2015 IEEE 10th International Conference on Industrial and Information Systems, ICIIS 2015, Dec.

18-20,2015, Sri Lanka

Improving Matrix Pencil and Hankel Total Least

Squares Algorithms for Identifying Dominant
Oscillations in Power Systems

D. Prasad Wadduwage U. D. Annakkage

Department of Electrical Engineering Department of Electrical and Computer Engineering
University of Moratuwa, Sri Lanka University of Manitoba, Canada
Email: prasad@elect.mrt.ac.lk Email: Udaya.Annakkage@umanitoba.ca

Abstract-Matrix Pencil (MP) and Hankel Total Least Squares to SCADA data. However, these time series data need to be
(HTLS) algorithms approximate an input signal as a sum of processed using appropriate algorithms for deriving the useful
decaying sinusoidal oscillations. This paper presents a simple information. The main contribution of this paper is to show that
technique to improve these algorithms for extracting only the the conventional MP and HTLS algorithms can be improved
true oscillations present in the input signal. The rationale behind
using a simple approach to extract only the dominant modes
the improvement is the observation that the true modes present
present in the input signal. Thus, the improved algorithms can
in the input signal appear consistently irrespective of the order
of the linear approximation used in MP and HTLS algorithms. It
be used for online oscillations monitoring purposes.
is shown that the modes appearing consistently can be extracted
This paper is organized as follows. Section II presents the
using a sorting method. The improved MP and HTLS algorithms
related literature on monitoring the power system oscillations.
presented in this paper are used for monitoring the low-frequency
Section III shows that the conventional MP and HTLS algo­
oscillations in power system in an online environment. The
performances of the proposed algorithms are demonstrated using
rithms can be improved using a simple approach to extract the
synthetic signals and simulated signals of different test systems. true modes present in the input signal. The proposed oscillation
monitoring algorithms in this paper are given in Section IV.
Keywords-Matrix Pencil, Hankel Total Least Squares, Prony,
The performances of the proposed algorithms are demonstrated
inter-area oscillations, online monitoring
in Section V. Finally, Section VI gives the conclusions of this
study.
I. INTRODUCTION

From a mathematical point of view, an electrical power II. LITER ATURE REVIEW
system can be considered as a nonlinear dynamic system
In power systems analysis, insufficient damping of inter­
[I]. It is an inherent property of such a system to exhibit
area oscillations is identified as a small-signal stability prob­
oscillations subsequent to a disturbance. If the system is stable,
lem [I], [2]. When monitoring the said stability problem,
the oscillations present in the post-fault system decay expo­
the related literature addresses two operating modes of the
nentially with time. The decaying rates of these oscillations
system, namely, the ambient operation and the ring-down
provide indications about the oscillatory stability [2]. If these
operation [5]. Under normal operating conditions, the power
oscillations are poorly-damped, the power system is in an alert
system is in a quasi-steady-state with small load variations.
state where a preventive action needs to be taken.
This can be treated as ambient noise from signal processing
Among different types of oscillations present in the power point of view. On the other hand, the power system variables
system, the inter-area oscillations (0.1 - I Hz), which are the exhibit noticeable oscillations subsequent to a large-magnitude
oscillatory modes involving the rotors of generators in different disturbance and this situation is referred to as a ring-down.
areas have gained the attention of power system engineers. Different algorithms proposed in literature for monitoring the
This is due to the fact that one of the root causes behind small-signal stability of the power system under these two
some of the grid failures such as the blackout in 1996 in the situations are summarized in [5]. This paper focuses on online
Western Interconnection [3] was an unstable inter-area mode. monitoring of the said stability under ring-down conditions in
Therefore, it is important to monitor these oscillations in online the system.
and alert the system operator when they are not acceptable.
The measurement based approaches proposed in the lit­
The power systems are being monitored using Supervisory erature for monitoring the power system oscillations can be
Control and Data Acquisition (SCADA) Systems. However, categorized as parametric and non-parametric methods. The
the current trend in large interconnected power systems such parametric methods first approximate the measured data using
as in USA and Canada is to install Phasor Measurement an appropriate model and the system dynamics are identified.
Units (PMUs) in widely dispersed locations to monitor their In contrast, a non-parametric method works on the data itself
systems. The PMUs provide the fundamental phasors of system to estimate the system characteristics [5]. Prony, MP, HTLS,
voltages and currents with respect to a common reference [4]. Kalman filter, eigen system realization algorithm are linear
Further, these data are available at a higher rate compared parametric methods where as the algorithms based on the

978-1-4799-1876-8/15/$31.00 ©2015 IEEE 13

 2015 IEEE 10th International Conference on Industrial and Information Systems, ICIIS 2015, Dec. 18-20,2015, Sri Lanka

Hilbert-Huang transformation, Wavelet transformation and the TABLE I: Conventional MP and HTLS algorithms
Teager-Kaiser energy operator do not use linear approxima­
tions for the measured data. The algorithms based on short MP HTLS

[ ][ ]
time Fourier transform is an example of a non-parametric
method. These algorithms are summarized in [5]. Some of the L - Pencil parameter p - Order of the exponential fitting
L > p, N = L+ 1'.1 - 1
algorithms proposed in literature have been implemented [6] I) Construct the data matrix Y as I) Construct the data matrix H as
for online oscillations monitoring. yeO) yeLl yeO) Y(1'.1 - 1)
Y(I) Y(L+ 1) Y(I) Y(1'.1)
This paper focuses on the application of the linear paramet­
ric methods for online monitoring of power system oscillations.
A common limitation associated with these techniques is to yeN - L - 1) yeN - 1) Y(L - 1) yeN - 1)
identify the true dynamic behaviour of the power system since 2) Obtain the SVD of Y, Y = USVT 2) Perform the SVD of H,

usually a higher order linear model is used in these methods. H = USVT

3) Determine the dominant singular 3) Determine the dominant
It is evident that all the oscillatory modes generated by this values ofS, ,
say .M. singular values of H, say k,.
higher order linear model do not reflect the true dynamic
4) Construct [V] = [VI V2 . • . V111 ] 4) Rearrange the matrix as;

[�' ;J
behaviour of the power system. Therefore, it is important to
identify the true oscillations. H = [u1 Un]

I�n
The linear parametric methods used in this paper are MP
and HTLS algorithms. These two methods approximate the
input signal as a sum of decaying sinusoids [7], [8]. Improving 5) Construct [V;] [V;] by deleting
and 5) Obtain l U'tj and [ul] by

MP and HTLS algorithms for extracting dominant oscillations last row and first row of [V'] deleting first and last rows of
has not been widely discussed in the literature. Our main respectively. [ U'] respectively.
contribution in this paper is to propose an improved MP 6) Obtain the eigenvalues of the matrix 6) Determine the total least
algorithm and an improved HTLS algorithm which extract only [Y lt [Y2] square solution of
the true dominant modes present in the power system. We have [Yl] = US' [v;r
contributed in [9] that the conventional Prony analysis can be [Y2] = US' [V;r 7) Determine the eigenvalues
improved using a simple approach to extract only the dominant of Z.
oscillations present in an input signal. This paper shows that N - Number of samples inside the data wmdow
the conventional MP and the HTLS algorithms can also be
improved using the same method we proposed in [9].
response. These two low-frequency modes have frequencies of
0.25 Hz, 0.39 Hz and damping ratios of 7%, 6.5% respectively.
III. M ATHEM ATIC AL PRELIMIN ARIES
x(t) = 2e-O.ll02t cos(1.5708t + 1.57r)+
2e-O.1596t cos(2.4504t + O.57r)
This section presents the theory of the conventional MP (1)
and HTLS algorithms and the improved algorithms.

This signal was analysed using the conventional MP algo­

A. Theory of conventional MP and HTLS algorithms rithm using a 16 s (four cycles of 0.25 Hz mode) long data
The MP and HTLS algorithms approximate an input signal window at 10 samples / s sampling rate. This sampling rate is

f=1 aie,,;t cos(27r fit +!/Ji) where, p is the

y(t) as y(t) = L
a recommended reporting rate of a PMU by the IEEE standard
order of the linear model, fi, (Ji ai and 1/Ji are the frequency, for synchrophasor measurements in power systems [4]. The
real part of the eigenvalue, strength and the phase angle of the pencil parameter L was set as 1f.
'ith mode respectively. This linear prediction model is associ­ First two columns of Table II shows the frequencies and
ated with a characteristic equation in discrete domain given the real parts of the eigenvalues of the modes estimated by
by Y(z) = 1 + L�=1 aiz-k. The roots of this characteristic the conventional MP algorithm in the main data window
equation give the discrete domain poles of the linear system. (window 1). It is clear that the conventional MP algorithm
These poles can then be transformed into continuous domain generates number of fictitious modes in addition to the true
as Ai = t;.IT In Zi, where Ai is the eigenvalue in continuous modes. Table II also shows the results of the conventional MP
domain and tJ..T is the sampling time step. Table I summarises algorithm applied on two sub-windows (window 2 and window
the solution procedure used in MP and HTLS algorithms. 3) of different lengths inside the main data window. Now, it
can be clearly seen that the true modes present in the input
It is the general practice in these methods to construct
signal appear consistently irrespective of the length of the data
the data matrices, Y and H, so that an overdetermined set
window on which the MP algorithm is applied.
of equations are solved to reduce the effect of measurement
noise on the mode parameter estimation [7], [8]. This is the As can be seen from Table I, the modes are determined
reason for generating fictitious modes during the calculation in the MP algorithm as eigenvalues of a matrix which is
in addition to the true modes presents in the input signal. derived using the data matrix Y. The size of this data matrix
is (N L) x (L + 1) which also changes with the length
-

B. Improved MP and HTLS algorithms

of the data window. However, the eigenvalues generated by
the MP algorithm represent the dynamic behaviour of a linear
Consider the synthetic signal given in Eq. (1) which shows system. Thus, the true dynamic behaviour of the linear system
a scenario of two low-frequency modes excited in a measured should be reflected by some of these eigenvalues consistently

14
 2015 IEEE 10th International Conference on Industrial and Information Systems, ICIIS 2015, Dec. 18-20,2015, Sri Lanka

irrespective of the size of the data matrix. This is why the true TABLE II: Rationale behind the improved MP algorithm
modes present in the input signal appear consistently irrespec­
tive of the length of the data window. Table I also shows that Window 1 (0-16 s) Window 2 (O-IS
- s)
. Window 3 (0-14 s)
the HTLS algorithm determines the modes as eigenvalues of a Fre. Real part Fre. Real part Fre. Real part
matrix Z which is constructed by mathematical manipulations (Hz) of ev (Hz) of ev (Hz) of ev
0.25 -0.1102 0.25 -0.1102 0.25 -0.1102
on a data matrix H. The conventional HTLS algorithm was 0.39 -0.1596 0.39 -0.1596 0.39 -0.1596
also applied on the same data windows (window 1, 2 and 3 as 0.1722 -368.79 0.1956 -373.72 0.2171 -365.45
shown in Table II) and it was also observed that the true modes 0.3477 -377.95 0.3619 -365.44 0.3776 -378.45
0.6814 -369.21 0.4427 -363.22 1.2683 -370.42
present in the input signal appear consistently irrespective of 1.4607 -367.96 0.7346 -368.33 1.3802 -366.47
the length of the data window. These consistently appearing 1.3658 -365.27 0.9211 -374.86 1.523 -369.02
modes can be extracted using a sorting method. 1.5384 -372.79 1.039 -366.24 1.9352 -375.39
1.7755 -371.06 1.8489 -374.80 2.0604 -360.85
2.1142 -366.61 2.0801 -362.73 2.4688 -370.16
In this study, the Euclidean distances between the individ­ 2.5510 -368.78 2.3193 -366.33 2.6731 -375.82
ual modes in the complex plane were compared to extract the 3.2635 -372.78 2.4157 -378.82 2.7902 -366.73
consistently appearing modes. Assume that PI and P2 are the 3.3126 -369.11 2.5332 -371.91 2.9374 -373.D7
3.9293 -368.76 3.1282 -376.86 3.3354 -362.19
number of modes identified by the MP or HTLS algorithms on 4.2077 -373.43 3.2683 -369.16 3.3614 -367.27
two data windows. The modes satisfying the constraint given 4.2337 -364.68 3.3686 -364.73 3.8124 -377.59
in Eq. (2) were extracted as the true modes among these PI 4.3966 -361.26 3.8008 -367.84 3.8144 -369.03
4.8480 -375.32 4.2094 -366.54 4.4269 -365.79
and P2. 5.0 -369.85 4.2701 -371.29 4.5403 -372.02
5.0 -362.76 5.0 -367.27

if, -lUi - fj)2+ (O" i - O"j)2::'; T,i = 1,2 ,··· ,PI

j = 1 2 ··· ,P
, ,
i
2
f + fj (2)
f =
2
O"i + O"j
0" = ----"-
2
Where, T is a threshold assigned for the distance in order
to identify the close modes, and f and 0" refer to the frequency
and the real part of the eigenvalue of the true mode.
As shown in this section, the conventional MP and HTLS
algorithms can be improved to extract only the dominant System is Secure
modes present in an input signal. The steps of the improved
MP algorithm is shown below.

1) Specify the main data window length and the sam­

pling time step.
2) Generate multiple sub-windows inside the main data
window by reducing the length of the main data
window in steps of 1 s. Fig. I: Proposed oscillations monitoring algorithm
3) Apply the MP analysis individually on each data
window and extract only the eigenvalues with positive
frequencies less than 5 Hz.
4) Extract the true modes as explained in this section important to distinguish between the ambient condition and the
and determine their frequencies and damping ratios. ring-down condition. Subsequent to an onset of a ring-down
oscillation in the power system, the energy of the measured
Let us call the above algorithm as the Shrinking Window
Improved MP algorithm. The conventional HTLS algorithm signal in the data window, Ek = L7�::-1 I x(iW dramatically
improved in the same way is referred to as the Shrinking changes compared to the mean energy in the ambient operation
Window Improved HTLS algorithm.
[6]. This feature can be used to detect the onset of a ring-down
oscillation and hence to initiate the improved MP/HTLS algo­
rithm. In this study the improved algorithm is initiated, if Ek is
IV. PROPOSED OSCILL ATION MONITORING AL GORITHMS
continuously changing and if Ek ::;:, 1. 05 x rnean(EI : Ek-l)
The improved MP and HTLS algorithms presented in or Ek ::.; 0.95 x rnean(EI : Ek-l).
Section III-B can be used for online monitoring of low­
After identifying a ring-down condition, the measured sig­
frequency oscillations as explained in this section. The pro­
nal is sent through a lowpass filter. The purposes of having the
posed oscillation monitoring algorithm shown in Fig. 1 was
filter are, i) separate the low-frequency oscillatory modes in the
implemented in MATLAB in the PC environment.
measured signal, ii) reduce the influence of the measurement
The proposed algorithms are expected to be operated noise on the mode estimation, and iii) enhance the true mode
under ring-down conditions of the power system. Thus, it is extraction capability of the improved algorithms under low

15
 2015 IEEE 10th International Conference on Industrial and Information Systems, ICIIS 2015, Dec. 18-20,2015, Sri Lanka

SNR conditions. In this study, the specifications used to design noisy conditions [10]. The Monte Carlo method uses indepen­
the lowpass filter were, i) 2 Hz pass-band corner frequency, ii) dent trials to generate different instances of the random noise.
5 Hz stop-band corner frequency, iii) 0.2 dB ripple in the pass­ In this study, 100 independent simulations were done on this
band, and iv) 20 dB ripple in the stop-band. These attenuation regard. Three measures were used to analyse the performances
levels are same as those proposed in [4] to improve the of the algorithms. These three measures were: 1) number of
dynamic performance of the PMD. A finite impulse response trials in which only the true modes were extracted (rd, 2)
(FIR) filter was designed for this application. The filter output number of trials in which only one of the true modes was
signal is then processed by the improved algorithms. extracted (r2) , and 3) number of trails in which true modes
and fictitious modes were extracted (r3). If, 11 +12+13 < 100,
The proposed algorithms determine whether the oscilla­
none of the modes were extracted in some trials.
tions present in the system has adequate damping or not. If the
oscillatory modes are poorly-damped or negatively-damped, Fig. 2 shows one instance of the synthetic signal at 5 dB
the operator is alerted to initiate the necessary actions. If not and 20 dB respectively and the output signal of the lowpass
the oscillations will decay and the system enters into an ambi­ filter. The length of the main data window was set as 16
ent condition again. After concluding the oscillatory stability s. Table III and Table IV show the true mode extraction
status, the algorithm is reset to check for the occurrence of a capability and the statistical significance of the estimated
ring-down condition again. mode parameters with respect to the true mode parameters
considering the 100 independent simulations respectively.
A. Implementation

It is recommended to set the length of the main data

�
·············· ··············
r
·············· ········· .....� .............
T .
window as four cycles of the dominant inter-area mode. In
the cases where there is no idea about the frequency of the
.
··············

t r �
·············· ·············· ········....

dominant mode, it is recommended to set the length of the first

data window as 10 - 20 s. The lengths of all the subsequent
data windows can be selected using the actual frequency of
the dominant mode determined by the algorithm. Further, it 10 15 20 25 30

is recommended to use two sub-windows inside the main

data window where these two sub-windows are generated by
reducing the length of the main data window in steps of
1 s. The threshold value is recommended to be set at 0.05
to extract the true modes. Note that all these settings were
determined using sensitivity analyses. The data are available
at the recommended PMU reporting rates for the application.
These are 10, 12, 15, 20, 30, 60, 100 and 120 samples/s [4]. 15 20 30
Time (s)
However, a sampling rate of 10 samples / s is sufficient to
monitor low-frequency inter-area oscillations. Fig. 2: Synthetic signal and lowpass filter output signal

V. RESULTS AND DISCUSSION

This section presents the performances of the proposed

oscillation monitoring algorithms using synthetic signals and TABLE TIT: Sensitivity analysis of the proposed algorithm with measurement
simulated signals. In this study, the pencil parameter L of the noise

MP algorithm and the parameter !vI of the HTLS algorithm

were set as .Jt. Therefore, the sizes of the data matrices Y and
SNR (dB) Algorithm "11 "12 "13

H in MP and HTLS algorithms are 2f x (.Jt + 1) and (2f X .Jt ) 5 MP 94 6 0

HTLS 92 5 3
respectively. Further, the number of dominant singular values MP 99 0
20
of these matrices were identified when the energy ratio (ER) HTLS 96 I
value in Eq. (3) is sufficiently close to unity.

TABLE IV: Statistical significance of the proposed algorithm with measure-

(3) ment noise

Where, O:i is the ith singular value and p is the total number Algorithm Fre. Standard Real part Standard Damping
(Hz) deviation of eigenvalue deviation ratio (%)
of singular values of the data matrix . 0.2501 0.0008 -0.1099 0.0054 6.97
MP*
0.3899 0.0014 -0.1605 0.0073 6.53
A. Case A - Synthetic signal HTLS* 0.2502 0.0009 -0.1104 0.0049 7.00
0.39 0.0014 -0.1594 0.0082 6.48
The synthetic signal given in Eq. (1) was used to test MP** 0.25 0.0002 -0.1102 0.0009 6.99
0.3899 0.0003 -0.1597 0.0017 6.50
the performances of the proposed oscillation monitoring al­
0.25 0.0003 -0.1102 0.0013 6.99
gorithms under noisy conditions. It is the usual practise in HTLS**
0.39 0.0004 -0.1598 0.0015 6.50
related literature to use a Monte Carlo method to evaluate *SNR - 5 dB
the performances of the mode identification algorithms under **SNR = 20 dB

16
 2015 IEEE 10th International Conference on Industrial and Information Systems, ICIIS 2015, Dec. 18-20,2015, Sri Lanka

>00
The perfonnance index /1 represents the probability of Riqdo...."D CKOlWioa
extracting only the true modes present in the input signal. ,>0

As shown in Table III, this probability is above 90% which 200

$111'_"), ".t.IC
�a1ioD
highlights the true mode extraction capabilities of the proposed 1>
i!. 1>0
algorithms. Further, Table IV shows that the presence of •
�
measurement noise does not degrade the performance of the 100

algorithms. The frequencies and the damping ratios of the �

>0
modes do not deviate significantly from their actual values.
This is due to the fact that the input signal is first filtered by �----�----�'0�--�"�---t��---t,,�--,, O
Ti ...cO)
the lowpass filter which suppresses the impact of measurement
noise. Fig. 4: Active power flow along the tie line

B. Case B - Two-area Jour-generator test system

Two-area four-generator test system shown in Fig. 3 is TABLE V: Mode parameter estimations: 2-area 4-generator test system
widely used in power system literature to analyse the inter-area
f
oscillations [1]. For this study, this test system was simulated
in RSCAD software with detailed generator models, exciters,
Data
window
Algorithm

MP
I Fre.
(Hz)
0.6416
Real part
of ev
-0.1635
I Darn in
ratIO %
4.0533
J I
(6.93-14.93) s
power system stabilizers and turbine-governor models to have HTLS 0.6416 -0.1635 4.0533
(12.8-20.8) s MP 0.6424 -0.1627 4.0269
an inter-area mode at frequency, 0.642 Hz and damping ratio, HTLS 0.6424 -0.1627 4.0287
4.02%. The test set up was arranged as mentioned below to
monitor this inter-area oscillation in an online environment.
mode determined by the proposed MP/HTLS algorithms agree
with those determined by the small-signal analysis.

C. Case C - 16-generator 68-bus test system

The 16-generator 68-bus test system shown in Fig. 5 is a

reduced order equivalent of the interconnected New England
test system (NETS) and the New York power system (NYPS)
[11]. This test system was modelled using detailed generator
models and four inter-area modes were identified using the
small-signal stability analysis at the given operating point.
The system has two poorly-damped inter-area modes with
AREAl
frequencies 0.52 Hz, 0.7 Hz and damping ratios 1.47%, 1.53%
respectively. These two modes correspond to oscillation of
Fig. 3: Two-area four-generator test system generator 14 against the generator 16 and group of generators
in NETS against generator 16 respectively. The observability
calculations reveal that both of these two modes are observable
in the active power flows along the tie lines connecting NETS
The analog output signals corresponding to the three-phase and NY PS. Fig. 6 shows the active power flow along one of
voltages (va, Vb, vc) and currents ('ia, 'ib, 'ic) as shown in Fig. the tie lines connecting the buses 60 and 61. The fault was
3 were taken from RTDS, real time digital simulator and applied at time; t = 5 s.
fed to an actual PMU. The PMU measurements were then
ll{�"�i _1
given to a phasor data concentrator to obtain the time-tagged :I
j-------------------------------, i----------------------i

I C ell I
::OCll:
j\TJS _'fl'l'S
measurements of the measured phasors. The active power
flow through the transmission line was calculated using the I I 11 I
'
1 1
II 11 I
measured phasors. r---�' I
i-' __ ;
Fig. 4 shows the active power flow through the tie line ""
.
,
subsequent to a clearing of a solid three-phase bus fault in "
,
I
,
the line 7-8 very close to the bus 7 after 5 cycles of the " .
,

fundamental frequency without any topology change in the

network. This signal was given to the oscillation monitoring
algorithm and the dc component of the filter output signal was
removed before analysing it using the improved MP/HTLS
algorithms. Table V shows the mode parameters over two data
windows selected soon after and a few seconds later clearing ,
,
the disturbance. .. _______________________________ .. __________________ ��E.::'}_:

In this case also, the improved MP/HTLS algorithms Fig. 5: 16 generator 68 bus test system
extracted only the dominant mode of the system. Further, the
values of the frequency and the damping ratio of the inter-area

17
 2015 IEEE 10th International Conference on Industrial and Information Systems, ICIIS 2015, Dec. 18-20,2015, Sri Lanka

22
200 VI. CONCLUSIONS

160 It has been shown in this paper that the conventional

� 140 MP and HTLS algorithms can be improved using a shrink­
� [20 ing window approach to extract the dominant low-frequency
& 100 modes in ring-down oscillations in power systems. Using
B 80
�
these improved MP and HTLS algorithms, the paper also
40
presented two oscillation monitoring algorithms to be used in
w.
the online environment. It has been shown that the true mode
w w � � w w m w � �
Time (s) parameters determined by the algorithms are less sensitive to
the amount of noise present in the input signal. The simplicity,
Fig. 6: Active power flow along the tie line
true mode extraction capability, accurate determination of the
mode parameters, and the acceptable computing time are the
indicators of the applicability of the proposed oscillation mon­
itoring algorithms for wide-area monitoring of power system
The noise corrupted simulated signal at 30 dB SNR was oscillations.
processed by the oscillation monitoring algorithms. The data
window length was set as four cycles of the 0.52 Hz mode and REFERENCES
100 independent simulations were done. Consider the two data [1] P. Kundur, Power System Stability and Control. New York: McGraw­
windows, (8.2-16.2) s and (20.6-28.6) s which were selected Hill Inc., 1993.
soon after and few seconds later clearing the disturbance. [2] P. Kundur, J. Paserba, V. Ajjarapu, G. Andersson, A. Bose, C. Canizares,
The true mode extraction capabilities of the improved MP N. Hatziargyriou, D. Hill, A. Stankovic, C. Taylor, T. V. Custem, and
algorithm over the two windows were, 11 = 96 and 11 = 98 Y. Vittal, "Definition and classification of power system stability," IEEE
Trans. Power Syst., vol. 19, pp. 1387 - 1401, May 2004.
and those of the improved HTLS algorithm were 11 = 92 and
[3] D. N. Kosterev, C. W. Taylor, and W. A. Mittelstadt, "Model validation
11 = 96 respectively. Table VI shows the average values of the
for the August 10, 1996 WSCC system outage," IEEE Trans. Power
mode parameters considering the 100 independent simulations Syst., vol. 14, pp. 967-979, Aug. 1999.
and their statistical significances. The results shown are over [4] "IEEE standard for synchrophasor measurements for power systems,"
the data window of (20.6-28.6) s. IEEE Std. C37.118.1, Dec. 2011.
[5] J. Thambirajaj, E. Barocio, and N. F. Thornhill, "Comparative review
of methods for stability monitoring in electrical power systems and
TABLE VI: Mode parameter estimations: 16-generator 68-bus test system vibrating structures;' lET Trans. Gener. Transm. and Distrib., vol. 4,
no. 10, pp. 1086-1103, 2010.
[6] 1. Quintero, L. Guoping, and V. M. Venkatasubramaniam, "An oscilla­
I I I
Algorithm Mode Fre. Std. Real part Std.
(Hz) of ev tion monitoring system for real-time detection of small-signal instability

I I I
Mode I 0.5214 0.0015 -0.0467 0.0031 in large electric power systems," in Proc. IEEE Power Eng. Soc. General
MP Meeting, pp. 1-8, Jun 2007.
Mode 2 0.6975 0.0026 -0.0678 0.0059
[7] T. K. Sarkar and O. Pereira, "Using the Matrix Pencil method to

I I I
Mode I 0.5215 0.0019 -0.0483 0.0061
HTLS estimate the parameters of a sum of complex exponentials," IEEE
Mode 2 0.6969 0.0036 -0.0672 0.0072
Antennas Propag. Mag., vol. 37, pp. 47-55, Feb. 1995.

[8] 1. 1. S. Gasca and 1. H. Chow, "Computation of power system low­

order models from time domain simulations using a Hankel matrix,"
It can be seen that the proposed algorithms determine the IEEE Trans. Power Syst., vol. 12, pp. 1461-1467, Nov. 1997.
mode parameters accurately. Further, we have elaborated in [9] D. P. Wadduwage, U. D. Annakkage, and K. Narendra, "Identification of
[9] that the damping estimations significantly deviate from the dominant low-frequency modes in ring-down oscillations using multiple
true values when a data window soon after the disturbance Prony models," lET Trans. Gener. Transm. and Distrib, vol. 9, pp. 2206-
is used for the analysis. This is due to the nonlinearity of 2214, Nov. 2015.
the system response. Thus, it is recommended to discard the [10] N. Zhou, 1. W. Pierre, and D. Trudnowski, "A stepwise regression
initial damping estimations of the algorithm before initiating method for estimating dominant electromechanical modes," IEEE Trans.
Power Syst., vol. 27, pp. 451-458, May 2012.
any preventive control actions.
[II] B. Chaudhuri and B. Pal, Robust control in power systems. New York:
Springer, 2005.
In this study, the true mode extraction capabilities of
the improved MP/HTLS algorithms were investigated using
simulated signals on a PC having Intel Core i7 processor (3.40
GHz) with 8 GB RAM. The computational burdens of the
algorithms were acceptable for updating the mode extraction
results at an interval of around half a second. Furthermore, it is
recommended to store the frequencies and the damping ratios
of the true modes calculated over a predetermined number of
data windows, say 3 windows and check the consistency of the
parameters before indicating to the operator. This consistency
check would avoid the possibility of giving false alarms in the
online environment. We have discussed in detail in [9] how
to handle the missing data and abnormal data of actual PMU
measurements.

18