Professional Documents
Culture Documents
Combined Use of PCA and Prony Analysis For Electromechanical Oscillation Identification
Combined Use of PCA and Prony Analysis For Electromechanical Oscillation Identification
Abstract—In this study, a new PMU-based method of identi- the most well-known [3]. However, due to the complexity that
fication of electromechanical oscillations is presented exploiting power systems have gained, they are hard to perform and valid
Principal Components Analysis (PCA). The proposed method only for the point of interest, limiting their use [3], [4].
makes it possible to carry out analysis on a limited number of
the so-called Princial Components (PCs), obtaining information
on the behavior of voltages and currents at buses following a As an alternative, system identification methods are also
small perturbation. The analysis is completed using the Prony able to provide modal information. With the recent deployment
analysis on the PCs, which is able to determine the amplitude,
frequency and damping of a mode. To assess the effectiveness of the Phasor Measurement Units (PMUs), that provide am-
of the method, the algorithm has been applied on a test system bient data of voltage and current acquired in various points of
based on the Kundur Two-area system as well as on a real event. the grid, interest arises on the development of new techniques
for system identification in real time. In between the methods
Index Terms—Electromechanical Oscillation, Mode identifi- that have been studied, the subspace identification methods are
cation, Power system dynamics, Principal component analysis,
Prony analysis. some of the most prominent techniques. They include seve-
ral algorithms whereby the state-space realization is directly
I. I NTRODUCTION obtained from the measured system inputs and outputs. The
linear system parameters are obtained from the row or column
Electromechanical oscillations have been widely studied in subspaces of certain matrices that vary according to the chosen
the literature since the development of the first interconnected method. Some examples of these applications can be found
power systems [1]. Over the years, the development of many in [4]–[8]. Alternative approaches are spectral analysis, [9]–
technologies and regulators affected the damping properties of [12], or the use of statistical properties in a hybrid model-
the system, what may lead to unstable conditions or inadequate measurement based method for estimation of the dynamic
performance of operations [2], such that the problem has matrix of the system [13], [14].
always been of interest for stability studies.
The electromechanical oscillation modes are usually clas-
sified as follows as interplant, local or interarea modes. The The present study is focused on the use of PMU da-
interplant modes are associated with oscillations between units ta for electromechanical oscillations identification, exploiting
in the same power plant, and their frequency range is between Principal Components Analysis (PCA), a very powerful tool
1.2 and 2.5 Hz. The local modes are associated with the that allows the reduction of the order of a system [15].
swinging of one generator or power plant against the rest The technique is based on multivariate statistics analysis that
of the power system, with frequency of oscillations ranging allows to transform a number of possibly correlated variables
from 1 to 2 Hz. The interarea modes are associated with into a smaller set of variables called Principal Components
the swinging of one geographically well defined group of (PCs) [16]. Then, the output of the PCA is evaluated through
synchronous machines against other groups. The frequency Prony Analysis, to identify the frequency and the damping of
of this mode of oscillations ranges from 0.2 to 1 Hz. the modes under study.
Traditionally, the analysis of such modes has been approa-
ched through small perturbation stability models and modal The remaining of this paper is organized as follows: in
analysis regarding an operation point around which the sy- section II-A PCA is explained while Prony analysis is pre-
stem is linearized. The so-called Linear Ringdown Analysis sented in II-B. Section III presents the proposed identification
methods are based on data following transient events and are algorithm. In section IV results on a test system based on the
widely spread and established, being the Prony method, ERA Kundur Two-area system and on a real event are presented,
(Eigensystem Realization Algorithm), and the Pencil Method while conclusions are presented in section V.
II. BASIS OF P RINCIPAL C OMPONENT A NALYSIS AND the rows of P are the new basis for the representation of the
P RONY A NALYSIS columns of X; these are the PCs directions, while Y is the
A. Principal Component Analysis matrix of the PCs.
The new set of variables (Y) are called scores, that are the
PCA is a statistical procedure which uses a vector space
representation of the initial data set in the PCs space.
transformation to reduce the dimension of a large data set
Coefficients of the PCs (P) are called loadings and can
to a small set that still contains most of the information in
be considered as the weights for each original variable when
the large data set. In linear algebra, the definition of this
calculating the PC.
transformation takes place expressing the new set of data as a
An important issue of PCA is the concept of uncorrelation
linear combination of the original one. Using a mathematical
between PCs in the new basis. This is proved considering the
projection, the original data set can be interpreted in just a few
variance of the data in the original basis which is defined, for
variables, called Principal Components (PCs). This analysis
a random variable Z, as:
allows a more easy examination of spot trends, patterns and
outliers in the data.
σZ2 = E[(Z − µ)2 ] (6)
Considering a data set represented by the following m × n
matrix X where µ is the mean.
The best way to uncorrelate the original data is to find
x11 x12 . . . x1n
x21 x22 . . . x2n the directions where variance is maximized and use these
X= .
(1) directions to define the new basis. Considering a 1×n row
.. .. ..
.. . . . vector r=[r1 ,r2 ,. . . ,rn ] scaled by its mean value µr , the variance
xm1 xm2 . . . xmn is given by:
where the m rows, which represent variables, contain n 1
σr2 = rrT (7)
columns each, which represent samples, the goal of PCA is n−1
to transform matrix X into a new matrix Y, with the same If we have a second 1×n row vector s=[s1 ,s2 ,. . . ,sn ] with
dimensions of X, so that zero mean, we can define the concept of covariance, which
can be considered as the measure of how much two variables
Y = PX (2) will change together:
P is a m × m matrix defined as follows: 1
2
σrs = rsT (8)
n−1
p11 p12 . . . p1m
p21 p22 . . . p2m Extending the approach to the matrix X with m rows and
P= . (3)
.. .. .. n columns, which contains all the samples of a particular
.. . . .
variable in a row, we can define the following symmetrical
pm1 pm2 . . . pmm
square matrix:
Assuming the following quantities:
x1 x1 T x1 x2 T x1 xm T
• pk are the rows of P of size 1×m, k=1,2,. . . ,m; ...
T x2 x2 T T
• xi are the columns of X of size m×1, i=1,2,. . . ,n; 1 1 x2 x1
... xm x2
CX = XXT = (9)
. . .. .
• yi are the columns of Y of size m×1, i=1,2,. . . ,n. n−1 n − 1 .. .
. . .
.
Equation (2) can be re-written as: xm x1 T xm x2 T ... xm xm T
p1
p1 x1 p1 x2 ... p 1 xn
The diagonal elements of the matrix are the variances while
p2 p2 x1 p2 x2 ... p 2 xn the off-diagonal elements are the covariances (the matrix is
PX = . x1 x2 ... xn = . . . =Y
..
.
.
.
.
.
. . .
. called Covariance Matrix). Covariance can be considered as
pm pm x1 pm x2 ... pm xn a measure of how two or more variables are correlated each
(4)
other. PCA definition aims at finding directions such that the
variables in the transformed matrix are as uncorrelated as
Considering a generic column of Y
possible which means that the covariance of different variables
p1 x i
has to be as low as possible.
p2 x i In order to determine the matrix of transformed data Y=PX,
yi = . (5) the following constraints on the covariance matrix CY :
..
• maximize the information of each variable, measured by
pm x i variance (diagonal elements);
We have now that the jth coefficient of yi (pj xi ) is a • minimize the covariance between the transformed
projection of xi onto the jth row of P. This indicates that the variables (off diagonal terms).
original data X are being projected on the columns of P. Thus, Thus, a matrix P that diagonalizes CY has to be find.
Assuming that {p1 , p2 , . . . , pm } are orthogonal, it is B. Prony analysis
possible to exploit linear algebra properties to find a solution Prony analysis was developed by Gaspard Riche, Baron de
for the problem. In particular: Prony in 1795 to explain the expansion phenomena of some
gases [17]. It has been shown to be a viable technique to
1 1 1 model a linear sum of damped complex exponential of signal
CY = YYT = (PX)(PX)T = (PX)(XT PT ) =
n−1 n−1 n−1 uniformly sampled. The Prony analysis is not only a signal
(10)
1 1
= P(XXT )PT = PSPT = PCX PT analysis technique but also a system identification method
n−1 n−1
widely used [18].
where The Prony analysis tries to fit a linear combination of
complex damped sinusoids to a signal as in (15).
S = XXT (11)
L
is a m × m symmetric matrix proportional to CX .
X
y(t) = Ai eσi t cos(2πfi t + φi ) (15)
From linear algebra i=1
where
S = EΛE−1 = EΛET (12) • Ai is the amplitude of component i;
• σ i is the damping coefficient of component i;
where E is a m × m orthonormal matrix, which leads to • fi is the frequency of component i;
E-1 =ET , whose columns are the eigenvectors of S, while Λ is a • φi is the phase of component i;
m×m diagonal matrix that has the eigenvalues of S as diagonal • L is the number of components.
elements. If we now choose as rows of P the eigenvectors of
According to Euler’s theorem, the cosine function can be
S, we obtain that P=ET . Therefore:
written as a sum of exponential so that
1 1
CY = PSPT = ET (EΛET )E (13) ej2πfi t+ ejφi e−j2πfi t e−jφi
n−1 n−1 cos(2πfi t) = + (16)
2 2
Since E is a orthonormal matrix, ET E=I, and CY becomes: Putting (16) in equation (15) and setting t=kTs , where Ts is
the sampling period, the samples of y(t) can be rewritten as
1
CY = Λ (14) L
n−1 X
y[k] = Ci µki (17)
It is worth noting that matrix Λ contains the eigenvalues of i=1
S, since they are proportional to the the eigenvalues of CX . where
This leads to the goal of this procedure: if we carry out
the eigenanalysis of CX , we obtain the variance of the newly Ai jφi
Ci = e ; (18a)
transformed variables CY . The variances gives information 2
about the relative importance of each principal component.
The largest variance corresponds to the first principal compo- µi = e(σi +j2πfi )Ts (18b)
nent, the second largest to the second principal component, and
with i=1,2,. . . ,L.
so on. Once we have obtained the eigenvalues and eigenvectors
Prony analysis consists of three steps:
of S=XXT it is possible to sort the eigenvalues in descending
order and place them on the diagonal of Λ. After that, we can 1) Solve the linear prediction model, that is made by
construct the matrix E by placing the associated eigenvectors the observed set of data.
Let us write (17) as a linear prediction model (LPM):
in the same order to form the columns of E.
In this way, the goal of obtaining a matrix CY of the
transformed data that is diagonal is achieved. Principal com- y[k] = α1 y[k − 1] + α2 y[k − 2] + · · · + αL y[k − L] (19)
ponents directions (the rows of P) are the eigenvectors of S,
while rows are ordered according to the ”importance” of each where y[k] is computed for k=L,L+1,L+2,. . . ,N-1.
principal component. By inverting the matrix representation, the linear
coefficients αL can be derived from (19),
The process described can be summarized as follows:
1) Data matrix X;
y[L]
y[L − 1] y[L − 2] ... y[0]
α1
y[L + 1] y[L] y[L − 1] ... y[1] α2
2) S=XXT ; y[L + 2] y[L + 1] y[L] ... y[2] α3
eigen-decomposition of S as S=EΛE-1 =EΛET ; =
3)
.
.
.
.
.
. .. .
.
.
..
. . . . .
4) selection of a proper number of PCs, according to the
y[N − 1] y[N − 2] y[N − 3] ... y[N − L − 1] αL
magnitude of the eigenvalues λ (20)
5) set P=ET ;
6) Y=PX. or, in a more compact form,
d = Dα (21)
where D is a (N-L)×L matrix. If N>2L, vector α is
estimated by solving the over-determined least square
problem (21). Vector α can be computed using a pseudo-
inverse:
α = (DT D)−1 DT d (22)
2) Find the roots of the characteristic polynomial
formed from the linear prediction coefficients.
f [Hz]
49.94
49.92
is carried out. PCA and Prony analysis are applied on the time [s]
A 4.321 4.158
0.2
ζ 0.353 0.310
0.15 f 0.685 0.660
0.1
0.05
0
1 2 3 4 5 6 7 8 9 10 11
Analyzing the sign of the loadings, it can be seen that buses
in area 1 (1-2-5-6-7) have an opposite behavior with respect
Figure 6: Two-area system: loadings of the first principal to buses in area 2 (3-4-9-10-11). The loading related to bus
components 8 is very small: it means that bus 8 has low participation in
the second principal component. Comparing loadings result
0.4
in Figure 7 with modes in Figure 3 shows that the proposed
0.3
0.2
approach gives information on how generators behave as well.
0.1
Applying Prony analysis on the second component magni-
0 tude A, damping ζ and frequency f can be found. Table II
-0.1
compares Prony analysis results and values computed by the
-0.2
modal analysis, i.e., the interarea modes M. 25 and M.26 of
-0.3
-0.4
Table I. It can be seen that the proposed method well estimates
the interarea oscillation.
1 2 3 4 5 6 7 8 9 10 11
Figure 7: Two-area system: loadings of the second principal B. ENTSO-E system: real event simulation
components In this section, the proposed method is applied to a real
event which resulted in interarea oscillations on the European
Let us now consider the frequency of the Center of Inertia Network of Transmission System Operators (ENTSO-E)1 .
(COI), a rotational analogy to the center of mass of an object, ENTSO-E represents 43 electricity TSOs from 36 countries
defined, according to [22], as across Europe extending beyond the EU borders.
Pn The geographical area covered by ENTSO-E’s members is
k=1 Hk SbGk fGk divided into five synchronous areas:
fCOIj (t) = P n (26)
i=1 Hk SbGk • Continental Europe
where, for each generator, Hk is the inertia, SbGk is the • Nordic Europe
nominal power and fGk is the electrical frequency. • United Kingdom
Figure 8 shows the shapes of both the first component and • Ireland
the frequency of COI: the two trends results very similar. The • Baltic Region
non perfect matching can be explained by the fact that the first Synchronous areas are groups of countries that are connec-
PC accounts only for the 78% of the total variance. ted via their respective power systems. The system frequency
2) Second component: The shape of the second component is 50 Hz and it is synchronous within each area; a disturbance
can be fitted by a damped sinusoid described by the formula: at one single point in the area will be registered across the
entire zone. Individual synchronous areas are interconnected
y(t) = Ae−λt cos(ωt + φ) through direct current inter-connectors.
ENTSO-E report [23] provides information about PMUs
It is worth noting that this behavior can be related to an
measurement locations depicted in Figure 9.
oscillation that shows up in the grid after the load variation.
In this real case:
• the PMUs are located only in few busses of the network;
0.2 50.02
• a moving window is applied to simulate real time
0.15
operation;
50
PC1
COI • the methodology to start the PCA analysis based on the
0.1
49.98
variance is applied and tested.
0.05
f [Hz]
49.96
0
during the considered perturbation. A significant oscillation at
49.94
-0.05 bus in Portugal (PT) can be clearly seen, which is in agreement
-0.1
49.92 to the fact that the perturbation was in the Iberian peninsula.
-0.15
0 5 10 15 20 25 30 35
49.9 1 The PCA algorithm was tested on some real recordings thanks to the
time [s]
Latent PC1
Threshold
5
variance
3
0
0 100 200 300 400 500
t [s]
Figure 9: European interarea modes Figure 12: Real event: PC variance and alarm setting
50.1
50.06
TKThessalonikiHz
50.08 TKTemelliHz
TKHamitabatHz
TKAtaturkHz
50.06
50.04 PTRecareiHz
ITLainoHz
ITBrindisiHz
GRStefanosHz
50.04 CHBassecourtHz
50.02
50.02
f [Hz]
f [Hz]
50
50
49.98
TKThessalonikiHz 49.98
TKTemelliHz
49.96 TKHamitabatHz
TKAtaturkHz
PTRecareiHz
ITLainoHz 49.96
49.94
ITBrindisiHz
GRStefanosHz
CHBassecourtHz
49.92
0 100 200 300 400 500 600
t [s] 49.94
235 240 245 250
t [s]
0.05
and Greece (GR) have opposite sign with respect to all the
others.
f [Hz]
-0.05
0.05
Analyzing the biplot graph reported in Figure 16, it is
f [Hz]
0
possible to cluster buses in three different groups:
-0.05
-0.1
150 200 250 300 350 400 450 500
• The Bus in Iberian area with highest PC1 (area α);
t [s]
f [Hz]
0.08
50
49.95
0.06
49.9
150 200 250 300 350 400 450 500
t [s]
0.04
Frequency
0.6
0.02
f [Hz]
0.4
0 0.2
0
150 200 250 300 350 400 450
-0.02
t [s]
0.8 Damping
-0.04
0.6
1/s
0.4
-0.06
0.2
0
-0.08
235 240 245 250 150 200 250 300 350 400 450 500
t [s]
Figure 14: Real event: PC1 and PC2 Figure 17: Real event: PCA+Prony analysis results for PC1
50.1
0.7
PC1 50.05
0.6
f [Hz]
50
0.5
0.4 49.95
0.3 49.9
150 200 250 300 350 400 450 500
0.2 t [s]
0.1
Frequency
0.8
0
0.6
f [Hz]
-0.1
0.4
-0.2
CHBassecourtHz GRStefanosHz ITBrindisiHz ITLainoHz PTRecareiHz TKAtaturkHz TKHamitabatHz TKTemelliHz TKThessalonikiHz 0.2
0
150 200 250 300 350 400 450 500
t [s]
0.5
PC2 0.6 Damping
0.4
0.4
0.3 1/s
0.2
0.2
0
0.1 150 200 250 300 350 400 450 500
t [s]
0
-0.1
-0.2
CHBassecourtHz GRStefanosHz ITBrindisiHz ITLainoHz PTRecareiHz TKAtaturkHz TKHamitabatHz TKTemelliHz TKThessalonikiHz
Figure 18: Real event: PCA+Prony analysisresults for PC2
0.4
TKAtaturkHz
TKTemelliHz
R EFERENCES
TKHamitabatHz
0.2
TKThessalonikiHz
GRStefanosHz
[1] J. Machowsky, J. Bialek, and J. Bumby, Power System Dynamics:
Stability and Control. Wiley, 2008.
Component 2