Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
1
AbstractProny analysis has been used to estimate oscillation
modes from ringdown responses in a power grid. When applying
Prony analysis, several factors must be considered to estimate the
modes accurately. In this paper, a general prediction model is
proposed for the Prony analysis. The influence of decimation
factors, model orders, and linear solvers on estimation accuracy
is studied using the Monte Carlo method with a goal of providing
a reference for applying Prony analysis to estimate
electromechanical modes.
Index Termsleast squares method, oscillation modes, phasor
measurement unit (PMU), Prony analysis, singular value
decomposition, total least squares method.
I. INTRODUCTION
y (t ) Al exp l t cos2f l t l
l 1
ci expi t
(1.a)
(1.b)
i 1
y[ k ] y kt ci zik for k 1, 2, , N
(2)
y k y[ k ] [ k ] .
(3)
i 1
y nd k
a1 y[(n 1)d k ] a2 y[(n 2)d k ] an y[k ]
y (n 1)d 1 y (n 2)d 1
y (n 1)d 2 y (n 2)d 2
y (n 1)d 3 y (n 2)d 3
y N 2d
y N d
y nd 1
a1 y nd 2
2 y nd 3
y N nd
y N
y1
y 2
y 3
Or Y a y
N nd n
or
(5)
(9)
(10)
z n a1 z n 1 a 2 z n 2 a n z 0 0
(6.a)
(11)
1
ln(zi ) for i 1, 2,, n
dt
imag (
)
i
Freqi
2
(6.b)
N n
n
vn1 1
*
SVD
V
1
SVD
0 V *
~ ~ ~
~ ~
Y U1 U 2 1 ~ ~1* and a PE V1 11U1* y
0 2 V2
(7)
(12)
(13)
(14)
Fig. 1. The time steps of the time series y[k] are marked out for constructing
each row of Eq. (6) for d=1 and d=3.
(8)
a LS pinv Y y
z20
zn0 c1 y[1]
1/ d
1/ d
z2
z1n/ d c2 y[2]
z1
(15)
( N 1) / d
z2( N 1) / d zn( N 1) / d cn y[ N ]
z1
y [1] z10
z 20
z n0 c1
1
/
1
/
d
d
z2
z1n/ d c2
y [2] z1
(16)
~ ( N 1) / d
z2( N 1) / d zn( N 1) / d cn
y [ N ] z1
(17.a)
1 N
2
(17.b)
k
N k 1
User choices on parameters and methods influence mode
estimation accuracy. According to the authors reviews, these
choices include but are not limited to model order n,
decimation factor d, and solvers of (6). The following two
sections will examine how these choices may influence mode
estimation accuracy.
concept and make it easy for readers to replicate and verify the
results. To simulate inter-area oscillations, there are two pairs
of complex modes in this simulation model: 1) {0.25 Hz,
7.0%} and 2) {0.39 Hz, 6.5%}, respectively in {Freq, DR}.
To simulate a PMU data stream, 30 seconds of ringdown are
generated at a rate of 60 samples/s. In addition to ringdown,
Gaussian white noise (t) is added in (18). The standard
deviation of (t) is adjusted to simulate a signal-to-noise ratio
(SNR) of 10 dB (as in Fig. 2).
y (t ) 2e 0.1102 t cos(1.5708 t 1.5 )
(18)
2e 0.1596 t cos( 2.4504 t 0.5 ) (t )
Applying (6)-(8) and (11)-(14) on the noisy ringdown
responses using the LS solver with n=24, d=12, the estimated
dominant modes are {0.25 Hz, 7.1%} and {0.39 Hz, 6.8%},
which are close to the true modes. Using the stepwise
regression method [12], the dominant modes are selected from
24 modes. The Prony fit reconstructed using (16) is shown in
Fig. 2. It can be observed that the Prony fit matches the
original ringdown components well. The average fit noise is
0.0678. The condition number of matrix Y in (6.b) is 33.
MAD i median
(mc)
mc{1, 2 ,,100}
i ,true
(19)
Fig. 3. Identified modes from ringdown responses with SNR=10 dB from 100
sets of ringdown data (zoom-in plots appear as insets).
Fig. 4. The MADs of 0.25 Hz mode estimates for different decimation factors
(d) using the LS solver (the MAD corresponding to Fig. 3 is delineated using
an arrow).
Fig. 5. The MADs of 0.39 Hz mode estimates for different decimation factors
(d) using the LS solver.
Fig. 7. The mean value of condition numbers for different decimation factors
(d) using the LS solver (the error bars show 1 standard deviation).
Fig. 8. The MADs of 0.25 Hz mode estimates from LS (dashed lines), TLS
(solid lines), and PE (dotted lines) solvers for the simple model.
Fig. 6. The mean value of average fit noise for different decimation factors
(d) using the LS solver.
is the MADs are comparable for the LS, TLS, and PE solvers
for all of the studied model orders. Fig. 10 shows the average
fit noise. Observe that the fit noises for LS, TLS, and PE
solvers are comparable. The higher order models and larger
decimation factors can continuously reduce the average fit
noise levels due to the high order noise model. Fig. 11 shows
the condition numbers of the Y matrix in (6.b). The condition
numbers are larger than those in the simple model. Other
observations are similar to the simple models. Similar
observations are made for n=12 and 48 and another mode at
{0.64 Hz, 3.9%}. To keep it concise, the related plots are not
included.
U.S. DOE and Dr. John Hauer at PNNL for their assistance.
The authors also appreciate the help from Mr. Bill Mittelstadt
of WECC and Dr. Dmitry Kosterev of BPA.
Fig. 11. The mean value of the condition numbers of matrix Y (the error bar
shows 1 standard deviation).
REFERENCES
Fig. 9. The MADs of 0.42 Hz mode from LS (dashed lines), TLS (solid
lines), and PE (dotted line) solvers for the 17-machine model
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
Fig. 10. The mean value of average fit noise using LS, TLS, and PE solvers.
[9]
V. CONCLUSIONS
Considering decimation factors, a general prediction model
is proposed for the Prony analysis. When applying Prony
analysis for inter-area mode estimation, it is shown that d and
n shall be increased until the fit noise cannot be significantly
reduced. The lower bound of d shall be determined by the
bandwidth of ringdown responses and (7). When the condition
number of matrix Y is small, a high order model is preferred
for estimating the modes. For high SNR, the difference
between the LS, TLS and PE solvers is insignificant.
[10]
[11]
[12]
[13]
[14]
VI. ACKNOWLEDGMENT
The authors would like to thank Mr. Phil Overholt of the