You are on page 1of 11

682 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO.

2, MAY 2007

Extraction of Dynamic Patterns From Wide-Area


Measurements Using Empirical Orthogonal Functions
A. R. Messina, Senior Member, IEEE, and Vijay Vittal, Fellow, IEEE

Abstract—An approach based on the empirical mode decompo- in which automatic and discrete control actions may take place
sition (EMD) technique and proper orthogonal decomposition is in a complex manner [1], [7].
proposed to examine dynamic trends and phase relationships be- In previous work [7], [8], the authors explored the use of
tween key system signals from measured data. Drawing on the
EMD approach, and the method of snapshots, a technique based Hilbert spectral analysis to characterize the temporal behavior of
on the notion of proper orthogonal modes, is used to express an critical system modes using both simulated and measured data.
ensemble of measured data as a linear combination of basis func- Central to the analysis of temporal variability, is the concept of
tions or modes. This approach improves the ability of the EMD empirical mode decomposition (EMD) [9], [10] which allows
technique to capture abrupt changes in the observed data. Ana- an arbitrary complex signal to be decomposed into its intrinsic
lytical criteria to describe the energy relationships in the observed
oscillations are derived and a physical interpretation of the system scales or mode functions (IMFs). This approach has been suc-
modes is suggested. It is shown that in addition to providing esti- cessfully applied to determine the instantaneous parameters of
mates of time dependent mode shapes, the analysis also provides a complex nonstationary data, and can be useful in characterizing
method to identify the modes with the most energy embedded in the abrupt changes in the observed oscillations [8]. Direct analysis
underlying signals. The method is applied to conduct post-mortem of each mode function, however, provides limited understanding
analysis of measured data of a real event in northern Mexico and
to transient stability data. of the underlying mechanisms causing the oscillations. In ad-
dition, it is not easy to extract from the set of IMFs, the em-
Index Terms—Empirical eigen functions, inter-area oscillations, bedded modal components and phase relationships within large
mode shape, phasor measurement units.
data sets.
In this paper the EMD technique and proper orthogonal
analysis [11] are used in a complementary manner to extract
I. INTRODUCTION
dynamic information from wide-area measurements obtained
using Phasor Measurement Units (PMUs). Proper orthogonal
decomposition is used to express the data as a linear combi-
T HE detection of temporal changes in the dynamic pattern
of system oscillations is a problem of great theoretical and
practical importance. Wide-area measurements provide the op-
nation of time-varying basis functions. This allows the energy
contained in the oscillations and the modes with the most
portunity to analyze and characterize inter-area swing dynamics energy to be clearly identified and provides a clearer picture
in complex interconnected systems. Extracting the dynamics of the time modulation of the modes of interest in the study of
of interest from the observed oscillations, however, is a com- nonlinear effects.
plex problem. This is particularly true in the study of measured, Paramount to this approach is the ability to compress the com-
inter-swing dynamics in which local and inter-area modes may plicated variability of the original data set into the smallest pos-
participate in the observed oscillations [1]. sible number of basic modes. The main idea is to simultaneously
Approaches such as Prony analysis, eigenrealization algo- extract dynamic patterns and phase relationships among critical
rithms, and block processing techniques have been successfully modes from wide-area simulations or measurements. A second
used to extract modal information from complex data sets objective is to reduce numerical problems encountered in the
[2]–[6]. As the number of measured signals increases, accurate definition of instantaneous parameters using numerical differ-
characterization of relevant modal behavior becomes difficult, entiation [8], [9].
especially in the presence of noise [3]. Also, the interpretation The viability of the technique is demonstrated on both tran-
may become less informative for nonstationary phenomena sient stability data, and post-mortem analysis of measured data
characterized by abrupt, randomly occurring periods of dif- of a real event in northern Mexico. Attention is focused on the
fering temporal scales as in the case of many forced oscillations identification of inter-area swing dynamics and the characteri-
zation of phase relationships between the modes containing the
most energy. Issues regarding the construction of reduced order
Manuscript received August 29, 2006; revised December 6, 2006. Paper no.
TPWRS-00562-2006.
models are also discussed.
A. R. Messina is with the Department of Electrical Engineering, The Center
for Research and Advanced Studies (Cinvestav), Guadalajara 45010, Mexico
(e-mail: aroman@gdl.cinvestav.mx).
V. Vittal is with the Department of Electrical Engineering, Arizona State Uni- II. THE EMPIRICAL MODE DECOMPOSITION TECHNIQUE
versity, Tempe, AZ 85287-5706 USA (e-mail: vijay.vittal@asu.edu).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. The Empirical Mode Decomposition (EMD) technique is a
Digital Object Identifier 10.1109/TPWRS.2007.895157 systematic method for numerically decomposing a time depen-
0885-8950/$25.00 © 2007 IEEE

Authorized licensed use limited to: CINVESTAV. Downloaded on April 13,2021 at 15:30:11 UTC from IEEE Xplore. Restrictions apply.
MESSINA AND VITTAL: EXTRACTION OF DYNAMIC PATTERNS 683

dent signal into its own intrinsic mode functions (IMFs) is optimal in the sense that the average least squares truncation
[9]. This decomposition is expressed in the form error,

(1) (4)

where the components represent the os- is minimized, where denotes ensemble average,
cillatory functions and is the residue, extracted through a and denotes the norm over . The ’s are
process called sifting. [7], [8], [10] provide details of the im- time dependent coefficients of the decomposition to be deter-
plementation used in this work. mined so that (3) results in a maximum for (4).
By construction, the IMFs are nearly orthogonal to each other, Following [11], assume that the field is decomposed into a
i.e., , for all , and the local mean value , and a fluctuating part
mean is close to zero. These properties are needed to define a
physically meaningful instantaneous frequency [9].
A critical step in the analysis of IMFs is the determination of (5)
embedded information. Previous work by the authors has shown
that the fundamental system characteristics can be split into an It follows that, a normalized basis function is optimal if the
oscillatory part and a trend as follows [7]: average projection of onto is maximized, i.e., [12]

(2) (6)

The optimization problem can be recast in the form of a func-


where the dominant IMFs, represent the tional for the constrained variational problem1
oscillatory behavior of interest directly related to physical char-
acteristics, while the rest of the terms are associated with fea-
tureless, often nonoscillatory characteristics. (7)
For signals with multiple dominant modes, extraction of
the relevant dynamics may be difficult. In addition, when the A necessary condition for the extrema is that the functional
signal-to-noise ratio of the analyzed signal is low, the anal- derivative vanishes for all variations , ,
ysis and interpretation of the IMFs is difficult. Moreover, the where is an arbitrary variation, i.e., .
method does not directly identify the frequency content in the It can be proved [11] that this condition reduces to
intrinsic oscillations. Hence, motivated by the near orthogo-
nality characteristics of the EMD decomposition, and in an
attempt to simplify the analysis procedure, the use of empirical
(8)
eigenfunction basis [11] to obtain further insight into the nature
of the IMFs is explored.
Defining , the problem
of minimizing (4) becomes that of finding the largest eigenvalue
III. EMPIRICAL ORTHOGONAL FUNCTION (EOF) ANALYSIS of the eigenvalue problem , subject to . In
other words, the optimal basis is given by the eigenfunctions
A. Analytical Aspects of EOF Analysis of (8) whose kernel is the autocorrelation function
. Among all linear decompositions, the EOFs
The proper orthogonal decomposition (POD) method [11], are optimal in the sense of capturing, on average, the most ki-
[12] is an optimal technique of finding a basis that spans an en- netic energy possible for a projection onto a given number of
semble of data, collected from an experiment or numerical sim- modes.
ulation. More precisely, assume that , , de-
notes a sequence of observations on some domain where
B. Calculation of POD Modes; The Method of Snapshots
is a vector of spatial variables, and is the time at which
the observations are made. The POD procedure determines em- For discretely sampled measured data, the integral time-av-
pirical orthogonal functions (EOFs), , , such erage can be approximated by a sum over the set of sampled
that the projection onto the first EOFs (a low order representa- data points [11]. In this case, the vectors
tion) , , represent
a set of snapshots obtained from the observed data at locations
.
(3) 1Given a function to maximize, f (P ), subject to the constraint g (P ) = 0,
0
the Lagrange function can be defined as F (P; j ) = f (P ) g (P ).

Authorized licensed use limited to: CINVESTAV. Downloaded on April 13,2021 at 15:30:11 UTC from IEEE Xplore. Restrictions apply.
684 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 2, MAY 2007

The set of data can then be written as the -dimension D. Numerical Computation of POMs
ensemble matrix, The procedure adopted to compute EOFs can be summarized
as follows.
1) Given a set of measurements, , compute the
covariance matrix . Remove the mean value by com-
.. .. .. puting .
. . . (9) 2) Solve the eigenvalue problem . Arrange the
eigenvalues in ascending order as .
3) Compute the POD basis vectors . Nor-
Under these assumptions, the actual
inte-
malize the vectors to ensure orthonormality.
gral (8) can be written as , where
4) Find the number of POD basis vectors that capture the de-
. Assuming the
sired amount of energy (%) by using (13). Compute the en-
EOFs to be of the form , where is a ergy of each mode from , .
coefficient to be determined, and substituting this expression 5) Compute the coefficients from (3) using the orthonor-
into (8), the problem of minimizing (4) can be recast as the mality conditions , as
problem of finding the largest eigenvalue of the linear equation where is the Kronecker delta.
6) Expand the data in terms of the truncated POD basis as
(10)

where is the autocorrelation (covariance) matrix defined as

and verify its accuracy.


(11) Refinements to this basic technique are discussed next and a
physical interpretation of the POMs is provided.
Matrix , is real and symmetric; it possesses a set of or- E. Physical Interpretation of POMs
thogonal eigenvectors, , with positive (real)
eigenvalues. The eigenvectors of are called proper orthog- The POMs have an interesting relationship with the normal
onal modes (POMs), and the associated eigenvalues are called (orthogonal) modes in mechanical systems [12], [13]. The key
the proper orthogonal values. point to observe is that both, the POMs and the mechanical
modes are orthogonal to each other; this suggests that in analogy
C. Energy Relationships with the mechanical counterpart, frequency (speed)-based sig-
Following Kerschen et al. [12], the energy of a vector se- nals might be used to approximate mode shapes.
quence building a matrix matrix is defined by the Frobe- To gain an insight into the physical interpretation of POMs,
nius norm consider an -dimensional system, , where
represents the mass (stiffness) matrix.
The free natural response is given by

(12)

Hence, the total energy of the vector sequence is equal to the


energy in its eigenspectrum . Defining the total energy (14)
as the energy of a given eigenvalue, can be
or
expressed as , . In this
paper, a practical criterion for choosing the number of relevant
modes is given by

where the ’s are the natural modes, and the terms repre-
sent the time modulation of the natural modes.
(13) Following [13], these displacements are then used to
build the ensemble matrix , where
, . Defining
where is the number of relevant modes. As explained later,
the eigenspectrum is dominated by a few terms, characterized
by the corresponding EOFs. The rest of the terms will contain .. ..
..
transient fluctuations, noise or could represent the contribution . . .
of some random process.

Authorized licensed use limited to: CINVESTAV. Downloaded on April 13,2021 at 15:30:11 UTC from IEEE Xplore. Restrictions apply.
MESSINA AND VITTAL: EXTRACTION OF DYNAMIC PATTERNS 685

and substituting (14) into the ensemble of data , gives

(15)

It thus follows that the correlation matrix may be written as

(16)

To verify that a modal vector is a POM, we post multiply (11)


by to obtain:

(17)
Fig. 1. Speed deviation of selected machines. (a) Case 1. (b) Case 2.
Further, using the orthogonality characteristics ,
for , (17) reduces to
Details of the system characteristics and base case condition
are given in [8].
Two contingency scenarios are selected for study.
(18) (i) Case 1. Simultaneous outage of the 400 kV line from
Chicoasen (MMT) to Juile (JUI) and the 400 kV line from
in which we have the equation shown at the bottom of the page. Temascal (TMD) to Puebla (PBD) on the 400 kV South-
The analysis shows that for a sufficiently large number of eastern-Central interface.
snapshots, , the terms vanish, and (ii) Case 2. Outage of the Merida (MDA) power station unit
. # 1 in the Peninsular system.
In other words the eigenvectors (POMs), , of , converge These cases illustrate both, linear and nonlinear behavior.
to the modal vector ; the columns of the left eigenvector are For clarity of illustration, 42 machines representing
the normalized time modulations of major generators in all six areas of the system are selected for
the eigen modes. analysis. Detailed numerical simulations for the above contin-
gencies were performed to generate the snapshots used in the
POD method. These solutions, depicted in Fig. 1, were com-
IV. APPLICATION TO TRANSIENT STABILITY DATA
puted using a time step of 0.0128 s and a time window of 30
To verify the ability of the method to extract the dominant s. For clarity, only a few plots representing dominant machines
features of complex oscillations, we first consider output data are shown. Fig. 2 provides the corresponding power spectra.
from transient stability simulations of a practical system. The For case 1 in Fig. 2(a), the power spectra show the presence of
system under study is the 6-area, 377-machine dynamic equiv- three dominant modes at about 0.30 Hz, 0.53 and 0.78 Hz; these
alent of the Mexican interconnected system used by the authors are the same inter-area modes discussed in [8]. The analysis of
in previous studies. case 2 reveals a dominant mode at 0.60 Hz associated with the

Authorized licensed use limited to: CINVESTAV. Downloaded on April 13,2021 at 15:30:11 UTC from IEEE Xplore. Restrictions apply.
686 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 2, MAY 2007

Fig. 2. Spectra of selected speed deviations. (a) Case 1. (b) Case 2.


Fig. 4. Temporal evolution of the dominant POMs for cases 1 and 2. (a) Case
1. (b) Case 2.

captures about 13% of the energy. The other modes capture less
that 7% of the total energy.
For case 2, the analysis identifies, essentially, two dominant
modes with about 80% and 17% of the total energy.
These results suggest that the time evolution of the system states
can be approximated by the reduced order model (ROM)
, where the ’s provide the temporal (ampli-
tude) evolution of the modes, and the ’s give the approximate
mode shape of the POMs.
The temporal variation of the POMs can be used to assess the
Fig. 3. Energy captured by the POMs for cases 1 and 2. strength and distribution of nonlinear effects as well as to re-
veal hidden structure of the data. Examination of the time evo-
lution of the eigenmodes for case 1, in Fig. 4(a), indicates that
interaction of machines in the south systems and two other peaks POM1, the leading mode, represents a nearly sinusoidal oscil-
at 0.33 Hz and 0.80 Hz. lation, i.e., , whilst POM2 evidences
a composite oscillation of more complex nature. In contrast, the
A. Computation of POMs temporal evolution of the three dominant modes for case 2, in
Fig. 4(b), shows three nearly sinusoidal oscillations suggesting
Following the general approach outlined in Section III, the an approximately linear behavior. These modes are numerically
correlation matrix (11) was obtained from the snapshots of the orthogonal, and have a physical interpretation in terms of linear
numerically integrated data i.e., , where normal modes in Section III-E.
for . Information on the nature of these modes can be gleaned from
Fig. 3 shows the eigenspectrum of selected sets of signals for the study of the spectra in Fig. 5. For case 1, comparison of
cases 1 and 2 computed to capture 99% of the signals’ energy. the power spectra of the POMs in Fig. 5(a) with Fig. 2(a) in-
Here, the horizontal axis shows the number of modes required dicates that the slowest mode at 0.28 Hz dominates the motion
to attain 99% of the average total energy; the vertical axis shows of POM1. The POD method captures a second mode [POM2
the energy in (13) captured by each POM. in Fig. 4(a)] with dominant peaks at 0.50 and 0.78 Hz that cor-
For case 1, proper orthogonal analysis of the records identifies respond to the linear modes in Fig. 2(a). POM 2 has no direct
eight modes in Fig. 3; the first or dominant mode, , counterpart in linear analysis, and can be understood by noting
captures nearly 67% of the total energy whilst the second mode that these two modes involve the interaction of machines in the

Authorized licensed use limited to: CINVESTAV. Downloaded on April 13,2021 at 15:30:11 UTC from IEEE Xplore. Restrictions apply.
MESSINA AND VITTAL: EXTRACTION OF DYNAMIC PATTERNS 687

Fig. 5. Spectra of the dominant POMs in Fig. 4. (a) Case 1. (b) Case 2.

Fig. 7. Mode shape of dominant POMs in Fig. 4(b). Case study 2. (a) POM 1.
(b) POM 2. (c) POM 3.

Fig. 6. Mode shape of dominant POM 1 in Fig. 4(a). Case study 1.

western and southeastern systems [8]. In addition, POMs 1 and


2 show evidence of interaction as suggested by the common fre-
quencies in the power spectra.
Fig. 8. Normalized mode shape of the 0.30 Hz mode 1 extracted using Prony
The analysis of the power spectra for case 2 in Fig. 5(b), con- analysis from transient stability output data. Case study 1.
firms that the POD technique is able to accurately extract the dy-
namics of the three dominant modes [see Fig. 2(b)] at 0.33 Hz,
0.60 Hz, and 0.78 Hz. In this case, these modes form a nearly phase of the speed deviation signals at the frequency of interest
de-correlated set of basis functions. considering one generator at a time.
Comparison of Figs. 6 and 8 for case 1, and Figs. 7(a)–7(c)
B. Mode-Shape Computation and 9(a)–9(c) for case 2, shows that EOF analysis identifies
both, the swing pattern as well as the machines most involved
Figs. 6 and 7 show the extracted mode shapes from EOF anal- in the oscillations (refer to Fig. 1). Due to the relatively large
ysis for selected modes in Fig. 4 using the ROM. For compar- number of signals that may be encountered in practical systems,
ison the mode shape computed using Prony analysis is shown in often involving several local and inter-area swing dynamics,
Figs. 8 and 9. Note that this is accomplished by computing the conventional analysis may be prohibitive.

Authorized licensed use limited to: CINVESTAV. Downloaded on April 13,2021 at 15:30:11 UTC from IEEE Xplore. Restrictions apply.
688 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 2, MAY 2007

Fig. 10. Schematic of the study area and its mechanical analogue. (a) Detail of
the measurement locations. (b) Mechanical analogue.

Fig. 9. Mode shape of dominant modes extracted using Prony analysis from
transient stability output data. Case study 2. (a) 0.60 Hz mode 2. (b) 0.32 Hz tests to interconnect the northwestern system to the Mexican in-
mode 1. (c) 0.78 Hz mode 3. terconnected system. These oscillations exhibited an oscillatory
pattern of changing characteristics.
Prior to this interconnection, the northwestern system oper-
A distinct advantage of the EOF analysis is that a large ated as an electric island. Following the interconnection of the
number of correlated data [limited by the eigenvalue problem systems, oscillations developed throughout the north, northern,
in (10)] can be simultaneously analyzed while capturing non- northeastern and southeastern systems causing the trip of load,
linear and temporal characteristics. In addition, all nonessential transmission resources and an independent generator. This
information such as noise is taken care of by the method. caused several changes in the observed dynamic patterns before
As noted above, while the mode shape is constant, the method the system was disconnected. Fig. 10(a) shows a schematic
allows for the identification of all temporal changes in the ampli- diagram of the study area indicating the actual location of
tude of the POMs; this can provide valuable information about PMUs used in this study whilst Fig. 10(b) show a mechanical
the essential stationary and nonstationary dynamics of the un- analog used in the studies below.
derlying process. Selected PMU recordings represent PMU measurements
(Macrodyne type PMUs 1690 series) at or adjacent to machines
in the northwestern-southeastern corridor. Data used in the
V. APPLICATION TO MEASURED MULTIMACHINE DATA study was acquired at a 2 ms sample rate over a 120 s window
period.
A. Data Used in the Study The time evolution of bus frequencies at four major sub-
stations along with relevant information about the temporary
The data used in this work was recorded using PMUs during interconnection preceding and following the oscillatory event
a real event in northern Mexico. Details of this event are given are shown in Fig. 11. For clarity of illustration, the time interval
in [7]. For completeness, we summarize the main events leading under study is divided into three subintervals: (06:27:42 to
to the observed oscillations. 06:28:06), (06:28:06 to 06:28:21), and (06:28:21 to 06:28:33)
On January 1, 2004 at 06:27:42 local time, undamped inter- which capture control actions and major trends in system
area oscillations developed in the northern systems following behavior.

Authorized licensed use limited to: CINVESTAV. Downloaded on April 13,2021 at 15:30:11 UTC from IEEE Xplore. Restrictions apply.
MESSINA AND VITTAL: EXTRACTION OF DYNAMIC PATTERNS 689

The analysis shows three time-varying modes whose frequen-


cies change slowly over time. Examination of the time evolu-
tion of the IMFs shows that the signals exhibit near time-sta-
tionarity for the three periods of interest, 06:27:42–06:28:06,
06:28:06–06:28:21, and 06:28:21–06:29:07. The first mode ex-
hibits a transition from about 0.25 Hz in the time period subse-
quent to the interconnection of the systems 06:27:42–06:28:06,
to about 0.61 Hz for 06:28:06–06:29:07. This transition coin-
cides with the outage of a major independent generator in the
northwestern system at 06:28:06 [see Fig. 10(a)]. This mode
is manifested throughout the northern systems and propagates
to the southeastern system through the north-south intercon-
nection. The second mode (IMF 2) starts at about 0.2 Hz to
Fig. 11. Temporal evolution of selected system signals. then decreases almost linearly for the time interval 06:27:42 to
06:28:06 and then increases to about 0.4 Hz between 06:28:06
and 06:29:07; the frequency then decreases to around 0.25 Hz.
The third mode exhibits nearly constant frequency content; it in-
creases from about 0.10 Hz during 06:27:42–06:28:06 to about
0.15 Hz.
Two problems are noted from this analysis: First, apart from
the nearly stationary period (06:27:42–06:28:06), it is difficult
to compare modal contents at separated system locations. Fur-
ther, visual inspection is subjective and produces a large number
of data. To circumvent these problems, the POD method was ap-
plied to identify dynamic trends and analyze phase relationships
between selected signals.
Drawing on the above EOF analysis, a systematic method for
eliminating IMF components that do not significantly contribute
Fig. 12. Temporal evolution of the dominant IMF at various measurement sites. to system behavior was developed. Two basis sets of signals are
being analyzed for feature extraction, namely frequency devi-
ations , and correlation measures be-
B. Determination of the Ensemble of Data tween selected sets of signals at different system locations, i.e.,
.
Based on measured data at various system locations , To assess the attractiveness of each selected set to investigate
the POD method was used to determine dynamic trends and to the phenomenon of concern a two stage-approach is used. In the
analyze phase relationships. Two procedures are investigated: first step the dominant modes are identified by using the above
the first is the conventional approach described in Section III-D approach; in the second step the system characteristics of the
(Procedure 1). The second approach is based on the combined resulting ROM are estimated using Hilbert /EMD analysis.
application of the EMD technique and EOF analysis (Procedure
2). The first approach allows the study of large sets of simula-
tions whereas the second can be used to study specific (probably C. Simulation Results—Procedure 1
nonlinear) nonstationary behavior arising from the interaction
of selected groups of machines in measured data. In this analysis, the ensemble matrix was obtained directly
Procedures 1 and 2 are used in the following analysis to from the PMU measurements for the case when the frequency
extract the dominant features of the observed oscillations. For (speed) signals at the Hermosillo , Mazatlan Dos
comparison and reference, the selected frequency signals are and Tres Estrellas substations, are used to build the en-
first decomposed into their underlying (nearly orthogonal) semble, i.e., .
modes (IMFs) using the EMD technique [3]. EOF analysis identifies two dominant POMs (not shown); the
To visualize the temporal dynamics of these modes and to first mode captures about 80% of the total energy whilst mode
allow comparison between them, the instantaneous frequency of 2 captures nearly 20% of the energy. Fig. 13(a) depicts the tem-
the three main IMFs at different system locations is computed poral evolution of these modes.
using the approach in [2]. Fig. 12 shows the time evolution of Comparison of the composite signal
the instantaneous frequency of the three dominant IMFs for the and the reconstructed signal in Fig. 13(b) shows that the two
bus frequency signals at the Hermosillo and Tres Estrellas sub- dominant modes accurately capture all salient features of the
stations at the end points of the northwestern-southeastern cor- phenomenon being investigated.
ridor obtained from the EMD technique. The time interval of To extract the underlying dynamics driving the oscillations,
interest is from 06:27:42 to 06:29:07 in which the system re- the approximate mode shape was determined from the eigen-
mained synchronized and the signals died out. vectors of the covariance matrix. This is illustrated in Fig. 14

Authorized licensed use limited to: CINVESTAV. Downloaded on April 13,2021 at 15:30:11 UTC from IEEE Xplore. Restrictions apply.
690 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 2, MAY 2007

Fig. 13. Time evolution of dominant PMOs and the reconstructed signal. (a)
Time evolution of dominant POMs. (b) Time evolution of the reconstructed
signal.

that shows the mode shapes associated with the two dominant
POMs for each interval of concern.
For the time interval 06:27:42 to 06:28:06, the analysis of
POM 1 shows an oscillation in which the frequency at Her-
mosillo and Mazatlan Dos is seen to swing in opposition to the
frequency at Tres Estrellas. The second mode reveals a local
(higher frequency) oscillation between the Mazatlan Dos and
Hermosillo signals. Fig. 14. Mode shape of dominant POMs for various time intervals. (a) Time
In the same vein, the analysis of time interval interval (06:27:42 to 06:28:06). (b) Time interval (06:28:06 to 06:28:21). (c)
Time interval (06:28:21 to 06:28:33).
06:28:06–06:28:21 shows, essentially, two highly co-
herent but out of phase groups (Hermosillo and Mazatlan Dos
and Tres Estrellas and Huinala). A third mode showing a local
interaction between Tres Estrellas and Huinala is also observed.
Similar conclusions can be drawn for the last interval, except
that the local oscillation is seen to disappear.
The analysis of relative frequency deviations in Fig. 15, on
the other hand, shows that the POMs form an excellent basis for
inter-swing dynamic characterization; it is observed that each
phase is clearly identified and that the time intervals of concern
exhibit nearly linear characteristics; this enables the damping
and frequency of the inter-swing dynamics to be easily and re-
liably computed. The analysis enables the precise identification
of dynamic trends, as well as the determination of time instants Fig. 15. Temporal evolution of the dominant POMs. Relative frequency devi-
at which structural or other operational changes impact system ations.
behavior.
Further insight into the temporal behavior of these modes
can be obtained from the analysis of instantaneous frequency in the bus frequencies show a dominant mode at about 0.25 Hz.
Fig. 16. It is apparent from this figure that for the interval subse- Further, the lower plot in Fig. 16 shows that the frequency of
quent to the interconnection of the systems, 06:27:42–06:28:06, the dominant POM increases to about 0.62 Hz. These results

Authorized licensed use limited to: CINVESTAV. Downloaded on April 13,2021 at 15:30:11 UTC from IEEE Xplore. Restrictions apply.
MESSINA AND VITTAL: EXTRACTION OF DYNAMIC PATTERNS 691

Fig. 16. Instantaneous frequency of dominant POMs.

Fig. 18. Temporal evolution of the basis functions for POM 1. (a) Temporal
evolution of POM 1. (b) Detail of the temporal evolution of POMs for selected
signals.

Fig. 17. Ensemble data from the EMD technique. Procedure 2.

correlate very well with EMD analysis in Fig. 12 showing the


accuracy and applicability of the procedure.

D. Simulation Results—Procedure 2
In this approach, a subset of individual signals is selected for
study. Each signal, , is decomposed into various dominant
IMFs as .
Then the ensemble matrix is constructed as
for
at each time instant, and the transposed measure- Fig. 19. Comparison of IMF1 and the reconstructed signal using POD. Bus
ment data matrix , is assembled as shown in Fig. 17. frequency signal at Hermosillo substation.
Since, in typical applications, the IMFs are relatively small
(typically, ) a significant reduction in computational time
results and the technique can be efficiently incorporated into oscillatory behavior, lasting from 0.6:28:21 to 06:28:33 shows
the basic EMD algorithms described in [3]. Fig. 18 shows plots a similar behavior to that observed during the first phase.
of the dominant mode versus time for the various signals of To verify the accuracy of the procedures and provide insight
interest. into the nature of the IMFs, Fig. 19 compares the time evolution
The analysis suggests the occurrence of three distinct of IMF1 to the reconstructed signal using the ROM,
phases of the observed dynamic oscillatory behavior related . For comparison, the time evolution of the re-
to different periods in the power system response. During constructed signal using Procedure 1 is also shown. The results
the first period following the interconnection of the system are practically identical.
(06:27:42–06:28:06), the Hermosillo and Mazatlan Dos signals Simulation results provide further insight into the character-
are seen to swing coherently in opposition to the Tres Estrellas istics of the EMD technique. The key important result from
bus frequency signal in Fig. 9(b). For the second interval, Fig. 19 is that IMF 1 can be interpreted as the leading POM; i.e,
06:28:06–06:28:32 this trend is reversed; the bus frequency the one that captures most of the energy in the signal. It should
signals at Hermosillo and Tres Estrellas substations oscillate be emphasized that whilst the EMD technique results in nine
approximately 180 out of phase with each other. IMFs and a residue (not shown), EOF analysis shows that IMF1
As pointed out above this behavior springs from control ac- preserves all the essential features of interest. This demonstrates
tions in the system including the loss of a generator in the north- the capability of the EOF analysis as a modal reduction tech-
western system and subsequent load shedding. The last phase of nique.

Authorized licensed use limited to: CINVESTAV. Downloaded on April 13,2021 at 15:30:11 UTC from IEEE Xplore. Restrictions apply.
692 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 2, MAY 2007

VI. CONCLUSION [5] R. W. Wies, A. Balasubramanian, and J. W. Pierre, “Combining least


mean adaptive filter and auto-regressive block processing techniques
for estimating the low-frequency electromechanical modes in power
In this paper, Hilbert analysis and the proper orthogonal anal- systems,” in Proc. IEEE Power Engineering Society General Meeting,
ysis techniques are used in a complementary manner to extract 2006.
dynamic information from measured data. Proper orthogonal [6] N. Zhou, J. W. Pierre, and J. F. Hauer, “Initial results in power system
identification from injected probing signals using a subspace method,”
analysis extends the applicability of the EMD technique by de- IEEE Trans. Power Systems, vol. 21, no. 3, pp. 1296–1302, Aug. 2006.
termining the energy and phase relationship embedded in the in- [7] A. R. Messina, V. Vittal, D. Ruiz-Vega, and G. Enriquez-Harper,
trinsic mode functions and the modes with the most energy and “Interpretation and visualization of wide-area PMU measurements
using Hilbert analysis,” IEEE Trans. Power Syst., vol. 21, no. 4, pp.
provides valuable information regarding the strength of non- 1763–1761, Nov. 2006.
linear effects in the system. [8] A. R. Messina and V. Vittal, “Nonlinear, non-stationary analysis of
The method is found to identify unambiguously the driving interarea oscillations via Hilbert spectral analysis,” IEEE Trans. Power
Syst., vol. 21, no. 3, pp. 1234–1241, Aug. 2006.
dynamics, and acts as a reduction technique to remove all [9] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.
nonessential information. Monitoring the time evolution of C. Yen, C. C. Tung, and H. H. Liu, “The empirical mode decomposition
the modes provides additional insight into the time-windows and the Hilbert spectrum for nonlinear and non-stationary time series
analysis,” Proc. Roy. Soc. Lond. A, vol. 454, pp. 903–995, 1998.
exhibiting abrupt changes. [10] P. Flandrin, G. Rilling, and P. Goncalves, “The empirical mode decom-
One of the primary applications of this technique is the sys- position as a filter bank,” IEEE Signal Process. Lett., vol. 11, no. 2, pp.
tematic determination of phase relationships between critical 112–114, Feb. 2004.
[11] H. Philip, L. L. John, and B. Gal, Turbulence, Coherent Structures, Dy-
modes and the identification of dynamic trends. In addition, the namical Systems and Symmetry. New York: Cambridge Univ. Press,
potential use of the method to determine optimal measurement 1996.
sites for the analysis of inter-area swing dynamics from mea- [12] G. Kerschen, J.-C. Galinval, A. F. Vakakis, and L. A. Bergman, “The
method of proper orthogonal decomposition for dynamical charac-
sured data is being investigated. Efforts to generalize the present terization and order reduction of mechanical systems: An overview,”
approach to the nonlinear case are being actively investigated Nonlin. Dynam., vol. 41, pp. 147–169, 2005.
and techniques to fill missing data are also being researched [13] B. F. Feeny and R. Kappagantu, “On the physical interpretation of
proper orthogonal modes in vibrations,” J. Sound Vibr., vol. 211, no.
using weightings in the POD formulation. 4, pp. 607–616, 1998.

A. R. Messina (M’85–SM’05) received the M.Sc. degree (Hons.) in electrical


engineering from the National Polytechnic Institute of Mexico in 1987 and the
REFERENCES Ph.D. degree from Imperial College, London, U. K., in 1991.
Since 1997, he has been a Professor at the Center for Research and Advanced
[1] J. F. Hauer and J. G. DeSteese, A Tutorial on Detection and Charac- Studies (Cinvestav), Guadalajara, Mexico. From 2005 to 2006, he was a Re-
terization of Special Behavior in Large Electric Power Systems Pacific search Associate at Arizona State University, Tempe.
Northwest Nat. Lab., Richland, WA, 2004, Rep. PNNL-14655.
[2] J. J. Sanchez-Gasca and J. Chow, “Performance comparison of three
identification methods for the analysis of electromechanical oscilla-
tions,” IEEE Trans. Power Syst., vol. 14, no. 3, pp. 995–1002, Aug. Vijay Vittal (M’82–SM’87–F’97) received the B.E. degree in electrical engi-
1999. neering from B.M.S. College of Engineering, Bangalore, India, in 1977, the
[3] D. J. Trudnowski, J. M. Johnson, and J. F. Hauer, “Making Prony anal- M.Tech. degree from the Indian Institute of Technology, Kanpur, India, in 1979,
ysis more accurate using multiple signals,” IEEE Trans. Power Syst., and the Ph.D. degree from Iowa State University, Ames, in 1982.
vol. 14, no. 1, pp. 226–231, Feb. 1999. He served on the faculty of the Department of Electrical and Computer En-
[4] J. W. Pierre, D. J. Trudnowski, and M. K. Donnelly, “Initial results gineering, Iowa State University, from 1982 to 2004. He is currently the Ira A.
in electromechanical mode identification from ambient data,” IEEE Fulton Chair of the Electrical Engineering Department, Arizona State Univer-
Trans. Power Syst., vol. 12, no. 3, pp. 1245–1251, Aug. 1997. sity, Tempe.

Authorized licensed use limited to: CINVESTAV. Downloaded on April 13,2021 at 15:30:11 UTC from IEEE Xplore. Restrictions apply.

You might also like