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

1, FEBRUARY 2011

323

Robust Detection and Analysis of Power System Oscillations Using the Teager-Kaiser Energy Operator
Innocent Kamwa, Fellow, IEEE, Ashok Kumar Pradhan, Senior Member, IEEE, and Geza Jos, Fellow, IEEE
AbstractCritical to real-time oscillations monitoring is early detection when otherwise dormant natural modes become a serious threat to grid stability. The next urgent issue is to determine the frequency and damping of the problematic modes when the signal is embedded in noise and the system contains closely spaced natural modes. The present paper addresses the detection issue using the Teager-Kaiser energy operator (TKEO) which has shown to be a fast predictor of the instability onset time when applied to the output signals of an orthogonal lter bank. In the system stability context, linear lter decomposition (LFD) is preferred rather than empirical mode decomposition (EMD), well known for its tendency to generate articial modes with no physical meaning. A narrowband LFD with a less than 0.2-Hz bandwidth is achieved in the range 0.05 to 3 Hz through a cosine-modulated lter bank design. The effectiveness of the scheme in accurately detecting and tracking the frequency and damping of oscillatory modes is demonstrated using Monte Carlo simulations of three closely spaced modes and a detailed analysis of an actual event recorded by Hydro-Qubecs WAMS in 2006. Index TermsFilter bank, interarea oscillation, power system identication, power system monitoring, power system oscillations, power system stability, Teager-Kaiser energy (TKE), wide-area measurement systems (WAMS).

I. INTRODUCTION

SCILLATION monitoring [1][5] is a topic of increasing recent attention, owing to a heightened interest in the prevention of widespread blackouts, which can result from uncontrolled power swings across large geographical areas. Basically, the entity responsible for this task should rst determine from power system response signals when/if there is an oscillation issue and then quantify the underlying threat by means of an accurate assessment of the frequency and damping of the oscillation. Since the early days of modal analysis of power system responses recorded by WAMS [6] or simulated using power system studies software [7], the two issues of detecting and quantifying oscillations have been generally mixed together
Manuscript received October 18, 2009; revised December 30, 2009. First published April 26, 2010; current version published January 21, 2011. Paper no. TPWRS-00823-2009. I. Kamwa is with Hydro-Qubec/IREQ, Power System and Mathematics, Varennes, QC J3X 1S1, Canada (e-mail:kamwa.innocent@ireq.ca). A. K. Pradhan is with the Department of Electrical Engineering, Indian Institute of Technology, Kharagpur 721302, India (e-mail: akpradhan@ee.iitkgp. ernet.in) G. Jos is with the Department of Electrical and Computer Engineering, McGill University, Montreal, QC H3A 2A7, Canada (e-mail: geza.joos@mcgill.ca). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TPWRS.2010.2046503

with few exceptions [8][10], but detailed analysis obviously makes little sense when there is no signicant oscillation activity in the electromechanical frequency range (typically 0.13 Hz). In this paper, a Teager-Kaiser energy operator (TKEO) [11][14] based criterion is proposed as a predictor of power oscillation problems. To the best of our knowledge, this is the rst time that the TKEO concept has been used in this context. Since the concept is known to perform poorly on multi-component signals [13], a multi-band pre-lter is rst applied to the raw input. This allows us to decompose the waveform into a set of largely orthogonal monochromatic components which are then subjected to time-frequency analysis using the energy separation algorithm (ESA). One of the benets of the linear lter bank is that, when the prototype is designed with a very narrow bandwidth, e.g., 0.2 Hz, any analysis method applied to the output will provide noise-resilient frequency and damping estimates. Recent authors have investigated closely related ideas using the EMD as lter bank [15] and discrete Hilbert transform (DHT) as the basis for instantaneous frequency and amplitude estimation [16], [17]. However, in spite of state-of-art improvements, even the best implementation of EMD tends to generate articial modes due to the frequency-mixing phenomenon [18][20]. This is especially true at low sampling rates, which are typical of todays WAMS, whose communication bandwidths are constrained [21]. Although the ESA and Hilbert transform are the shortest paths to amplitude and frequency information [14], [17], they do not provide damping information without additional processing. Furthermore, they are still prone to errors when the signal is not perfectly monochromatic, which is hard to achieve for systems with closely spaced natural modes. Application of parametric methods to the lter bank outputs will, in this case, result in a better frequency resolution while providing damping information [22]. For this reason, the eigensystem realization algorithm (so-called ERA method) proposed in [23] will be adopted in this research. In addition to its intrinsically multi-signal capability, it performs very well for ringdown and stationary ambient noise signals when the SNR ranges from moderate to high. The main innovation of the present paper resides in the application of TKEO to the output signals of a linear lter bank for robust location in time and frequency of the onset of power system oscillations. Locating the oscillation frequency can be useful for control and identication methods that require a good initial value of the natural mode within a certain a priori frequency range. It is also a good catastrophe-tracking predictor for

0885-8950/$26.00 2010 IEEE

324

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 26, NO. 1, FEBRUARY 2011

the control room, especially when supplemented with an accurate assessment of the underlying modal frequency and damping using a high-resolution parametric method like ERA. The proposed multi-band lter consists of narrow band-pass cells with a center frequency uniformly distributed from 0.1 to 3 Hz. Designed using cosine modulation [24] of Taylor spectral windows [25], the linear bandpass prototype FIR-lter possesses attractive attributes like steep stop band (60 dB), fast-decaying side-lobes, and DC rejection. The DC attenuation can be further improved using a 100-sample FIR DHT [1]. The paper is organized as follows. In section two, the essential ideas of the proposed online modal analysis are summarized. Then in Section III, the Teager-Kaisers energy operator is reviewed. Its tracking and selectiveness capabilities are demonstrated on synthetic multi-sine waveforms signals including a 0.025 Hz/s chirp signal. Section IV is devoted to the design and assessment of a narrow-band linear lter bank for facilitating the application of the TKEO to closely spaced natural modes. Next, the ERA-based multi-band modal analysis (MBMA) is developed in Section V for providing full modal amplitude, frequency, and damping information in tracking mode. Monte Carlo simulations of a parametric signal with closely spaced modes [26] demonstrate the higher reliability of the LFD TKEO based modal-analysis approach, compared to a conventional parametric method. Finally in Section VII, the analysis of an actual unstable event recorded by the Hydro-Qubec WAMS [27] in 2006 conrms that the MBMA scheme can detect real-life oscillation problems, quantify their severity, and report the results every 5- to 40-s data blocks, depending on the frequency band with signicant oscillation activity, which is much faster that the 120 s required by typical sliding Fourier or autocorrelation based block-processing methods [4], [28]. II. OUTLINE OF THE PROPOSED MODAL ANALYSIS METHOD The power system response signals can generally be described as the sum of the combined amplitude (AM) and frequency (FM) modulated primitive signals:

Fig. 1. Application of ESA on multi-component AM-FM signals.

(1) , and are, respectively, the amplitude, frewhere quency, damping, and phase of the th signal component. The corresponding complex signal is of the form

(2)

Fig. 1 illustrates the overall concept of rst decomposing the multi-component power system signal (1) using a lter bank, and then performing a time-frequency analysis on each channel component using ESA or, alternatively, the Hilbert transform [16]. If the EMD is used for pre-processing, it yields intrinsic mode functions (IMF) [18], which are essentially single-component sinusoids if and only if the natural modes are well spaced on the frequency axis and the sampling rate is sufciently high. However, most of the time, such is not the case and the EMD will create IMFs that contain articial modes due to the mixing phenomenon [18]. To avoid the EMD pitfalls for closely spaced modes, the Gabor lter bank could be used, as in [13] and [29], but it is not sharp enough for the present application. We will design in Section IV a narrow-band linear lter bank which is far more suited for power system oscillations signals at the relatively low sampling rates of todays WAMS [21]. When accurate damping information is required or the lter bank output signal is not monochromatic, a more detailed modal analysis, using a parametric method for instance, should be preferred to the ESA. The overall scheme is shown in Fig. 2, assuming that ERA [7], [23] is the modal analysis tool, even though any alternative modal estimation method could equally well be used [4], [22], [30]. An optional DHT [1] is rst applied to the incoming signal to reject quasi-steady state components more efciently. The lter bank then splits the possibly multi-component signal as in (1) into N essentially in the present paper). orthogonal components (with After a linear lter bank decomposition of the incoming signal, each channels energy is computed using TKEO and the (typically ) energy-dominant signals are selected for parametric modal analysis based on an energy threshold test. The modal analysis is then performed on sliding non-overlapping data blocks whose size can be adjusted per channel, according to the channel center frequency. For instance, a 5-s data window sufces for lters centered in the range 1 to 3 Hz while a 20-s data window could be preferable for analyzing output signals of lters in the range 0.1 to 1 Hz. III. OVERVIEW OF THE TEAGER-KAISERS ENERGY OPERATOR Teager [11] developed a nonlinear signal operator which was subsequently shown by Kaiser [12] to be highly effective for

where is the Hilbert transform of . It should be , and can be slowly stressed that all parameters time-varying without posing any signicant complication to the proposed algorithms.

KAMWA et al.: ROBUST DETECTION AND ANALYSIS OF POWER SYSTEM OSCILLATIONS

325

Fig. 3. Illustration of the TKEO and ESA concepts on simple signals.

Fig. 2. Overview of the MBMA scheme with a TKEO-based criterion to detect channels with the strongest oscillatory activity.

with . By applying backward difference, time ESA was developed in [13], [14]

to both

and its , the discrete-

assessing the energy and detecting AM and FM signals in the form (1)

(3) and the rst and second derivatives with, is the instantaneous energy of . In fact, deof , while the th component of the AM-FM signal in (1) noting by and assuming a constant initial phase , it is proven that [12], [13] (8)

(4) with a negligible approximation error under quite general and realistic conditions. This motivated the energy separation algorithm (ESA) for tracking the instantaneous amplitude and frequency as follows [13], [14]:

The formulation shows that ESA can provide the instantaneous amplitude and frequency with a single sample delay in sharp contrast with the DHT [14], [16], [17] and other blockprocessing schemes. However, involvement of the signal derivative in (8) also points to a potential sensitivity to noise. Fig. 3 illustrates the nearly perfect performance of the ESA on a unit magnitude, monochromatic damped, and un-damped sinusoids of the same 1-Hz frequency:

(9) , for a pure sine; , for a constant where , for a sine with a linearly damping sine and varying damping. The discrete energy in Fig. 3 is divided by T to match the same energy scale of the continuous operator. It is clear from (4)(8) that the ESA is dened specically for a single-component signal. Therefore, its proper application requires a pre-processor to decompose any incoming multi-component signal into monochromatic components using a lter bank, as illustrated in Fig. 1 [29].

(5) Note that, in the case of a monochromatic signal with constant and , the relationships (5) are exact. When the continuous energy operator (3) is sampled with a sampling period T, the following equivalent discrete energy operator is obtained [13]:

(6) It results in IV. LINEAR MULTI-BAND SIGNAL DECOMPOSITION In order to design a realistic narrow-band lter bank better suited than EMD for decomposing power systems signals in

(7)

326

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 26, NO. 1, FEBRUARY 2011

the form (1) with closely spaced modes, let us consider the following denition, which comes from the cosine-modulated lter bank theory [24]:

(10) and is the imwhere pulse response coefcients of a linear-phase low-pass FIR prototype lter. Furthermore, the number of lter channels is seis the length of the prototype lter. If properly lected so that chosen, the scaling factor and the center frequency of the th lter can be located at a pre-specied frequency. The nal impulse response data of the th band-pass lter, after a re-scaling for a unit magnitude and zero phase at the center frequency, is
Fig. 4. Example of cosine-modulated lter bank designed from Taylor window-based low-pass prototypes. (a) Taylor window based lter bank. (b) Filter bank with DHT pre-lter.

(11) is the response of the th lter cell at where its center frequency. The lter bank gain is shown in Fig. 4(a). The corresponding low-pass prototype is based on the Taylor window as dened in [25]

(12) with the lter coefcients being

Given a 40-Hz sampling rate, the prototype window has a 400-sample length for lters 1 to 4 and 200 samples for lters 5 to 9. These numbers were selected as a trade-off between lter delay and narrow-band behavior. It should be noted that the component lters are essentially orthogonal as they overlap -dB attenuation frequency. For instance, the -dB at their crossing frequencies of the third lter are 0.4 Hz and 0.6 Hz on the low and high sides, respectively. To improve the DC rejection of the lter bank, especially that of the rst lter, a DHT pre-lter can be applied to the signal, as in [1]. The effect of a 100-sample FIR-based DHT designed in Matlab is shown in Fig. 4(b), where it is seen that the attenuation at 0.01 Hz has increased from 20 dB to 40 dB. Fig. 5 shows side-by-side the performance of EMD and LFD on the following signal: (13)

Fig. 5. Comparison of EMD and LFD on a benchmark signal consisting of two closely-spaced modes (13). (a) EMD. (b) Filter bank.

studied in [16] to demonstrate and mitigate the EMD frequency mixing effects. Our EMD results were obtained using state-of-art software [19], but the masking technique of [16] was not implemented because we were looking for decomposition methods with no a priori knowledge about the signal. For both EMD and LFD, only the three dominant components are shown with a 40-Hz sampling rate. The periodogram of the LFD signals correctly predicts two components at 0.5 and 0.8 Hz while IMF #6 contains four modes, including two modes not found in the original signal. In addition, it is not monochromatic, which will diminish the ESA and DHT performances. IMF #4 is monochromatic but its frequency is articial. All these artifacts are avoided by the LFD, thanks to its linear and narrow-band behavior.

KAMWA et al.: ROBUST DETECTION AND ANALYSIS OF POWER SYSTEM OSCILLATIONS

327

To further demonstrate the tracking capability of the lter bank, let us consider the following unit magnitude chirp signal: (14) with

(15)

The input signal and results of its multi-band analysis are shown in Fig. 6(a) and (b), respectively. Each channel signal is represented by a different color. It appears that, during the frequency sweep, the active lter channel changes according to the instantaneous chirp frequency at the given time and this change is reected in the energy operator. However, during a channel transition, there is a frequency change from 0 to the sweep signal value. The criterion for detecting the coincidence of the sweep and channel frequencies is based on the TKEO at that time, with the detection threshold dened as follows: (16)

V. TKEO-BASED MULTI-BAND MODAL ANALYSIS A. ERA-Based Regular Modal Analysis After a linear lter bank decomposition of the incoming signal, each channels energy is computed as in (6) and the (typically ) energy-dominant signals are selected for parametric modal analysis based on an energy threshold test such as (16). The ERA based modal analysis [7] is then performed on data blocks whose size can be adjusted per channel, according to the channel center frequency. Typically, a data window of two to four cycles of the center frequency sufces for analyzing the corresponding lter output signal. The ERA method at the core of our modal analysis approach basically assumed that the digital signal samples are the impulse responses of a hypothetical proper and linear system, whose discrete state-space representation (F, G, C) is determined through minimal realization of the impulse responses as in [23]. An implementation of ERA method adapted to single-input multiple-output (SIMO) signals is summarized in the Appendix. Once the continuous time-domain state matrices (A,B,C) are derived from (F,G,C), using for instance, the inverse Tustin transform, the Prony decompositions of the output signals are obtained as follows. 1) Transform (A,B,C) into its diagonal form assuming for presentation simplicity, distinct complex eigenvalues: (17) with the diagonal matrix of the , eigenvalues of A; and

Fig. 6. Multi-band analysis of the chirp signal in (14): each color corresponds to the signal output of a corresponding band-pass lter. Amplitude and frequency are computed with ESA. (a) 0.025 Hz/s chirp signal. (b) Multi-band analysis.

, the corresponding left-eigenvector matrix. , 2) Compute the vectors of residues, for each natural mode (18) where designates the th output signal. , the Prony parameters are dened 3) Given the residues , according to the for each output signal following sum of damped sinusoids expression:

(19)

B. Demonstration on Multi-Modes Synthetic Signal The multi-band modal analysis (MBMA) combining TKEObased detection with ERA-based parameter estimation will be tested rst on a noise-free multi-component signal with several closely spaced natural modes. The signal is the result of the Prony analysis of a measured power system response initially reported in [6] and later used in [7] as a typical case for assessing

328

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 26, NO. 1, FEBRUARY 2011

Fig. 7. Standard ERA-based modal analysis and validation results of the synthetic signal in (20).

Fig. 8. Multi-band decomposition of the synthetic signal in (20).

TABLE I MALIN-ROUND MOUNTAIN MODIFIED NATURAL RESPONSE LINE CURRENT TO HVDC MODULATION [6], [7]

modal analysis methods. The parameters shown in Table I contribute to the signal as follows:

Fig. 9. Frequency responses of the ERA models associated with the synthetic signal in (20). The INPUT plot is the result without LFD.

(20) with obvious relationships between the modal parameters in . (19) and (20), giving a single output signal The results of a conventional ERA analysis of this signal (20) are presented in Fig. 7. The sampling frequency is 40 Hz, with an assumed system . After Prony decomposition of the state-space order model as in (19), a ltering is performed to reject as non-relevant all modes with a damping higher than 30% and amplitude lower than 0.1% of the maximum modal amplitude. Under these assumptions, conventional ERA-based modal analysis is quite accurate in spite of the many closely spaced modes present in the signal (20). The time and energy signals from the four dominant channels in the lter bank are shown in Fig. 8. From time-domain plots,

the red and black signals are not perfectly monochromatic despite the narrow bandwidth of the lters. The energy information points to the fact that a large magnitude in the time domain does not necessarily translate into large energy, which is a product of frequency and magnitude. ERA-based modal analysis of the four dominant lter output signals was performed under conditions similar to those in Fig. 7, but assuming a reduced model order of 6. The four frequency responses obtained with these models are shown in Fig. 9 along with the overall frequency response that resulted from the conventional full-band ERA application (Fig. 7). Again, the red and black signals highlight two peaks, meaning that ESA and DHT cannot accurately extract their amplitude and frequency. In addition, the full-band ERA yields ve magnitude peaks, the same number as the MBMA with four dominant channels but the peaks from the latter are sharper, leading to more accurate frequency location. Finally, Fig. 10 portrays the damping and frequency estimation errors incurred by the multi-band modal analysis. The fre-

KAMWA et al.: ROBUST DETECTION AND ANALYSIS OF POWER SYSTEM OSCILLATIONS

329

Fig. 10. MBMA frequency and damping errors on the synthetic signal in (20). Fig. 11. Comparative reliability of the conventional and multi-band modal analysis of the synthetic signal in (21). There are 4000 trials for each SNR value.

quency bias is minor but the error on damping is more signicant when the lter channel used to extract the underlying frequency is not monochromatic (i.e., 0.65 Hz and 0.82 Hz). This aspect could be further improved by doubling the number of lters in the bank, as this would greatly improve the narrow-band behavior of the lter channels. C. Monte Carlo Study of MBMA Performance One claimed attribute of the proposed scheme is its noise-resiliency, which will now be assessed by the usual means of Monte Carlo studies. The synthetic signal chosen for this purpose is from [26] and is described as follows:

Considering only the cases for which the modal analysis was successful (i.e., exactly three modes were obtained after ltering non-relevant modes), Fig. 12 depicts the dispersion of the estimated parameters around their mean values, giving a 20-dB SNR. The three plots on the left (12-a-c) present the frequency against damping for the three modes, while the three plots on the right express (12-d-f) the frequency against the magnitude. It should be noted that the amplitude estimate from the MBMA is compensated according to the frequency response gain of the corresponding channel lter as follows:

(21) (23) where where is the frequency response magniis the tude of the th lter at the natural frequency , and ERA amplitude estimate based on the th signal. All plots in Fig. 12 show a more compact error distribution in the MBMA case compared with the conventional ERA method, especially for the two lowest frequency modes. In addition, given the higher reliability index of the former, Fig. 12 contains 2512 over 4000 runs which have converged for the multi-band scenario in contrast to only 1714 convergent points in the standard ERA scenario. VI. ANALYSIS OF AN ACTUAL EVENT The eld record considered here describes the dynamic process preceding and following the trip of a single 350-MW generating unit at a 16-unit power plant in the Hydro-Qubec system. The event was recorded by the WAMS [27] for about 1-min duration on March 11, 2006, starting at 16 h 49 min 20 s local time. The signals shown in Fig. 13 are very remote from the unit in trouble, which explains why the magnitude of the unstable voltage and angle swings is not signicant. Selecting the angle shift signal as target for the MBMA, Fig. 14 shows the subset of lter-bank output signals to be used in assessing the event. These are the dominant-energy signals according to the TKEO criterion. The energy of all other angle-driven for the entire observation channel signals is less than period. By contrast, the energy of the lters outputs #5 and #6 starts to rise immediately at the onset of the oscillation. The

(22)

The above signal was sampled at 40 Hz and white noise was added so as to achieve a specied signal-to-noise ratio (SNR) from 20 dB to 100 dB. At each SNR level, 4000 trials were generated. Modal analysis of each noisy record was then performed using the conventional ERA and the new scheme combining multi-band preprocessing, TKEO-based oscillation detection, and ERA. The selected post-ltering criteria and aswere the same for both approaches. sumed model order The rst performance metric is the ratio of the number of cases for which the algorithm obtained exactly three modes (after ltering) over the number of trials (4000). When the algorithm is unable to t the actual number of modes (3), it is assumed to have failed in that case, even if this is not always true, but since the same ERAs implementation is applied with and without the multi-band pre-processing, this assumption allows for a fair comparison. According to Fig. 11, both methods are sensitive to noise but the MBMA is more robust. At a 60-dB SNR, it is perfectly reliable while the conventional ERA shows a 75% reliability level only. Even at a 100-dB SNR, the latter is still unable to reach a 100% reliability target.

330

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 26, NO. 1, FEBRUARY 2011

Fig. 14. Multi-band analysis of the angle shift shown in Fig. 13. Amplitude and frequency computed with ESA.

Fig. 12. Conventional versus multi-band analysis results of the synthetic signal in (21) at a 20-dB SNR.

Fig. 15. Tracking of the dominant mode based on a 6-s, frame-by-frame analysis of channel # 5 and 6 signals.

and the damping of the plant local mode was restored immediately to more acceptable values. Throughout this event, the frequency of the mode never changed (ERA and ESA concluded the same), except that the damping moved from negative to positive.
Fig. 13. PMU signals on March 11, 2006: single unit trip at a remote power plant. Response variables from the Churchill Falls PMU, with angle-shift referenced near Montreal [27].

VII. CONCLUSION In this paper, the TKEO is introduced for the rst time as a tool for early detection of power system oscillations. In contrast to other parametric and high-pass ltering-based detection methods, TKEO makes no assumption about the signal. It is intrinsically a fast non-parametric detection method with a smooth behavior when applied to the output signals of a linear narrow-band lter bank. It is found that, although EMD is not appropriate for this task, a bank of cosine-modulated orthogonal lters appears to be very effective in decomposing the incoming multi-component power system response in terms of its monochromatic components while rejecting both random noise and DC components. The ESA is used to obtain an initial estimate of the frequency and amplitude of the ltered signal once an oscillation problem is detected. This can be sufcient in many circumstances, but it is shown that the ERA provides noise-resilient damping of the modes over short time frames. Several multiple-component synthetic signals available from the literature are analyzed in both deterministic and Monte Carlo

ESA is then used for quick visual screening of the amplitude and frequency of the oscillation. The channel #6 signal was fed to the MBMA tool. A separate analysis was then performed on a contiguous, non-overlapping 6-s buffer of data, for a total of ten snapshots over the approximately 60-s period of the record. Fig. 15 shows the frequency and damping obtained in tracking mode using the ERA. The fact that something is going wrong on the system is clearly identied in the rst snapshot by the negative damping associated with the dominant mode of the channel analyzed. The local mode associated with the unit in trouble has a frequency of about 1.25 Hz (same as the value obtained with ESA) and its damping was %. However, all PMU signals are needed in order to pinpoint more precisely the area in which this unit is located. After growing quickly at rst, the oscillation amplitude stabilizes between 10 and 40 s. Around this time, the unit in trouble tripped

KAMWA et al.: ROBUST DETECTION AND ANALYSIS OF POWER SYSTEM OSCILLATIONS

331

setups to conrm the superiority of multi-band modal analysis over the conventional approach using the same ERA implementation. Analysis of an actual power system event also highlights the efcacy of the new detection and MBMA framework in real-life situations. Overall, the results presented in the paper allow us to believe that the proposed approach could be useful in analyzing ambient noise, as well as ringdown signals recorded by WAMS or simulated using power system stability analysis software. VIII. APPENDIX: THE EIGENSYSTEM REALIZATION ALGORITHM The purpose of ERA is to determine numerically the matrices (A, B, C) or (F, G, C) in the following equivalent continuous and discrete state-space representations of a single input, -output linear system:
Fig. 16. Description of the impulse response based ERA.

(24) Their sampled impulse response is given by a sequence of matrices such as:

distribution parameters and of the matrices and , which allow us to change the time span of the window to be used in the identication procedure. C. Construction of the Hankel Matrices and

(25) In these expressions, and are, respectively, the state, and 1, while output, and input vectors of dimensions are the sampling instants. For MBMA purposes, we can consider that represents a block-data output over a time window of durafrom the lter bank channel . The modes found in the signal are exactly the tion state matrix A. eigenvalues of the A. Top View of ERA Fig. 16 recaps the ERA for state-space system identication, which similarly to the N4SID method [6], is based on the construction of two Hankel matrices, and . B. Identication Parameters sampling period T; p measured signals, concatenated in ; dimension n of the state representation (A,B,C); and , which in dimension of the Hankel matrices our case are square;

The elements of the Hankel matrices are measurements that have been organized in terms of the selected identication parameters. For a given set of parameters, construction of the two matrices is almost identical, the only difference being the is offset by one sample compared to . For position of is , example, if the rst sample used for calculating then the rst sample used for will be . For and with the notations given, the Hankel matrices and are obtained as shown in (26) at the bottom of the page, where the sufx is dropped from in order to t the Hankel matrix more conveniently in a single column. D. Remarks matrices depend on the parameters T, , and The . If , then and are symmetrical. , adjacent data are used. If and must be chosen so that The parameters , where represents the number of samples. For a value of , the parameters and allow the width equal to to be of the observation window , and , increased. So, if

. . .

. . .

. . .

. . .

(26)

332

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 26, NO. 1, FEBRUARY 2011

this means that the identication has been done using the , rst 200 samples. If, on the other hand, , this means that one sample out of two and is used by the identication procedure and that the last sample used is the 400th. In both cases, the identication procedure uses 200 samples, the difference being that in the rst case, it is the rst 200 samples, whereas in the second, it considers one sample out of two up to the 400th. E. State Representation Computation The discrete state, input and output matrices F, G, and C, respectively, of the discrete system are computed using matrices and . Decomposition of into singular values yields the ardiagonal matrix S, which contains the singular values of ranged in descending order, and the matrices U and V containing the left and right singular vectors, respectively. The identication parameter n is therefore used to select the n highest singular containing the n highest values. Thus, we obtain the matrix singular values and the matrices of the associated singular vecand from which the matrices (F,G,C) of the state tors representation of the discrete system are derived as follows:

(27) Once the matrices F, G, and C have been computed, the continuous system (A, B, C) is obtained using the discrete to continuous transformation (i.e., the inverse Tustin transform). ACKNOWLEDGMENT The authors would like to thank F. Levesque for fruitful discussions on tracking modal analysis and to G. Blais and J. Bland for retrieving and analyzing eld data from the HydroQubec WAMS. REFERENCES
[1] S. Zhang, X. Xie, and J. Wu, WAMS-based detection and early-warning of low-frequency oscillations in large-scale power systems, Elect. Power Syst. Res., vol. 78, pp. 897906, 2008. [2] P. Korba, Real-time monitoring of electromechanical oscillations in power systems: First ndings, IET Gen., Transm., Distrib., vol. 1, no. 1, pp. 8088, Jan. 2007. [3] D. H. Wilson, K. Hay, R. F. B. Maclaren, D. J. Hawkins, A. Dunn, A. J. Middleton, A. Carter, and W. Hung, Control centre applications of integrated WAMS-based dynamics monitoring and energy management systems, . Paris, France, Aug. 2008, CIGRE Session. [4] D. J. Trudnowski, J. W. Pierre, N. Zhou, J. F. Hauer, and M. Parashar, Performance of three mode-meter block-processing algorithms for automated dynamic stability assessment, IEEE Trans Power Syst., vol. 23, no. 2, pp. 68690, May 2008. [5] I. Kamwa, J. Bland, and D. McNabb, PMU-based vulnerability assessment using wide-area severity indices and tracking modal analysis, in Proc. IEEE/PES Power Systems Conf. Expo. (PSCE06), Atlanta, GA, Oct. 29Nov. 1 2006, pp. 139149. [6] J. F. Hauer, Application of Prony analysis to the determination of modal content and equivalent models for measured power system response, IEEE Trans. Power Syst., vol. 6, no. 3, pp. 10621068, Aug. 1991. [7] I. Kamwa, R. Grondin, J. Dickinson, and S. Fortin, A minimal realization approach to reduced-order modeling and modal analysis for power system response signals, IEEE Trans. Power Systems, vol. 8, no. 3, pp. 10201029, Aug. 1993.

[8] J. F. Hauer and F. Vakili, An oscillation detector used in the BPA power system disturbance monitor, IEEE Trans Power Syst., vol. 5, no. 1, pp. 7479, Feb. 1990. [9] R. Andrew, G. Ledwich, and P. OShea, A Kalman ltering approach to rapidly detecting modal changes in power systems, IEEE Trans. Power Systems, vol. 22, no. 4, pp. 16981706, Nov. 2007. [10] N. Zhou, D. J. Trudnowski, J. W. Pierre, and W. A. Mittelstadt, Electromechanical mode online estimation using regularized robust RLS methods, IEEE Trans. Power Syst., vol. 23, no. 4, pp. 16701680, Nov. 2008. [11] H. M. Teager and S. M. Teager, Evidence for nonlinear sound production mechanisms in the vocal tract, in Speech Production and Modelling, ser. NATO Advanced Study Institute Series D, W. J. Hardcastle and A. Marchal, Eds. Boston, MA: Kluwer, Jul. 1729, 1989, vol. 55, Bonas, France. [12] J. F. Kaiser, Some useful properties of Teagers energy operator, in Proc. ICASSP-93, 1993, vol. 3, pp. 149152. [13] P. Maragos, J. F. Kaiser, and T. F. Quatieri, Energy separation in signal modulations with application to speech analysis, IEEE Trans. Signal Process., vol. 41, no. 10, pp. 30243051, Oct. 1993. [14] A. Potamianos and P. Maragos, A comparison of the energy operator and the Hilbert transform approach to signal and speech demodulation, Signal Process., vol. 37, pp. 95120, 1994. [15] P. Flandrin, G. Rilling, and P. Goncalves, Empirical mode decomposition as a lter bank, IEEE Signal Process. Lett., vol. 11, no. 2, pp. 112114, Feb. 2004. [16] D. S. Laila, A. R. Messina, and B. C. Pal, A rened Hilbert-Huang transform with applications to interarea oscillation monitoring, IEEE Trans. Power Syst., vol. 24, no. 2, pp. 610620, May 2009. [17] T. J. Browne, V. Vittal, G. T. Heydt, and A. R. Messina, A comparative assessment of two techniques for modal identication from power system measurements, IEEE Trans. Power Syst., vol. 23, no. 3, pp. 14081415, Aug. 2008. [18] R. T. Rato, M. D. Ortigueira, and A. G. Batista, On the HHT, its problems, and some solutions, Mech. Syst. Signal Process., vol. 22, pp. 13741394, 2008. [19] Y. Kopsinis and S. McLaughlin, Improved EMD using doubly-iterative shifting and high order spline interpolation, EURASIP J. Appl. Signal Process. vol. 2008, pp. 18. [Online]. Available: http://www. hindawi.com/journals/asp/2008/128293.abs.html. [20] EMD Software. [Online]. Available: http://perso.ens-lyon.fr/patrick. andrin/emd.html and http://www.see.ed.ac.uk/~ykopsini/emd/emd. html. [21] K. E. Martin et al., Exploring the IEEE Standard C37.118-2005: Synchrophasors for power systems, IEEE Trans. Power Del., vol. 23, no. 4, pp. 18051811, Oct. 2008. [22] P. OShea, A high-resolution spectral analysis algorithm for powersystem disturbance monitoring, IEEE Trans. Power Syst., vol. 17, no. 3, pp. 676680, Aug. 2002. [23] J.-N. Juan and R. S. Pappa, An eigensystem realization algorithm (ERA) for modal parameter identication and model reduction, J. Guid., Control, Dynam., vol. 8, no. 5, pp. 620627, 1985. [24] A. Viholainen, J. Alhava, and M. Renfors, Efcient implementation of complex modulated lter banks using cosine and sine modulated lter banks, EURASIP J. Appl. Signal Process. vol. 2006, pp. 110. [Online]. Available: http://www.hindawi.com/journals/asp/2006/058564.abs.html. [25] D. C. Rife and G. A. Vincent, Use of the discrete Fourier transform in the measurement of frequencies and levels of tones, Bell Syst. Tech. J., vol. 49, no. 2, pp. 197228, Feb. 1970. [26] M. Crow, Computational Methods for Electric Power Systems. Boca Raton, FL: CRC, 2003, sec. 7.4, pp. 218231. [27] I. Kamwa, J. Bland, G. Trudel, R. Grondin, C. Lafond, and D. McNabb, Wide-area monitoring and control at Hydro-Qubec: Past, present and future, in Proc. 2006 IEEE/PES General Meeting, Montreal, QC, Canada, Jun. 1822, 2006. [28] M. Glickman, P. OShea, and G. Ledwich, Estimation of modal damping in power networks, IEEE Trans. Power Syst., vol. 22, no. 3, pp. 13401350, Aug. 2007. [29] A. C. Bovik, P. Maragos, and Th. F. Quatieri, AM-FM energy detection and separation in noise using multiband energy operators, IEEE Trans. Signal Process., vol. 41, no. 12, pp. 32453265, Dec. 1993. [30] P. M. T. Broersen, Automatic spectral analysis with time series models, IEEE Trans. Instrum. Meas. vol. 51, no. 2, pp. 211216, Apr. 2002. [Online]. Available: http://www.mathworks.com/matlabcentral/leexchange/1330.

KAMWA et al.: ROBUST DETECTION AND ANALYSIS OF POWER SYSTEM OSCILLATIONS

333

Innocent Kamwa (S83M88SM98F05) received the Ph.D. degree in electrical engineering from Laval University, Laval, QC, Canada, in 1988. Since then, he has been with the Hydro-Qubec Research Institute (IREQ), Power System Analysis, Operation, and Control, Varennes, QC, where he is currently a Principal Researcher in bulk system dynamic performance. He has been an Associate Professor of electrical engineering at Laval University since 1990. Dr. Kamwa has been active for the last 13 years on the IEEE Electric Machinery Committee, where he is presently the Standards Coordinator. A member of CIGR and a registered professional engineer, he is a recipient of the 1998, 2003, and 2009 IEEE Power Engineering Society Prize Paper Awards and is currently serving on the Adcom of the IEEE System Dynamic Performance Committee.

Geza Jos (M82SM89F06) received the M.Eng. and Ph.D. degrees from McGill University, Montreal, QC, Canada. He has been a Professor with McGill University since 2001 and holds a Canada Research Chair in Power Electronics applied to Power Systems. He is involved in fundamental and applied research related to the application of high-power electronics to power conversion, including distributed generation and wind energy, and to power systems. He was previously with ABB, the Ecole de technologie suprieure, and Concordia University. He is involved on a regular basis in consulting activities in power electronics and power systems. Dr. Jos is active in a number of IEEE Industry Applications Society committees, IEEE Power Engineering Society working groups, and CIGRE working groups. He is a Fellow of the Canadian Academy of Engineering and of the Engineering Institute of Canada.

Ashok Kumar Pradhan (M94SM09) received the Ph.D. degree in electrical engineering from Sambalpur University, Burla, India, in 2001. He has been with the Department of Electrical Engineering, Indian Institute of Technology, Kharagpur, India, since 2002. Presently, he is an Associate Professor. His research interests include power system dynamics and Relaying.