IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 26, NO.

3, AUGUST 2011 1733

Wavelet-Based Analysis of Power System

Low-Frequency Electromechanical Oscillations
José L. Rueda, Member, IEEE, Carlos A. Juárez, Graduate Student Member, IEEE, and
István Erlich, Senior Member, IEEE

Abstract—This paper addresses the use of the continuous performing eigenvalue calculation on the basis of a state-space
wavelet transform (CWT) in the study of power system low-fre- representation of the system, linearized about a given operating
quency electromechanical oscillations (LFEOs). Based on a point [5]. Notwithstanding, it is worth mentioning that the scope
modified complex Morlet mother-wavelet function, an approach is
of the analysis is limited to the region where the linear approx-
proposed to exploit the relationship between system low-frequency
oscillation features and the Morlet CWT of a system ringdown imation is valid [6].
signal for detection of modal parameter changes as well as for Over recent years, eigenanalysis from the perspective of non-
modal frequency and damping computation. Also, several guide- linear phenomena has been attempted by using the method of
lines for selecting the center frequency and bandwidth parameters, normal forms (MNF) [7]. MNF provides a systematic means of
the scaling factor and the translation factor, are proposed in order transforming a set of higher order nonlinear differential equa-
to provide reliable modal identification estimates. The efficiency
tions into a simpler form by using a nonlinear variable transfor-
of the proposed approach is confirmed by applying it to synthetic,
simulated and measured signals. mation [7]. Linearization-based eigenanalysis and MNF-based
eigenanalysis are considered offline methods, since they depend
Index Terms—Continuous wavelet transform, electromechanical on approximate data and modeling processes (i.e., model-based
modes, modal identification, power system signal records, small
signal stability. techniques) [8].
Alternatively, the extraction of oscillation parameters can
be accomplished via several measurement-based techniques.
I. INTRODUCTION These techniques are a good complement to model-based tech-
niques, since they rely on the actual response of a power system

I N a liberalized electricity market framework, power sys-

tems are prone to be operated closer to their technical
limits, due to several factors such as greater competition be-
(i.e., signal records obtained from system measurements or
from time domain simulations). Moreover, measurement-based
techniques are of interest for monitoring LFEOs in real time
tween agents, market pressures, reduced generation reserve [9]. As discussed in [10]–[12], typical measurement data can be
margins, increased demands for large power transactions from classified into three typical categories: 1) ringdown, 2) ambient,
the power producers to consumers over long electrical dis- and 3) probing.
tances and physical limitations in the transmission network [1]. Ringdown data occurs when the system is excited by some
Besides, these factors increase the possible sources of system large disturbance, (e.g., adding/removing large loads, gener-
disturbances, thus making the system more vulnerable to small ator tripping, severe short circuits) and results in a transient
signal stability problems [2]. These problems usually mani- response that can be easily observed and distinguished from
fest themselves as poorly damped, sustained or even growing ambient noise. This transient is usually assumed to be a sum
low-frequency electromechanical oscillations (LFEOs), typi- of several damped sinusoids [10]. Ambient data is related
cally in the range of 0.1 to 2 Hz, which can potentially lead to the fact that load switching throughout the day, which is
to unstable system operation with major consequences (e.g., random in nature, is persistently exciting the power system.
grid breakup, large-scale system blackout) [3]. Hence, accurate This disturbance commonly appears as noise with very small
determination of oscillation properties is of great importance magnitudes, and hence, they cannot be easily distinguished
for proper security evaluation and for controlling purposes [4]. from measurement noise [11]. Probing data is obtained when
Various techniques have been developed to provide informa- low-level pseudo-random noise is intentionally injected into
tion of troublesome LFEOs. Amongst these is the linearization- the system for testing its performance [12].
based eigenanalysis which commonly involves the determina- To date, several measurement-based techniques have been
tion of oscillation parameters (e.g., frequency, damping ratio) by developed, each one with its own advantages and limitations.
Previous comparisons between some of the existing techniques
Manuscript received July 26, 2010; revised August 04, 2010, August 05, have been reported in [13]–[15]. It has been pointed out that
2010, October 21, 2010, and November 26, 2010; accepted December 28, 2010. some techniques, such as Prony analysis [16], are appropriate
Date of publication February 04, 2011; date of current version July 22, 2011. for ringdown data when linear and stationary signal portions are
This work was supported in part by the University Duisburg-Essen and in part
by the German Academic Exchange Service. Paper no. TPWRS-00603-2010. considered. For ambient data and probing data, the stochastic
J. L. Rueda and I. Erlich are with the Department of Electrical Power framework has been widely employed. Thus, for instance, sev-
Systems, University Duisburg-Essen, Duisburg, Germany (e-mail: jose.rueda@ eral techniques are based on auto-regressive (AR) [17], auto-re-
uni-duisburg-essen.de; [email protected]).
C. A. Juárez is with the Instituto de Energía Eléctrica, Universidad Nacional
gressive moving average (ARMA) [18], auto-regressive moving
de San Juan, San Juan, Argentina (e-mail: [email protected]). average exogenous (ARMAX) [12], and stochastic subspace
Digital Object Identifier 10.1109/TPWRS.2010.2104164 models [11]. However, caution needs to be exercised when se-
0885-8950/$26.00 © 2011 IEEE
1734 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 26, NO. 3, AUGUST 2011

lecting the identified oscillation parameters since the estimates II. THEORETICAL BACKGROUND
depend on the model order. Additionally, the computation time
may become significant since these techniques use a significant A. Continuous Wavelet Transform
number of matrices. Also, alternative techniques have attempted
to employ a whitening filter [19] and Fourier-based methods Mathematically, the continuous wavelet transform (CWT) of
[20], [21], for estimating stationary oscillation parameters from a continuous time signal is expressed by the following inner
ambient data. product in the Hilbert space [27]:
Characterization of the temporal variation of the fundamental
oscillatory modes (i.e., detection of nonstationarities) has been
(1)
previously addressed in [22] and [23]. In [22], a Kalman
filtering-based technique attempts to detect mode parameter
changes based on the statistics of the power spectral density of where is a dilation or scale parameter, is a translation or time
the Kalman filter innovation (i.e., the difference between the shift parameter, and is complex conjugate of a mother-
measured power system output and the estimated Kalman filter wavelet function .
output). Nevertheless, prior knowledge of the power system It can be noted that the CWT is the sum over all time of the
transfer function, which may not be available in real applica- signal multiplied by scaled, shifted versions of the mother
tions, is necessary to formulate the system Kalman estimator. wavelet (which are also called son wavelets). Thus, the CWT
In [23], Hilbert spectral analysis (HSA) has been employed to possesses localization properties in both time and frequency do-
track the evolving dynamics of modal parameters. However, mains and consequently provides valuable information about
the ability of this approach to estimate modal parameters and at different levels of resolution and measures the similarity
to identify their temporal changes is obviated when applied between and each son wavelet [28]. This indeed implies
to signals with multiple modal components since it cannot that the CWT can be used for feature discovery by selecting a
adequately distinguish two separate modes unless there is a suitable mother-wavelet function (which behaves like a window
large difference in either frequency or damping ratio [14]. function in both frequency and time domains) [29]. Further-
more, it is worth mentioning that the ability of multi-resolution
The wavelet transform constitutes an alternative mathemat-
in the CWT can automatically filter out the noise from and
ical tool to deal with nonstationary analysis of LFEOs since
thus no additional filters are needed [24].
its multi-resolution property is particularly useful for recog-
nizing the features of a signal in both the time and frequency do-
B. Mother-Wavelet Function
mains [24]. Hence, several techniques have been recently devel-
oped for supervision of various power system disturbances [25] There are several different types of mother wavelet functions
and for monitoring power system oscillatory performance [26]. satisfying the admissibility condition, such as the Mexican hat,
Notwithstanding, further studies are still needed for better ex- Gabor, and Morlet, which can be selected according to the na-
ploitation of the wavelet transform concept in order to properly ture of the signal to be analyzed. It has been found that the com-
extract the characteristics of typical nonstationary power system plex Morlet wavelet is well-suited for the analysis of ringdown
signals. In this connection, this paper provides an approach to signals since it is able to unfold these signals in time and fre-
the analysis of LFEOs based on CWT. The contributions of the quency allowing for the decoupling of oscillatory modes [30].
paper to other previously reported wavelet-based modal identi- This property is especially useful for modal identification pur-
fication techniques consist basically in the use of a suitable for- poses as demonstrated in several applications reported in civil
and in mechanical engineering for the analysis of multi-degree
mulation of the Morlet mother-wavelet function to improve the
of freedom systems in [31] and [32].
time-frequency resolution and the proposal of several guidelines
The complex Morlet wavelet is defined as
for selecting the center frequency and bandwidth parameters,
the scaling factor and the translation factor. These will facili-
(2)
tate the identification of modal parameter changes within a time
window and also will enable estimation of modal frequency and
damping with good accuracy. where is the wavelet central frequency parameter and is a
normalization factor that is commonly assigned the value of
The remaining part of the paper is organized as follows: In
to ensure that the wavelet has unit energy [30]. Also, the com-
Section II, a brief review of the CWT as well as a discussion
plex Morlet mother-wavelet has been formulated using different
about its use as feature detection tool for LFEOs analysis is pro- values of [33]. By substituting (2) in (1), it can be noticed
vided. Then, mathematical derivations for the analysis of modal that constitutes a multiplier to the computed wavelet trans-
frequency and damping using ringdown data based on CWT are form. Hence, a value of unity is used in this paper for in
presented in Section III. The efficacy of the proposed approach order to avoid alteration of the wavelet transform modulus and
is demonstrated in Section IV by application to a synthetic test erroneous estimation of modal parameters. The Fourier trans-
signal, simulated transient signals and actual measured data. form of is then given by
Section V provides a discussion and suggestions for future re-
search. Finally, conclusions are summarized in Section VI. (3)
RUEDA et al.: WAVELET-BASED ANALYSIS OF POWER SYSTEM LOW-FREQUENCY ELECTROMECHANICAL OSCILLATIONS 1735

Fig. 1. Energy spectra of the basic complex Morlet wavelet. Fig. 2. Energy spectra of the modified complex Morlet wavelet.

and hence its dilated Fourier transform is III. PROPOSED CWT-BASED APPROACH

A. Modal Parameter Identification

(4) Power system ringdown signals characterizing LFEOs are
usually modeled as a sum of damped sinusoids, each one having
Note that is maximum at and the com- the following form [33]:
plex Morlet wavelet can be considered as a linear bandpass filter
whose bandwidth is proportional to or to the central fre- (7)
quency [33]. Besides, it can be interpreted that only the th os-
cillatory mode is strongly related to the scale [31]. or alternatively
The energy spectrum of is given by

(5)
(8)
Fig. 1 shows the energy spectrum of for both
and which are two typical inter-area oscillation fre- for lying in some interval. The exponential decay constant
quencies [33]. From visual inspection, it can be noted that the and the angular frequency correspond to the real and the imag-
spectrum for contains a zero frequency component inary components, respectively, of the th mode expressed as
that causes a singularity in the wavelet admissibility condition . M is the mode relative amplitude and is the mode
phase shift.
(given by ). Moreover, for a given value Let us represent by its Taylor series in the neighborhood
of the dilation parameter , the spectrum of the complex Morlet of the reference point
wavelet has a fixed bandwidth and is therefore limited in fre-
quency resolution which may prevent clear identification of in-
(9)
dividual oscillatory modes. Thus, in this paper, for the analysis
of power system LFEOs, the complex Morlet wavelet defined
in (2) is further modified by introducing a bandwidth parameter The use of (8) and (9) constitutes a good approximation of the
which controls the shape of the former wavelet. Therefore, analytic signal provided that the phase of the signal varies
the complex Morlet wavelet is reformulated as follows: much faster than the amplitude (i.e., asymptotic signal) [31]. By
substituting (8) into (1), the Morlet wavelet transform of is
(6) expressed as

The energy spectrum of is shown in Fig. 2 for a fixed

and for equal to 2 and 10 Hz. There is no
zero frequency component when . Hence, an important (10)
value of gives a narrower spectrum, thus allowing a better
resolution for the detection of low-frequency oscillatory modes.
Also, the modified complex Morlet wavelet provides a suitable Taking into account that the Morlet mother-wavelet is a wave-
compromise in terms of localization in both time and frequency form of effectively limited duration that has an average value of
for a given signal. zero (i.e., fast decay wave), the amplitude , as expressed in
1736 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 26, NO. 3, AUGUST 2011

(9), can be truncated at the first term around (i.e., the point If , then and
where the son wavelet reaches its maximum), then can be approximated by

(11)

(21)
By substituting , and substituting (6)
into (11) If the th mode related to the th scale gives a significant
contribution to (21), the term obtains its max-
imum at ; therefore

(22)
(12)
By substituting for , (22) can be rewritten in the form of
Rearranging (12), we get (13), shown at the bottom of the time-varying amplitude and phase angle
page, which can be rewritten as
(23)
(14)
By taking the logarithm of the modulus of

where (24)

(15) Thus, the real part of the th mode can be estimated from
the slope of straight line portions of the logarithm of the CWT
modulus. Similarly, the imaginary part of the th mode can be
estimated from the slope of the phase of the CWT, which is given
(16) by

(25)
By completing the square on the exponential arguments of
(15) and (16) and by solving the integrals It should be noted that the slopes of logarithm of the
modulus and phase of the CWT constitute the basis for
(17) identification of modal parameter changes within a time
(18) window since the straight line fits with
and versus time can capture a mode
By substituting (17) and (18) into (14) damping change and a mode frequency change, respec-
tively. By using linear regression analysis with a set of
data given by
, the real and the imaginary part of
(19) the th mode can be computed as follows:

Let , then

(20) (26)

(13)
RUEDA et al.: WAVELET-BASED ANALYSIS OF POWER SYSTEM LOW-FREQUENCY ELECTROMECHANICAL OSCILLATIONS 1737

oscillations within a Gaussian window. This improves the

location detection capabilities of the CWT [30].
In Section II, it was pointed out that an important value of
the bandwidth parameter provides a suitable resolution for
(27) the detection of low-frequency oscillatory modes and also pre-
cludes nonfulfillment of the admissibility condition. In this con-
The frequency and damping ratio of the th mode are then nection, sensitivity analysis performed on several typical power
given by and , respectively. system signal records indicated that it is appropriated to choose
So far, the previous analysis can be considered as the CWT- greater than 9.
based analysis of a single-mode signal case. As previously men- It is worth to mention that the CWT in practice is computed
tioned, multiple-mode signals are represented by a linear combi- for a predefined range of scales. Thus, it is required to determine
nation of single modal components like the one shown in (7). In a proper range that allows identification of individual oscillatory
Section II, it has been shown that the CWT constitutes a time- modes. When the complex Morlet mother-wavelet function is
frequency signal decomposition procedure that allows decou- dilated by a scaling factor , the center frequency becomes .
pling of signal individual modal components. Hence, since the Hence, if the underlying signal sampling period is , the scale
complex Morlet mother-wavelet function has compact support is associated to the modal frequency as follows:
in both time and frequency domains, the CWT of a multi-mode
(29)
signal can be expressed as follows:
Therefore, the range of scales can be determined to contain
(28) the scale . In the case of determining the value of , it is taken
at the peak of , since the spectrum of a ringdown
signal exhibiting LFEOs reaches a peak in a neighborhood of the
Again, the key idea is to choose a scaling factor that maxi- frequency associated to a dominant oscillatory mode. This peak
mizes the son wavelet, so that the frequency window covers only corresponds to that of the wavelet transform of the ringdown
the mode associated with while all the other modes and the signal.
signal noise are automatically filtered out. Also, for all modes Finally, for estimation of the real and the imaginary part of
of interest, the linear regression analysis is to be performed as the th mode, the translation factor is taken in the range of
well as for the single-mode signal case. positive small values where the plots of and
versus time are approximately linear, since
B. Guidance in Selection of Parameters for CWT Analysis the time window in this range of contains the information of
Morlet CWT-based modal identification is approximately modal parameters.
performed on the basis of (24) and (25) which depend on
the center frequency parameter, the bandwidth parameter, the IV. TEST RESULTS
scaling factor, and the translation factor. Hence, several guide- In this section, different signal records are used to demon-
lines are provided in this section for proper choice of these strate the efficacy of Morlet CWT-based modal identification.
parameters in order to yield good modal identification results. For this purpose, several routines have been written in Matlab.
The center frequency can be approximated to be near Prony analysis is performed using the Dynamic System Identi-
the typical modal frequencies (e.g., 0.2–0.5 Hz). The typical fication Toolbox (DSI) [34].
ranges of modal frequencies are usually known beforehand, or
at least, they can be preliminary estimated based on systematic A. Synthetic Test Signal
analysis via Fourier spectra analysis of power system signal The proposed CWT-based modal identification approach is
records, model-based eigenanalysis, and sensitivity analysis. first applied to a three-mode synthetic signal with the following
In addition, modal frequencies can also be solely determined parameters:
based on wavelet coefficients charts, wavelet contour plots, Mode 1) ; ; ; ;
and wavelet scalograms, since, from (1), it should be noted Mode 2) ; ; ; ;
that the son wavelets exhibit a clear relationship between Mode 3) ; ; ; .
scales and frequency. Lower scales correspond to the most The sampling rate is 20 Hz. Additionally, a sudden detri-
compressed wavelets which allow detecting rapidly changing mental change in damping (i.e., ) as well as a
signal features related to high-frequency components whereas slight shift in modal frequency (i.e., ) was in-
higher scales correspond to the most stretched wavelets which troduced to Mode 1 after 15 s in order to account for modal
allow detecting slowly changing signal features related to parameter changes. Fig. 3 shows a 30-s snapshot of the tem-
low-frequency components. Thus, for instance, the scalogram poral evolution of the synthetic ringdown signal whereas the
of the Morlet CWT of a signal with dominant LFEOs will contour plot of its Morlet CWT magnitude is shown in Fig. 4 for
produce relatively large wavelet coefficients at scales where the a range of scales between 2 and 30. The Morlet CWT possesses
oscillation in the wavelet correlates best with the low-frequency three salient peaks at the scales 21 [which is, from (29), approx-
signal features. Moreover, it is worth mentioning that Morlet imately equivalent to 0.38 Hz], 14 (equivalent to 0.57 Hz), and
wavelets with low values of contain only a few significant 7 (equivalent to 1.14 Hz) as highlighted with dashed lines in
1738 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 26, NO. 3, AUGUST 2011

Fig. 5. Contour plot of the Morlet wavelet transform magnitude for the syn-
Fig. 3. Three-mode synthetic transient. =04
thetic test signal with f : and f =7 .

Fig. 4. Contour plot of the Morlet wavelet transform magnitude for the syn- Fig. 6. Magnitude plot of the Morlet wavelet transform for the synthetic test
=04
thetic test signal with f : and f = 25
. signal with a = 21,f =04: , and f = 25.

Fig. 4. These peaks are taken as the scaling factors whose asso- method. From Fig. 6, note that there is a change in the slope of
ciated frequency windows cover each mode separately. Besides, the straight line segments of the logarithm of the CWT mod-
it is interesting to note that the highest peak is closely related ulus at approximately (see the gray dashed line in
to the most poorly damped 0.4-Hz mode. Indeed, this indicates the figure). This change corresponds to the above-mentioned
that the Morlet CWT is especially useful for identifying those damping decrement. Besides, a slight change is noticeable in
dominant oscillatory modes that may constitute possible threats the slope of the phase of the CWT depicted in Fig. 7 since only
to power system security. a slight variation in mode frequency has occurred (see the gray
Fig. 5 illustrates the contour plot of the Morlet CWT magni- dashed line in the figure). Based on (26) and (27), the new fre-
tude for . Note that there is no evidence of the 0.6-Hz quency and of mode 1 are estimated by taking in the range
and 1.10-Hz modes. This indicates that the spectral resolution 19–24 s. The identified frequency and are 0.38 Hz and 2.57%,
of the Morlet CWT is inadequate to resolve individual oscilla- which are quite close to those of and , respectively.
tory modes when values of lower than 9 are chosen, thus This demonstrates that the proposed Morlet CWT-based modal
supporting the criterion for selecting that was mentioned identification approach can be used to detect modal parameter
in Section III. Further analysis for other different values of changes as well as to accurately estimate what their new values
lower than 9, which are not shown here in the interests of space might be.
and brevity, also supports this observation. Similarly, for the 15-s time window prior to the modal pa-
The magnitude and phase plots of the Morlet CWT of the rameter changes and by considering the peaks of
synthetic signal for the first mode with are illustrated corresponding to different scaling factors for the three modes
in Figs. 6 and 7, respectively. Besides, the case with no detri- and different ranges of , the frequency and of each of the three
mental changes in frequency and damping (i.e., without non- modes are also estimated based on linear regression analysis of
stationarities) are illustrated in both figures with black solid the magnitude and phase of the CWT. The identification results
lines in order to show the tracking capability of the proposed are summarized in Table I, comparing with the above theoretical
RUEDA et al.: WAVELET-BASED ANALYSIS OF POWER SYSTEM LOW-FREQUENCY ELECTROMECHANICAL OSCILLATIONS 1739

Fig. 8. Three-area, six-machine power system.

Fig. 7. Phase plot of the Morlet wavelet transform for the synthetic test signal
with a = 21 ,f =04 : , and f = 25.

TABLE I
IDENTIFIED MODAL PARAMETERS FOR THE SYNTHETIC TEST SIGNAL

values. It is clear that the proposed approach can be equally ap-

plied to multi-mode signals with an acceptable level of accuracy.

B. Simulated Power System Example

Fig. 9. TASM—Simulated ringup data. Sampling rate = 20 Hz.
The proposed method is also tested with a three-area, six-ma-
chine benchmark power system (TASM) obtained by modifying
the well known two-area, four-machine dynamic test system efficacy of the proposed approach for multi-mode signals. Fur-
widely used for stability studies [5]. The single line diagram is thermore, white Gaussian noise is added to this signal as mea-
depicted in Fig. 8. All generators are modeled using the sub- surement noise to evaluate the performance of the proposed
transient model and equipped with fast static exciters and a method with respect to varying levels of noise. For this pur-
simple thermal turbine-governor system. This system exhibits pose, a Monte Carlo type of simulation is used to assess the ac-
five OMs: one inter-area mode, in which the generating units G1 curacy of frequency and damping estimates for both inter-area
and G2 oscillate against G3 and G4, a second inter-area mode, modes before and after (i.e., time at which a detri-
in which G5 and G6 oscillate against the rest of the system, and mental change is made in the system) by simulating 400 cases
three local modes, one in each area, associated with the oscilla- with independent realizations of noise at each different level of
tions of the generating units within each area. signal-to-noise ratios (SNRs). Figs. 10 and 11 illustrate a com-
Due to its topology, the inter-area modes are easily excited parison in terms of the mean square error (MSE) between the
by assuming a 4% step in the exciter reference voltages of the proposed method and the KT Prony method (which, compared
synchronous machines connected at buses 1, 11, and 21. Also, a to simple Prony analysis, allows performing modal identifica-
1% load increase at buses 4, 14, and 24 is applied at 15 s in order tion at relatively low SNR thresholds [38]).
to induce detrimental changes in the small signal stability per- For mode 1 (the most dominant and lowest frequency inter-
formance of the system. Time-domain simulations are accom- area mode), it is evident from Figs. 10 and 11 that the proposed
plished using the functions of the Power System Toolbox (PST) method outperforms the KT Prony method at any level of SNR
[37]. Ringup data in the form of active power flow on the line for both time windows (i.e., the lowest MSEs in frequency and
220-101 is shown in Fig. 9. The results of analyzing this signal damping), thus confirming that Morlet CWT-based method is
in the 15-s intervals prior and subsequent to the detrimental more resilient to noise than the KT Prony method. This may be
changes in damping and frequency are summarized in Table II, due to the fact that the ability of multi-resolution in the CWT can
comparing with the results obtained with linear eigenanalysis automatically filter out the noise. Also, note that the MSEs in
and the Kumaresan-Tufts Prony method (KT Prony). From the frequency and damping in the 15- to 30-s time window are lesser
table, it can be seen that the estimates from Morlet CWT are than those in the 0- to 15-s time window. This is probably related
closer to those from linear eigenanalysis, hence confirming the to the fact that the mode has become even more prominent in
1740 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 26, NO. 3, AUGUST 2011

Fig. 10. MSEs in frequency for (a) the 0–15 s time window and (b) the 15–30 s Fig. 11. MSEs in damping for (a) the 0–15 s time window and (b) the 15–30 s
time window. MCWT stands for acronym for Morlet continuous wavelet time window. MCWT stands for acronym for Morlet continuous wavelet
transform-based modal identification whereas KT Prony stands for acronym transform-based modal identification whereas KT Prony stands for acronym
for Kumaresan-Tufts Prony-based modal identification. for Kumaresan-Tufts Prony-based modal identification.

the second time window due to the detrimental change that has TABLE II
IDENTIFIED MODAL PARAMETERS FOR THE SIMULATED RINGUP SIGNAL
occurred in the system.
For mode 2 (the highest frequency inter-area mode), it can be
seen from Figs. 10 and 11 that, in general terms, the MSEs in
frequency and damping in both time windows are higher than
those of mode 1. This difference is even more noticeable in the
0- to 15-s time window and may be attributable to the lower pre-
dominance of the mode in the signal combined with its relatively
limited duration because of its faster decay. From Figs. 10(a)
and 11(a), note that below the threshold , the pro-
posed Morlet CWT method outperforms the KT Prony method,
but by contrast, the KT Prony method gives better results above
the threshold. This is reasonable since heavily damped modes method would provide better parameter estimates for less dom-
are well filtered with the KT Prony method. inant modes at high SNRs.
Similarly, both Morlet CWT and KT Prony methods are ap-
plied on the ringdown data shown in Fig. 12, which reflects C. Application to Real Measurement Data
the effect of adding power oscillation damping controllers to Data was obtained from [34], corresponding to growing os-
the system. Results are summarized in Tables III and IV. From cillations that occurred in the Western Systems Coordinating
Table III, one can see that the Morlet CWT-based method per- Council (WSCC) system on August 10, 1996. A record of the
forms the best as in the previous example. From Table IV, it is real power flowing on a major transmission line is shown in
further confirmed that the proposed method is more resilient to Fig. 13. In this case, the structure of the WSCC system became
noise as compared to the KT Prony method and that KT Prony such that it exhibited small-signal instability due to pronounced
RUEDA et al.: WAVELET-BASED ANALYSIS OF POWER SYSTEM LOW-FREQUENCY ELECTROMECHANICAL OSCILLATIONS 1741

Fig. 14. Contour plot of the Morlet wavelet transform magnitude for the mea-
sured power flow signal record with f = 0 23
: and f = 14 .

Fig. 12. TASM—Simulated ringdown data. Sampling rate = 20 Hz.

TABLE III
IDENTIFIED MODAL PARAMETERS FOR THE SIMULATED RINGDOWN SIGNAL

TABLE IV
MSE IN FREQUENCY AND DAMPING

Fig. 15. Magnitude plot of the Morlet wavelet transform for the measured
power flow signal record with a = 21,f = 0 23
: , and f = 14
.

that a low-frequency oscillatory mode is becoming consider-

ably more dominant in this system time-response as the tran-
sient progresses. Figs. 15 and 16 show the magnitude and phase
plots of the Morlet CWT, respectively. Examination of Fig. 15
indicates a prominent change in mode damping at around
, whereas a slight change in mode frequency can be seen
in Fig. 16. The resulting estimates for the straight line seg-
ments in the 20-s intervals prior and subsequent to the damping
and frequency change are summarized in Table V. In the same
table, comparisons with estimates obtained from Prony analysis
show a remarkable agreement, confirming the efficacy of Morlet
CWT-based modal identification.

V. DISCUSSION
Fig. 13. Power flow recording—August 10, 1996 WSCC system. The sampling By using different signals, the proposed Morlet CWT-based
rate is 20 Hz. method has demonstrated to be a viable alternative tool for iden-
tifying modal parameter changes and also for estimating modal
frequencies and damping ratios with good accuracy. Compared
0.26-Hz oscillations. Detailed analyses of this event and mod- with KT Prony method, effect of noise on estimation accuracy
eling issues have been explained clearly in [35] and [36]. of Morlet CWT is greatly reduced, especially for signals ex-
The contour plot of the Morlet CWT magnitude for this signal hibiting prominent poorly damped oscillatory modes. Notwith-
is illustrated in Fig. 14. Note that the peak around the scale 21 standing, the method does not outperform the KT Prony method
(equivalent to 0.22 Hz) becomes wider over time, indicating at all when estimating the parameters of lesser dominant and
1742 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 26, NO. 3, AUGUST 2011

• Further comparisons of the proposed method with existing

Fourier-based sliding window (FSW) methods will allow
extracting further conclusions with regard to noise sensi-
tivity, since FSW methods have a reputation of being re-
silient at low SNRs [42].

VI. CONCLUSION
In this paper, an approach to the analysis of low-frequency
power system oscillations based on continuous wavelet trans-
form is presented. The complex Morlet CWT has been used to
decompose power system ringdown signals into single oscilla-
tory modes. A mathematical framework for modal parameter
estimation and change detection in ringdown signals has been
Fig. 16. Phase plot of the Morlet wavelet transform for the measured power provided. Several guidelines are also proposed for proper choice
flow signal record with a= 21 ,f = 0 23
: , and f = 14 . of the center frequency parameter, the bandwidth parameter, the
scaling factor, and the translation factor, since these are crucial
TABLE V
for good identification results. The efficiency of the proposed
IDENTIFIED MODAL PARAMETERS FOR THE MEASURED POWER FLOW SIGNAL approach has been demonstrated using a synthetic test signal,
simulated transient signals, and actual measured data.

ACKNOWLEDGMENT
The authors would like to thank Prof. G. Plonka of the Faculty
of Mathematics of the University Duisburg-Essen, for several
insightful discussions about wavelet transform theory and signal
spectral analysis.

relatively well damped or even heavily damped modes. Further REFERENCES

research is being pursued in order to overcome this drawback
[1] V. Vittal, “Consequence and impact of electric utility industry restruc-
and to deal with some of the following issues. turing on transient stability and small-signal stability analysis,” Proc.
• Modal identification of ringdown signals containing very IEEE, vol. 88, no. 2, pp. 196–207, Feb. 2000.
dense modes. Unlike Fourier bandpass filtering, the width [2] K. Morison, L. Wang, and P. Kundur, “Power system security assess-
ment,” IEEE Power Energy Mag, vol. 2, no. 5, pp. 30–39, Sep./Oct.
of the bandpass filter in Morlet CWT is inversely propor- 2004.
tional to scale: the width of Morlet CWT filters decreases [3] G. Andersson, P. Donalek, R. Farmer, N. Hatziargyriou, I. Kamwa,
with increasing scale. Thus, the broader the support of a P. Kundur, N. Martins, J. Paserba, P. Pourbeik, J. Sanchez-Gasca, R.
Schulz, A. Stankovic, C. Taylor, and V. Vittal, “Causes of the 2003
signal in time, the narrower its support in frequency. The major grid blackouts in North America and Europe and recommended
converse relationship also holds [39], [40]. In this connec- means to improve system dynamic performance,” IEEE Trans. Power
tion, a systematic and adaptive technique is needed to find Syst., vol. 20, no. 4, pp. 1922–1928, Nov. 2005.
[4] L. Grigsby, Power System Stability and Control. Boca Raton, FL:
a suitable compromise in terms of frequency and time res- Taylor & Francis, 2007.
olutions in order to avoid possible negative effects on esti- [5] P. Kundur, Power System Stability and Control. New York: McGraw-
mation of damping ratios. Hill, 1994.
[6] A. Sarasua, “Analysis and modelling of nondamped oscillations in elec-
• By analyzing the convolution operation in (1) in light of trical power systems,” Ph.D. dissertation, National Univ. San Juan, San
the mother wavelet function in (6), it should be pointed Juan, Argentina, 2005.
out that, although the Morlet CWT is focused at a given [7] J. J. Sanchez-Gasca, V. Vittal, M. J. Gibbard, A. R. Messina, D. J.
Vowles, S. Liu, and U. D. Annakkage, “Analysis of higher order terms
time and represents the signal content in that vicinity, the for small signal stability analysis,” in Proc. IEEE Power Eng. Soc. Gen-
window extends equally into the past and future [41]. The eral Meeting, Jun. 2005, vol. 2, pp. 1745–1753.
span of this analysis window depends on both the mother [8] M. J. Gibbard, N. Martins, J. J. Sanchez-Gasca, N. Uchida, V. Vittal,
and L. Wang, “Recent applications of linear analysis techniques,” IEEE
wavelet function and scale being analyzed, near the ends of Trans. Power Syst., vol. 16, no. 1, pp. 154–162, Feb. 2001.
the signal. Hence, the wavelet’s analysis window may ex- [9] J. F. Hauer, D. J. Trudnowski, and J. G. DeSteese, “A perspective on
tend significantly beyond the length of data, deteriorating wams analysis tools for tracking of oscillatory dynamics,” in Proc.
IEEE Power Eng. Soc. General Meeting, Jun. 2007, pp. 1–10.
the identification accuracy (this phenomenon is known as [10] D. J. Trudnowski and J. W. Pierre, “Overview of algorithms for esti-
end-effect [24]), especially for mode relative amplitudes mating swing modes form measured responses,” in Proc. IEEE Power
and phase shifts when short time-snapshots of a signal & Energy Soc. General Meeting, 2009, pp. 1–8.
[11] H. Ghasemi, “On-line monitoring and oscillatory stability margin pre-
record (i.e., shorter than 10-s intervals) are considered. To diction in power systems based on system identification,” Ph.D. disser-
make up for this inadequacy, an integrated approach will be tation, Univ. Waterloo, Waterloo, ON, Canada, 2006.
built by employing the Morlet CWT in conjunction with an [12] N. Zhou, D. J. Trudnowski, J. W. Pierre, and W. A. Mittelstadt,
“Electromechanical mode online estimation using regularized robust
optimization problem that is formulated in a least-square RLS methods,” IEEE Trans. Power Syst., vol. 23, no. 4, pp. 1670–1680,
sense. Nov. 2008.
RUEDA et al.: WAVELET-BASED ANALYSIS OF POWER SYSTEM LOW-FREQUENCY ELECTROMECHANICAL OSCILLATIONS 1743

[13] J. J. Sanchez-Gasca and J. H. Chow, “Performance comparison of three [38] R. Kumaresan and D. Tufts, “Estimating the parameters of exponen-
identification methods for the analysis of electromechanical oscilla- tially damped sinusoids and pole-zero modeling in noise,” IEEE Trans
tions,” IEEE Trans. Power Syst., vol. 14, no. 3, pp. 995–1002, Aug. Acoust. Speech Signal Process., vol. ASSP-30, no. 6, pp. 833–840, Dec.
1999. 1982.
[14] T. J. Browne, V. Vittal, G. T. Heydt, and A. R. Messina, “A compara- [39] M. Misiti, Y. Misiti, G. Oppenheimand, and J. M. Poggi, Wavelet
tive assessment of two techniques for modal identification from power Toolbox™ 4 User’s Guide. Natick, MA: MathWorks, 2010.
system measurements,” IEEE Trans. Power Syst., vol. 23, no. 3, pp. [40] P. Kang and G. Ledwich, “Estimating power system modal parameters
1408–1415, Aug. 2008. using wavelets,” in Proc. 1999 5th Int. Symp. Signal Processing and Its
[15] D. J. Trudnowski, J. W. Pierre, N. Zhou, J. F. Hauer, and M. Parashar, Applications, pp. 563–566.
“Performance of three mode-meter block-processing algorithms for au- [41] T. Kijewski and A. Kareem, “On the presence of end effects and as-
tomated dynamic stability assessment,” IEEE Trans. Power Syst., vol. sociated remedies for wavelet-based analysis,” J. Sound Vib., vol. 256,
23, no. 2, pp. 680–690, May 2008. no. 5, pp. 980–988, 2002.
[16] J. F. Hauer, “Application of prony analysis to the determination of [42] M. Glickman, “Disturbance monitoring in distributed power systems,”
modal content and equivalent models for measured power system re- Ph.D. dissertation, Queensland Univ. Technol., Brisbane, Australia,
sponse,” IEEE Trans. Power Syst., vol. 6, no. 3, pp. 1062–1068, Aug. 2007.
1991.
[17] N. Zhou, J. W. Pierre, D. J. Trudnowski, and R. T. Guttromson, “Robust
RLS methods for online estimation of power system electromechanical
modes,” IEEE Trans. Power Syst., vol. 22, no. 3, pp. 1240–1249, Aug. José L. Rueda (M’10) was born in 1980. He received
2007. the Electrical Engineer diploma from the Escuela
[18] R. W. Wies, J. W. Pierre, and D. J. Trudnowski, “Use of ARMA block Politécnica Nacional (EPN), Quito, Ecuador, in 2004
processing for estimating stationary low-frequency electromechanical and the Ph.D. degree in electrical engineering from
modes of power systems,” IEEE Trans. Power Syst., vol. 18, no. 1, pp. the Universidad Nacional de San Juan, San Juan,
167–173, Feb. 2003. Argentina, in 2009.
[19] J. W. Pierre, D. J. Trudnowski, and M. K. Donnelly, “Initial results From September 2003 until February 2005,
in electromechanical mode identification from ambient data,” IEEE he worked in Ecuador, in the fields of industrial
Trans. Power Syst., vol. 12, no. 3, pp. 1245–1251, Aug. 1997. control systems and electrical distribution networks
[20] K. K.-P. Poon and K.-C. Lee, “Analysis of transient stability swings in operation and planning. Currently, he is pursuing
large interconnected power systems by Fourier transformation,” IEEE postdoctoral research at the Institute of Electrical
Trans. Power Syst., vol. 3, no. 4, pp. 1573–1581, Nov. 1988. Power Systems (EAN, by the acronym in German), as a part of a one-year
[21] P. O’Shea, “The use of sliding spectral windows for parameter esti- scholarship financed by the University Duisburg-Essen. His current research
mation in power system disturbance monitoring,” IEEE Trans. Power interests include power system stability and control, system identification,
Syst., vol. 15, no. 4, pp. 1261–1267, Nov. 2000. power system planning, probabilistic and artificial intelligent methods, and
[22] R. A. Wiltshire, G. Ledwich, and P. O’Shea, “A Kalman filtering ap- wind power.
proach to rapidly detecting modal changes in power systems,” IEEE
Trans. Power Syst., vol. 22, no. 4, pp. 1698–1706, Nov. 2007.
[23] A. R. Messina and V. Vittal, “Nonlinear, non-stationary analysis of
interarea oscillations via Hilbert spectral analysis,” IEEE Trans. Power
Syst., vol. 21, no. 3, pp. 1234–1241, Aug. 2006. Carlos A. Juárez (GS’07) was born in 1983. He re-
[24] S. Chen, J. Liu, and H. Lai, “Wavelet analysis for identification of ceived the Electrical Engineer degree from Univer-
damping ratios and natural frequencies,” J. Sound Vib., vol. 323, no. sidad Centroamericana José Simeón Cañas, San Sal-
1, pp. 130–147, Mar. 2009. vador, El Salvador, in 2005. Since 2007, he has been
[25] S. Huang, C. Hsieh, and C. Huang, “Application of Morlet wavelets to pursuing the Ph.D. degree in electrical engineering
supervise power system disturbances,” IEEE Trans. Power Del., vol. in the Instituto de Energía Eléctrica (IEE), Univer-
14, no. 1, pp. 235–243, Jan. 1999. sidad Nacional de San Juan, San Juan, Argentina, as
[26] S. Bruno, M. De Benedictis, and M. La Scala, “Taking the pulse of a part of a four-year scholarship financed through the
power systems: Monitoring oscillations by wavelet analysis and wide German Academic Exchange Service (DAAD, by the
area measurement system,” in Proc. IEEE PES Power Syst. Conf. acronym in German) Regional Program.
Expo., 2006, pp. 436–443. He has worked in the Electronic and Informatics
[27] S. Mallat, A Wavelet Tour of Signal Processing. San Diego, CA: Aca- Department of Universidad Centroamericana José Simeón Cañas during 2005
demic, 1998. and 2006. His main fields of interest are control and supervision of power sys-
[28] W. J. Staszewski, Identification Damping MDOF Syst. Using Time- tems, power system stability, and signal processing.
Scale Decomposition, vol. 203, no. 2, pp. 283–305, Jun. 1997.
[29] J. C. Goswami and A. K. Chan, Fundamentals of Wavelets: Theory,
Algorithms and Applications. New York: Wiley, 1999.
[30] P. S. Addison, J. N. Watson, and T. Feng, “Low-oscillation complex István Erlich (SM’08) was born in 1953. He re-
wavelets,” J. Sound Vib., vol. 254, no. 4, pp. 733–762, Jul. 18, 2002. ceived the Dipl.-Ing. degree in electrical engineering
[31] J. Lardies and S. Gouttebroze, “Identification of modal parameters and the Ph.D. degree from the University of Dresden,
using the wavelet transform,” Int. J. Mech. Sci., vol. 44, no. 11, pp. Dresden, Germany, in 1976 and 1983, respectively.
2263–2283, Nov. 2002. After his studies, he worked in Hungary in the
[32] B. F. Yan, A. Miyamoto, and E. Brühwiler, “Wavelet transform-based field of electrical distribution networks. From 1979
modal parameter identification considering uncertainty,” J. Sound Vib., to 1991, he was with the Department of Electrical
vol. 291, no. 1-2, pp. 285–301, Mar. 2006. Power Systems of the University of Dresden. In
[33] A. R. Messina, Inter-Area Oscillations in Power Systems. New York: the period of 1991 to 1998, he worked with the
Springer, 2009. consulting company EAB, Berlin, Germany and
[34] BPA/PNNL. Dynamic System Identification Toolbox. [Online]. Avail- the Fraunhofer Institute IITB Dresden, respectively.
able: ftp://ftp.bpa.gov/pub/WAMS_Information/. During this time, he also had a teaching assignment at the University of
[35] C. W. Taylor, “Improving grid behaviour,” IEEE Spectr., vol. 36, no. Dresden. Since 1998, he has been a Professor and Head of the Institute of
6, pp. 40–45, Jun. 1999. Electrical Power Systems at the University of Duisburg-Essen, Duisburg.
[36] D. N. Kosterev, C. W. Taylor, and W. A. Mittelstadt, “Model validation Germany. His major scientific interest is focused on power system stability
for the August 10, 1996 WSCC system outage,” IEEE Trans. Power and control, modeling and simulation of power system dynamics, including
Syst., vol. 14, no. 3, pp. 967–979, Aug. 1999. intelligent system applications.
[37] J. Chow and G. Rogers, Power System Toolbox Version 3.0. Dr. Erlich is a member of VDE and the Chairman of the International Feder-
Cobourg, ON, Canada, Cherry Tree Scientific Software, 2008. [On- ation of Automatic Control (IFAC) Technical Committee on Power Plants and
line]. Available: http://www.eagle.ca/~cherry/. Power Systems.