate modal
decomposition of x(t).

@ 1990 IEEE
0885-8950/90/02~80/Wl.~
 81

For simplicity, suppose that there is just one output This can be expressed more compactly as
from the system and that it is of form
ZB = Y . (7B)
y(t) = c x(t). (3)
If the,zi are found then the eigenvalues Xi can
Prony methods and their recent extensions are be calculated from them via (6B). The zi are
designed to directly estimate the parameters for the necessarily the roots of some nth-order polynomial
exponential terms in (U)and/or (3), by fitting a with (unknown) coefficients ai, and thus satisfy
function
z" - (alzn-1 + a2zn-2 + ... + anzO = o (8)
Q
g(t) = 1 Ai exp(qt) cos(2nfjt + +i) (4) Next construct the (1 x N) matrix
i=l
'i = [-an -+-I ... -ai 1 o ... 01 = [-a 101 (9)
to an observed record for y(t). In doing this it may
also be necessary to model offsets, trends, noise, and apply it to (7): after some minor reordering,
and other extraneous effects in the signal.
XY = y(n) - [aly(n-1) + ... + a,g(0)1
Let the record for y(t) consist of N samples
y(tk)=y(k), k=O,l,...,N-l, that are evenly spaced = 'iZB
by an amount At. The strategy for obtaining a (10)
Prony solution (PRS) can be sumnarized as follows:

STEP 1: construct a discrete linear prediction

model (LPM) that fits the record. =o
STEP 2: find the roots of the characteristic where the last step follows from the fact that each
polynomial associated with the LPH of step zi satisfies (8). Since the initial time has been
1. selected arbitrarily, (10) can be applied repeatedly
to form
STEP 3: using the roots of step 2 as the complex
modal frequencies for the signal,
determine the amplitude and initial phase
for each mode.
These steps are performed in z-domain. For power
system applications the eigenvalues would usually be
translated to s-domain, consistent with equations
(1 1 4 5 ) .
Prony's main contribution is at step 1. Its
development here parallels that of [2], pp. 378 and
379. For initial purposes, assume N=2n and that the
signal record is noise-free. Solution of this equation provides the z-domain
polynomial in (a), which is then rooted for the zi
The notation is simplified if (4) is recast in the and, through (6B), the eigenvalues Xi. This
exponential form completes steps 1 and 2 of the basic Prony method.
Step 3 is the solution of (7) for the complex
P amplitudes Bi.
$(t) = 1 ~i exp(Xit) (5)
i=l Practical use of this approach requires some
extensions and refinements. A major problem is that
paralleling (-2). At the sample times tk this can the true system dimension, n, may be unknown or may
be contracted to be so large that any fitted model must be a
reduced-order approximation. Additional modes may be
P needed to fit signal offsets or noise. The array
$(k> = 1 Bi 21, (6A) dimensions in (11) will usually differ from those
i=l shown, and tend to "overfit" the signal by using a
generous number of samples and fitted modes.
zi = exp(XiAt). (6B) Singular-value analysis and other mechanisms are then
used to adjust model features at each solution step.
The inmediate objective is to find the Bi and zi
that produces y(k)=y(k) for all k. For insight as to The reconstructed signal y(t) will usually fit y(t)
how this can be done, apply (6A) at each tk and inexactly. An appropriate measure for the quality of
form the following equation: this fit, used here, is the signal-to-noise ratio
-I-

B1zY + . . . +Bnzon B1

Biz: + . . . +Bnzt 1 1
z2 -- 1
zn B2

.zN-1
- L BN-
82

SNR 20 log I ly(k)-3(k)l I/ lY(k) II (12) I I

where 11.11 denotes the usual root-mean-square norm

and the SNR is in decibels (dB)
0.67 HZ MOO€

111. STRATEGIC ISSUES

The immediate objective is to assess Prony analysis
as a tool for power system applications, and to s’ + (wl.4)’
conduct or promote further development as needed.
Ahere are many aspects to this. The methodology 0.71 HZ MODE
itself permits many variations. Then, for a given
record, there are questions as to PRSCKSO MOOEL . I
POLES* 0.0000 JrO.66666667 HZ
0.0000 + J-O.71928S71 HZ
o preprocessing, to mitigate the effects of noise
or hidden inputs Fig. 1. Model #l for prs check series #O
o how much of the record to use in the model fit
0.71 Hz mode has a damping of -0.005. The unit for
o the order of the fitted model damping here is inverse seconds, divided by 2n to
be dimensionally consistent with frequency in Hz.
o pruning and/or adjustment of the fitted model
Model #l was introduced in 1201, as a proposed
The power system models ordinarily used in stability benchmark for damping estimators. The modal
studies pose serious difficulties for Prony analysis frequencies are realistic for the western power
(and for other methods that rely upon system, and there is a risk that the modal damping
linearization). They are strongly nonlinear, and of may be so as well. Figure 2 illustrates that the
dimehsions well beyond those of models that any direct examination of such damping can require very
fitting procedure would reliably construct. Linear long simulations for correct results. A hoped-for
characterizations can be developed for power system benefit of Prony analysis is the minimization of
transient behavior, but, if they are formed with simulation length.
respect to the actual operating point, they should be
expected to vary with time. It appears that Prony 500
analysis performs such a characterization implicitly. I C
The evaluation of Prony analysis in this context
requires special attention to the factors below:
o The Prony model will normally be of far lower
order than the dynamic process underlying the
signal record.
o For large-signal cases, PRS results will change
with the magnitude of the system disturbance or
input.

o For large-signal cases, PRS results will change

with record length and position.

o For large-signal cases, there is probably a

maximum record length beyond which information
is lost in the PRS.
Fig. 2. Response of PRSCKSo model #l to brief
o For small-signal cases, substantial changes in pulse input
record length should not produce material
changes in PRS results. Prony analysis was performed on the Figure 2 signal
for successively longer values of the record length,
These are being examined through a series of test TBAR. In all cases the sample spacing was At4.10
cases, which range from very simple to rather sec. The solution logic, though permitted to develop
difficult. Salient results are presented in the a model of order 90, in each case recognized that
sections that follow. just two significant (complex) modes were present and
pruned the initial LPM accordingly. Model parameter
The methodology is also being applied to system estimates were accurate to 4 significant digits for a
operating records, in a parallel effort. In this 2 second record, and improved with record length.
environment the problems with noise and hidden inputs Like results were obtained for Model 2.
are much more severe, and alternative methods (such
as lattice filters [17-191) may be also be tested.

V. PRS CHECK SERIES 14

I V . PRS CHECK S E R I E S NO
The basic model for PRS check series #4 has the
This series of checks is based upon two elementary frequency response shown in Figure 3A. It represents
models. The first of these, shown in Figure 1, has a linear characterization of bus frequency response
two complex modes. Both have zero damping. Model #2 to small-signal control action during a major
has the same structure and parameters, except that disturbance (see (211, Section V). Having 14 complex
the 0 . 6 7 Hz mode has a damping of 0.02 and the modes, it is of order n=28. Some modes are closely
 83

indicates an excellent fit, as can also be seen from

the superposed curves. The corresponding PRS table
reproduces the parent model data to 6 or 7
-20 significant digits. The damping, for an eigenvalue
X=u+jw, is displayed as -X/(2n).

m
-30
.SlGYRL-TO-NOISE
-.. -
MODE ORMPING
RRTlO I 120.8700 FOR NMOOES
FRO lHZl REL W T
-
111
PHASE RHPLITUOE

-
5
2
a.
4 0 1
2
3
9
0.0239805
0.0318909
0.0596168
0.0508920
0.0367901
0.0912621
0.3095262
0.4608535
0.2690819
0 9999797
30.119507
0.18656311 116.559008
-9 998126
013286152 -119:239292
0.01911529
0.0100207
0.02111690
0.0176505
5 0.0938688 0.5357990 0.5398253 0.052731 0.0287265
-50 6 0.0227360 0.5975186 0.5201265 76.9631113 0.0279370
7
8
0.0857313
0.0257093
0.7275286
0.7926979
y!:$b: ; ~ ~ : ~ ~ ~ ~ < ~
0.0353275
0.0537119
9 0.0633062 1.0928168 0.1917231 -811.687450 0.0076122
-60
10 0.0560979 1.1506912 0.1709997 -21.813123 0.0091579
11 0.0176936 1.2396901 0.07857211 -86.217689 0.0092203
12 0.0265690 1.2919787 0 0757616 -28 926462 0.0090693
13 0.0384194 1.5015666 0: 1254782 167:056919 0.0067397
19 0.19969U1 1.6059325 0.9596698 -61.027965 0.0299209
-m

TABLE I. PRS table for PRSCKS4 model #I

TBAR = 20 seconds (200 samples)
Fig. 3A. Frequency response for PRSCKS4 model 8 1

VI. ANALYSIS OF KEMANO GENERATOR RESPONSE

The Kemano generator, in northwestern British

Columbia, attaches to the main power system through a
long radial connection (see Figure 5). Kemano has a
frequency-domain "signature" that is conspicuous
throughout the western power system.

I , ~ " ' ' ' I' ~

"""'"s~ ,, , -0.1
IO 40
TIME IN SECONDS

Fig. 3B. Impulse response of PRSCKS4 model bl

POUT 1 / 1 TSPRN=COOOO.0~00~~.~~
SNR = 120.87

-MEIISUIIEO S16*11L
-A--- PION1 M M E L

TlHE I N SECONDS

Fig. 4. Prony fit for PRSCKS4 model bl TBAR = Fig. 5. General structure of the Western U.S.
20 seconds (200 samples) power system

spaced, as may be deduced from the beating effects in

the impulse response of Figure 3B. Figure 6A shows the frequency response of Kemano
generator power to the modulation of reactive load at
Figure 4 and Table I show typical SIGPAKZ outputs. the Malin 500 kv bus. This data was produced by
The Figure 4 plot header indicates that the signal is Fourier analyzing the ringdown signal of Figure 6B,
named POUT 1/1, that it extends from 0.6 to 20.5 and dividing the results by the Fourier transfom of
Sec., and that the SNR for the PRS is 120.87 dB. This the applied pulse. The single dominant peak makes
84
G KEMANO 13.8 I COULEE 2 13.8 I 4 ELECTRICAL POLlER Kemano a good initial case for testing SIGPAKZ
R22/Rl1 : INPUT =MAL MVR IHlNOLlOl x I.0000
0 , 0 against transient stability output data.
4

Full PRS tables for TSP or field test signals tend to

be large, showing many modes that are weak or
accessory to the fitting process. Signal offsets
generally produce modes near 0 Hz or near the Nyquist
-20
frequency 1/(2At). Table I1 shows a sequence of
(D
PRS tables in which such modes have been supressed.
Line 7 of the header indicates that all modes with a
relative weight below (TRIMTAB E) 0.20 have been
trimmed from the display. The normalization is
-YO performed with respect to the mode within a user-
specified frequency range.
-50 Linearized characteristics of stability output
signals should be expected to vary with time (e.g.,
according to system stress). It is useful to examine
-60
such a signal by sliding a data window along it and
FREOUENCT I N MERTZ processing the window contents (a sub-record) at
preset positions. This provides an indication of
Fig. 6A. Frequency response of Kemano electrical dynamic variability, assists the identification of
power to shunt reactive control at Malin dominant modes, and checks the solution process for
consistency. Table 11 and Figure 7 show results for
a sliding-window analysis in.which a 15 second (225
sample) window was advanced 10 samples (0.667
seconds) between successive solutions. These
indicate that, consistent with the peak in Figure 6A,
Kemano is strongly involved in a mode that is
essentially ,fixed near 0.628 Hz, with a damping pear
0.0212-0.0215. Kemano also participates in a strong
mode near 0.77 Hz, which is either more variable or
less accurately estimated.

G KEMRNO 13.8 I COULEE 2 13.8 1 I) E L E C T R I C A L POHER

NO. s O L u r I a N s = 32 TRIMLOC = O . I O E * O O

2000
O o o TIME I N -CYCLES * o

Fig. 6B. Kemano ringdown for brief Myar load

pulse at Malin

C ..PRONY
t aS1GNRL-IO-NOISE R A T I O
C D I S P L A Y IRIM L E V E L -
-
S-TABLE FOR PRS NUMBER 1: TSPRN C 91.00.
69.9550 FOR NMOOES
O.lOOOOE*OO
56
- - 990.003

c
C MOOE DAMPING FRO IHZI REL HT PHASE RMPLITUOE
C 9 0.1505101 0.1082269 0.1106290 59.908092 3.2976931
C 5 0.02111010 0.6289195 1.0000000 -178.260696 23.09401171
C 6 0.10327711 0.6833275 0.2172331 -118.116087 5.0167906
E 7 0.0761593 0.7818811

-
C .*PRONY S - M O L E FOR PRS NUMBER 2: I S P A N
C .SIGNAL-IO-NOISE RATIO
E 89.00.
75.6260 FOR NMOOES
E O I S P L A Y T R I M L E V E L - 0.10000E*00
56
-
0.81871U2 7.611981

-
18.90711213
980.003

C MOOE ORMPING FRO IHZI REL HI PHASE AnPLiTunE REAL RXIS x 104
C 5 0.3891222 0.5557588 0.1753990 95.627236 3.6571974
C 6 0.0212116 0.6286833 1.0000000 -26.579108 20.8513308
Fig. 7. PRS locus for sliding-window fits to
-
C 7 0.0711585 0.78191197 0.5791028 -161.512752 12.0813190
C .*PRONT 5-TABLE FOR PRS NUMBER 3: TSPRN
C .SIGNRL-TO-NOISE
C
RATIO
D I S P L A Y T R I M LEVEL.
-
O.lOOOOE*OO
E 129.00. 1020.003
67.9670 FOR NMOOES 57 - Kemano ringdown

C HOOE OWPING FRO 1HZl RE1 M I Pm!x WlI~IJnE

C 6 0.0213526 0.6287557 1.0000000 124.125559 19.2038697
C 7 0.0721698 0.7764332 0.4830101 35.096527 9.2756624 VII. MODE SHAPE VIA PROW ANALYSIS
C
-
9 0.2060205 0.860U318 0.1321199 -127.891332 2.1531365
C *.PRONY S-TABLE FOR PAS NUMBER 9: TSPAN
C .SIGNRL-TO-NOISE
5 RRIIO
O I S P L R Y T A l N LEVEL-
- C 169.00. 1060.003
65.2870 FOR NMOOES
O.looMK*OO
56 - Figures 6-10 and Table I11 extend the analysis of
Section VI to other "index" generators in the western
MOOE OAMPING FRO IHZI REL H I PHASE AMPLITUOE system. The 0.628 Hz "Kemano Mode" is visible at
C S 0.0212092 0.6280298 1.0000000 -82.598303 17.368V915
C 6 0.0595462 0.706L285 0.1168129 68.606910 2.0288556 each machine, both as a sharp peak in the frequency
C 7 0.0877736 0.7698999 0.5908038 -193.795953 10.26157011 response and as a consistent mode in the PRS locus.
5
-
IO 0.2326938 1.1937891 0.1068583 -118.91125311 1.8559677

C
C
.mPRONY 5-TRBLE FOR PRS NUMBER 5s TSPRN
.SIGNRL-TE-NOISE RATIO
O I S P L R Y T R l N LEVEL.
- C 2011.00. 1100.003
72.8790 FOR NMOOES
O.IOOOOE+OO
56 - The PRS tables show Grand Coulee and Shasta about 180
degrees behind Kemano for this mode, and Palo Verde
about 370 degrees behind.
c
C MOO€ DAMPING FRO I H Z I REL H I PHASE ANPLIIUDE
C 6 0.0215031 0.6284879 1.0000000 66.599927 16.1728315
C 7 0.06998611 0.7700647 0.2988102 55.219579 4.0239653 Except for Kemano, all signals show a significant
C 9 0.1962575 0.9072315 0.1158Y86 -38.773916 1.8735996
mode near 0.76 Hz. The indicated damping for this
mode is about 0.012-0.018--except at Grand Coulee,
TABLE 11. PRS tables for sliding-window fits to for which the damping is ten times higher. This
Kemano ringdown result is persistent enough to raise the possibility
 85

G COULEE 2 1 3 . 8 I COULEE 2 1 3 . 8 I '4 ELECTRICRL POYER Continuation of this analysis would provide a
R22/RII : INPUT =HAL MVR IUINDYDI cx 1.003
0 progressively better view of the phase distribution
0
-c m for the eigenvectors associated with key modes (for
+rm1 which machine electrical power is implicitly being
-10 taken as a state variable). Insight into the
--so interaction mechanisms for each mode can be extended
by examining complex power flow on major lines 1161.
v)
-20

' PIRSE .
-30
-U0
--U00

-YO m
w
Y
-50 m

'
U
Y

-60 1,.
0.0
, , ,, , , , , , ., , , . . , , ,
0.5
FREOUENCT I N HERTZ
1.0
* .. I . . .. I I
1.5
. I . 8 . - 8 1-360
2.0
-50
Loo z
=
Y
m
n.
-60

Fig. EA. Frequency response of Grand Coulee

electrical power to shunt reactive --I200
-10
control at Malin

. . . . . . . . .0.5. . . . . . . . . .1..0 . . . . . . . . .1.5. . . . . . . . . .2.0r

0.0
FREOUENCT I N HERTZ

Fig. 9A. Frequency response of Shasta electrical

power to shunt reactive control at Balin

0
, , , . , , , . .

500
, , , . . , . I . I
1000
, . I I I . I I . . ,
1500
~ ~ a - 1
1
2000
-20

l l M E I N CTCLES

Fig. 8B. Grand Coulee ringdown for brief W a r

load pulse at Malin

T I H E I N CYCLES

Fig. 9B. Shasta ringdown for brief Mvar load

pulse at Malin

REAL A X I S x IO-'

Fig. 8C. PRS locus for sliding-window fits to

Grand Coulee ringdown
REAL A X I S I 10.'

of 2 modes at essentially the same natural

frequency. The PRS tables indicate that Shasta is Fig. 9C. PRS locus for sliding-window fits to
swinging against Palo Verde. Shasta ringdown
86

G P R L O V R O I 24.0 I COULEE 2 1 3 . 8 I U ELECTRICRL POWER

-
R22/RLI : INPUT =HRL HVR IWINOHOI cx 1.003
C COULEE 2 GENERATION
-10
C *.PRONY
C
6
*SIGNRL-TO-NOISE -
5-TABLE FOR PRS NUMBER I : TSPRN
RATIO
D I S P L R Y T R I M LEVEL-
76.9200
O.lOOOOE*OO
C 99.00,
FOR NMOOES * 57
990.001

-20
kC MODE
5
ORMPING
0.1719990
FRO IHZ)
0.58119895
REL UT
0.66U3203
PHASE
-29.8U1l629
AMPLITUDE
8.5027202
C 6 0.0213050 0.6285593 0.2825898 2.2.11899 3.6169033
C 7 0.0935907 0.7885799 0.95768SO 55.072929 5.8579688
C 9 0.1081152 1.0315957 0.3U159911 -87.653386 U.37217U5
-30 C IO 0.0271096 1.1117995 0.3803902 -l93.306719 9.8680235
m

-
z
-U0
5 12
C 9.PRON'I
F
0.0191389 1.9166899

.
0.2276238
S-TABLE F O R PRS UWBER 2 4 TSPRU 1 89-00.
C 6IGNRL-TO-NOISE RATIO
DISPLRY T R I M LEVEI-
- -96.593526

62.2790 FOR NMODES I 57

0.10000E*00
2.9133867
980.003

-
z
U.
c?
2 MODE ORMPING FRO IHZI R E 1 UT PHRSE AMPLITUDE
C 5 0.2252735 0.5069690 0.8897686 -l69.3l2289 5.1761101
C 6 0.0208683 0.6289520 0.5996029 153.159691 3.2153256
-50 c 7 o.io39~11 0.7586371 0.8187653 -72.668157 4.790002~
C IO 0.0390882 1.1175969 1.0000000 106.202965 5.8502755
C 11 0.0570722 1.2150878 0.3609350 163.975121 2.1086990
C 12 0.0203290 l.111SO188 0.47971S6 -60.196025 2.8064682
-60
C 13 0.0993880 1.5523321 0.1311808 -169.160932 0.7679939

-
C SHRSTR CENERRTION

-70 C *SIGNAL-TO-NOISE
C
-
C *=PRONY S-TABLE FOR PRS NUMEER 1: TSPAN
RRTlO
D I S P L R Y T R I M LEVEL-
78.6570
0.10000E*00
C 99.00.
FOR NMOOES = 57
990.001

FREOUENCY I N HERTZ C MODE ORMPING FRO l H Z l REL UT PHASE AMPLITUDE

C 3 0.1661619 0.1983997 0.1290521 -9.279089 0.38991159
C 6 0.0208332 0.6295718 0.1511991 -0.712765 0.9691701
Fig. 10A. Frequency response of Palo Verde C 7 0.0165329 0.7600899 0.1393316 -87.189098 0.9323598
II I . 1930360 1.0000000 57.916009 3.1030989
-
C 0.0559060
electrical power to shunt reactive
C *.PRONY S-TRBLE FOR PRS NUMBER 2: TSPRN C 89.00. 980.003
control at Malin C .SIGNAL-TO-NOISE R R T I O .I 67.8190 FOR NMOOES = 56
D I S P L R Y T R I M LEVEL- 0.10000E*00
kC MODE
9
OAMPING
0.1596379
FRO IHZl
0.99852911
REL UT PHRSE
0.1101860 -126.258983
AMPLITUDE
0.2969681
C 5 0.020IU13 0.6286822 0.1516459 152.893U9l O.UO80208
C 6 0.0139213 0.7565785 0.12951120 103.788616 0.3350996
C 8 0.0873923 0.9806037 0.3299831 -65.325787 0.8799096
C 9 0.0609597 1.1160129 l.0000000 -3.692968 2.6906155
C 13 0.1598299 1.5760160 0. I369206 159.3731158 0.3689007
C
c
-
PRLO VERDE GENERATION

C .SlGNRL-TO-NOISE
C
RATIO
D I S P L A Y T R I M LEVEL-
-
C *=PRONY S-TRBLE FOR PRS NUMBER I: TSPRN C
71.0880 FOR NMOOES
O.IOOOOE+OO
-
411.00.
56
940.001

E MODE DRMPING FRO I H Z I

0.2537908
R E 1 UT PHASE
0.3889187 -125.115988
AMPLITUDE
1.3952108
C 2 0.0789932
C 3 0.0927809 0.3159691 0.1138085 9l.010910 0.3991532
= - 3 C 9 0.1099590 0.9582658 0.3959828 -127.855330 1.3696758
a . - c 5 0.0215992 0.~29u4vv 0.6143338 167.280802 2.1276230
5 : D C 6 0.0222897 0.7580719 0.5303981 108.103273 1.8369283
C 8 0.6530660 0.9975308 1.0000000 1.5908'49 3.9633010
C 9 0.0902369 0.9586993 0.7930519 -196.192701 2.7965775
C 10 0.0326951 1.0931821 0.3666928 16.532271 1.2699675
C II 0.0977920 1.2093322 0.1863706 -5.395985 0.6959575
C 0.0979320 1.3501917 0.2851505 -99.868328 0.9875622
-
12
kC
C
*SIGNAL-TO-NOISE -
=*PROHI 5-TABLE FOR PRS NUMBER 2: TSPAN
RRTIO
O I S P L R Y T R I M LEVEL-
62.9910
O.IOOOOE*OO
FOR NMDDES -
I: 89.00,
56
980.003

C
C MODE OAMPING FRO I H Z I REL UT PHASE RMPL ITUOE
1500 2000 C 2 0.0892829 0.2331630 0-1595Y70 -20.028538 0.7Y01033
TIME I N CYCLES C 3 0.0551856 0.3072306 0.1606265 -173.803898 0.7951106
C 5 0.0187972 0.6302957 0.3667230 -96.289759 1.7011966
C 6 0.0023850 0.76U6566 0- 1266312 - 1 1 O ~ O I J u 2 0.587U139
C 7 0.0590879 0.8138000 0.6002199 -110.655109 2.78112636
C 9 0.1960338 1.0095698 1.0000000 39.799072 U.6387779
Fig. 10B. Palo Verde ringdown for brief Mvar C IO 0.0360597 1.0991668 0.2893185 -99.696175 1.3188909
C I1 0.0589623 1.2081369 0.1717699 -96.377798 0.7968022
pu1,seat Malin C 12 0.1070655 1.3505293 0.1828986 -87.170322 0.8189261

TABLE 111. PRS tables for sliding-window fits to

generator ringdowns

VIII. EVALUATION OF REMEDIAL ACTION SCEEMES VIA PROW

ANALYSIS
This section treats two separate large-signal
stability cases involving loss of the 3100 MW
Celilo-Sylmar WDC Intertie. In the reference case
only standard remedial actions are taken, and growing
oscillations occur at the Malin 500 kv bus. These
same actions are taken in the experimental case. In
addition, a 1000 MW discharge is taken from a
fictional superconducting magnetic energy storage
(SMES) unit near the Sylmar terminal of the Pacific
HVDC Intertie. In this scenario the unit can serve
as a spinning reserve equivalent for 2 minutes.

RERL R X l S x IO-' Comparison of Figures 11A and 11B shows that the SMES
power injection into the grid has reduced the level
of the oscillations, and that it may have improved
Fig. 1OC. PRS locus for sliding-window fits to their damping. A rough indication of their modal
Palo Verde ringdown content is provided by the (unscaled) autospectra of
Figures 12A and 12B, which show that the injection
has also affected the natural frequency of the
 1.1

-
J
z
2
1.0
c
a
L
f

0.9

Fig. 12B. Autospectrum of Malin voltage swings

(1000 MW SMES pulse)
0.8 I , , ' ' ' ' . ' ' , . ' ' ' ' . . ' ' , ' ' ' . ' ' . ' . I . . ' ' " ' . '
0 500 I000 1500 1

T I M E I N CYCLES

Fig. 11A. Malin voltage swings following 3100 mw

hvdc bipole loss (no SMES pulse)

1.1

U
J

c)

q- 1.0

Fig. 13A. PRS locus for sliding-window fits to

Malin voltage (no SMES pulse) TBAR = 15
seconds (225 samples)
0.8 l l l l l l l l l l l l l l l l l . ~ . l ~ ~ ~ . ~ ~ . ~ ~ I ~ ' ' ~ ~ ' ' . ~
0 500 1000 1500 2000

T I M E I N CYCLES

Fig. 11B. Malin voltage swings following 3100 mw

hvdc bipole loss (1000 MW SMES pulse)
-
N
I
-
v)

x
a
*
-
a
z
0

-
a
z

REAL A X I S I IO-'

Fig. 1 U . Autospectrum of Malin voltage swiogs Fig. 13B. PRS locus for sliding-window fits to
(no SMES pulse) Malin voltage (1000 MW SHES pulse) TBAR
= 15 seconds (225 samples)

IX. CONCLUSIONS
critical modes near 0.27 Hz and 0.70 Hz. The
associated PRS tables and PRS loci (Figures 13A,B) Prony analysis--and SIGPAKZ in particular--have
display these effects more clearly, in a quantitative already proven useful in the direct modal analysis of
format well suited to the refinement of pulse level power system response signals at BPA. The next phase
and timing. of development will emphasize application guidelines
88

and computational efficiency. The latter is strongly 131 J.F. Hauer, "An Overview of BPA Software for
affected by model order, which incurs n-cubed Signal Processing and Other Systems Analysis
complexity costs during singular-value decompositions. Tasks Involved in Identification and Control of
Power System Dynamics," letter to the WSCC 0.7
While Prony analysis may well assume a place Hz Oscillation Ad Hoc Work Group, Nov. 16, 1987.
alongside eigenanalysis and Fourier methods, it will
certainly not replace them. Each such tool has its 141 J.F. Hauer, "Description of Reduced-Scale FFT
own merits and applications, and provides a different Software Package SIG-PAKR," BPA working notes,
view into dynamic system behavior. May 23, 1988. Distributed to the WSCC 0.7 Hz Ad
Hoc Work Group, June 6, 1988.

1151 T. Kailath,"Linear Systems. Prentice-Hall:

Englewood Cliffs, 1980.

REFERENCES 1161 J.F. Bauer, "Initial Results in the Use of Prony

Methods to Determine the Damping and Modal
G.R.B. Prony, "Essai experimental et analytic Content of Power System Dynamic Response
...," J . 1'Ecole Polytech. (Paris), vol. 1, Signals," BPA Technical Report, October 5, 1988.
pp. 24-76, 1795.
[171 B. FriedlandFr, "Lattice methods for spectral
F.B. Hildebrand, "Introduction to Numerical estimation," Proc. IEEE, pp. 990-1017, Sept.
Analysis." New YorK: McGraw-Hill, 1956. (See 1982.
esp. Chapter 9.)
1181 S. Li and B.W. Dickinson, "Application of the
S.M. Kay and S.L. Marple, "Spectrum analysis-a lattice filter to robust estimation of AR and
m o d e m perspective," Proc. IEEE, pp. 1380-1419, ARMA models," IEEE Trans. Acoustics, Speech, and
Nov. 1981. Signal Processing, pp. 502-512, April 1988.

A.J. Poggio, M.L. Van Blaricum, E.K. Miller, and [191 P. Strobach, "Recursive covariance ladder
R. Mittra, "Evaluation of a processing technique algorithms for ARMA system identification," IEEE
for transient data," IEEE Trans. Antennas and Trans. Acoustics, Speech, and Signal Processing,
Propagation, pp. 165-173, January 1987. pp. 560-580, April 1988.

M.L. Van Blaricum and R. Mittra, "Problems and [20] J.F. Hauer, "Fast Damping Estimators for Use
solutions associated with Prony's method for with the Transient Stability Program," BPA
processing transient data," IEEE Trans. Memrandum, Feb. 20, 1987. Comrmnicated to the
Antennas and Propagation, pp. 174-182, January WSCC 0.7 Hz Ad Hoc Work Group on October 22,
1978. 1987.
D.W. Tufts and R. Kumaresan, "Singular value [21] J.F. Hauer, "Reactive Power Control as a Means
decomposition and improved frequency estimation for Enhanced Interarea Damping in the Western
using linear prediction," IEEE Trans. U.S. Power System-A Frequency-Domain
Acoustics, Speech, and Signal Processing, pp. Perspective Considering Robustness Needs ,"
671-675, Aug. 1982. invited paper in Application of Static Var
Systems for System Dynamic Performance, IEEE
D.W. Tufts and R. Kumaresan, "Frequency Publication 87TH0187-5-PWR, pp. 79-92.
estimation of multiple sinusoids: making linear Proceedings of a symposium presented at the IEEE
prediction perform like maximum likelihood," PES 1987 Winter Meeting, New Orleans, LA, and at
Proc. IEEE, pp. 975-989, Sept. 1982. the IEEE PES 1987 Surmner Meeting, San Francisco,
CA.
R. Kumaresan and D.W. Tufts, "Estimating the
parameters of exponentially damped sinusoids and 1221 J.F. Hauer, "Robust Damping Controls for Large
pole-zero modeling in noise," IEEE Trans. Power Systems," IEEE Control Systems Magazine,
Acoustics, Speech, and Signal Processing, pp. pp. 12-19, January 1989.
833-840, Dec. 1982.

R. Kumaresan, "On the zeros of the linear

prediction-error filter for deterministic John F. Hauer (S'59-SM'87) was born in Washington
signals," IEEE Trans. Acoustics, Speech, and State in 1936. He received the B.S. degree (sumna)
Signal Processing, pp. 217-220, Feb. 1983. at Gonzaga University in 1961, and the Ph.D. degree
at the University of Washington as a National Science
[lo] R. Kumaresan, D.W. Tufts, and L.L. Scharf, "A Foundation Graduate Trainee in 1968. Both degrees
Prony method for noisy data: choosing the were in electrical engineering.
signal components and selecting the order in
exponential signal models," Proc. IEEE, In 1961 and 1962 he was with the General Electric
pp. 230-233, February 1984. Company, working in the area of nuclear reactor
controls while enrolled in the Advanced Engineering
[Ill R. Kumaresan, L.L. Scharf, and A.K. Shaw, "An Training Program. In 1963 and 1964 he developed
algorithm for pole-zero modeling and spectral spacecraft navigation and guidance methods at the
analysis," IEEE Trans. Acoustics, Speech, and Boeing Company, for use in Lunar Orbiter design and
Signal Processing, pp. 637-640, June 1986. mission control. His subsequent doctoral research
addressed methods for designing safety factors into
[121 H.P. Frisch and F.H. Bauer, "Modem Numerical thrusting trajectories for interplanetary flight.
Methods for Classical Sampled Systems Analysis- From 1968 to 1975 he was a member of the Computing
SAMSAN I1 User's Guide." NASAIGoddard Space Science faculty at the University of Alberta, where
Flight Center, Greenbelt, Maryland, 1984. he became an associate professor in 1972. His
[COSMIC Program #GSC-12827] teaching responsibilities and research activities
 89

there centered upon constrained optimization of Labs., Seattle, the Research Triangle Institute,
dynamic systems. Since 1975 he has been with the Green Mountain Geophysics, and Ball Aerospace
Bonneville Power Administration, where his work deals Corporation.
with the identification, analysis and control of
power system dynamics. Dr. Scharf is a past member of the ASSP ADCOM and the
Editorial Board of Signal Processing. He is a past
Dr. Hauer is a member of the IEEE Power Engineering Associate Editor for the IEEE Transactions on
and Control Systems Societies. Acoustics, Speech, and Signal Processing. He was the
Technical Program Chairman for the 1980 International
Louis L. Scharf received the PH.D degree in Conference on ASSP. He is a member of Eta Kappa Nu.
.Electrical Engineering from the University of
Washington, Seattle, in 1969.

From 1969 to 1971, he was a member of the Technical Cedric J. Demeure was born in Bruxelles, Belgium, on
Staff at Honeywell's Marine Systems Center in September 11, 1960. He received the Engineering
Seattle. He was a Faculty member at Colorado State degree from the Ecole Nationale Superieure des
University, Fort Collins, from 1971 to 1981, where he Telecomnunications, Paris, France, in 1983, the M.S.
last served as a Professor of Electrical Engineering degree from the University of Rhode Island, Kingston,
at the University of Rhode Island, Kingston. He is RI, in 1984 and the Ph.D degree from the University
currently a Professor of Electrical and Computer of Colorado (Boulder) in 1988, all in Electrical
Engineering at the University of Colorado, Boulder, Engineering. In 1985, he worked for the radar branch
where he teaches and conducts research in signal of Thomson-CSF, Pairs, France, as research engineer
processing. In 1974 he was a Visiting Associate in signal processing.
Professor at Duke University, Durham, NC. In 1977 he
was at the University of South Paris, Orsay, where he He is currently working in the Department of
was a member of the Technical Staff in the CNRS Electrical and Computer Engineering at the University
Laboratoire des Signa- et Systemes, Gif-sur-Yvette. of Colorado (Boulder), as a research associate and a
In 1981 he was a Visiting Professor Superieure des lecturer. His main research interests are Digital
Teleconrrmnications, Paris. He has served as a Signal Processing, Statistical Modeling, and fast
Consultant to Honeywell Inc., the Applied Physics algorithms.