Abstract
The subsynchronous oscillations (SSOs) related to renewable generation seriously affect the stability and safety of the power systems. To realize the dynamic monitoring of SSOs by utilizing the high computational efficiency and noise-resilient features of the matrix pencil method (MPM), this paper proposes an improved MPM-based parameter identification with synchrophasors. The MPM is enhanced by the angular frequency fitting equations based on the characteristic polynomial coefficients of the matrix pencil to ensure the accuracy of the identified parameters, since the existing eigenvalue solution of the MPM ignores the angular frequency conjugation constraints of the two fundamental modes and two oscillation modes. Then, the identification and recovery of bad data are proposed by utilizing the difference in temporal continuity of the synchrophasors before and after noise reduction. The proposed parameter identification is verified with synthetic, simulated, and actual measured phase measurement unit (PMU) data. Compared with the existing MPM, the improved MPM achieves better accuracy for parameter identification of each component in SSOs, better real-time performance, and significantly reduces the effect of bad data.
WITH the development of renewable generations in modern power systems, subsynchronous oscillations (SSOs) caused by the resonances between power electronic devices and series compensators or weak grids are of frequent occurrence [
There has been much research on identifying SSOs using instantaneous data, which can accurately obtain SSO parameters [
SSO parameter identification based on synchrophasors provided by phasor measurement units (PMUs) and wide area measurement systems (WAMSs) can realize dynamic and synchronized monitoring of SSOs by using the synchronous measurement mechanism and high reporting rate of synchrophasors [
In addition, non-DFT methods based on synchrophasors have also been studied. Two classic modal parameter extraction (MPE) algorithms, i.e., Prony analysis [
As mentioned before, the supersynchronous component may exist during the SSO. Due to the coupling frequency of the subsynchronous and supersynchronous components, the positive and negative spectra of which are aliasing, the parameters obtained by the above methods are incorrect because the supersynchronous component is not considered. In the primary research of our team [
MPM is a classic MPE algorithm with advantages in computational efficiency and noise-resilient features [
To take advantage of the computational efficiency and noise-resilient features of MPM, the MPM is improved by solving issues in the matrix pencil solution and identification and recovery of bad data in this paper. The features of the proposed improved SSO parameter identification based on MPM are twofold.
1) By analyzing the principle of the MPM and the disadvantage of the matrix pencil solution, the angular frequency fitting equations corresponding to fundamental and oscillation frequencies are established based on the characteristic polynomial coefficients of the matrix pencil, in which the angular frequency conjugation constraints of the two fundamental modes and two oscillation modes are considered. Multi-order matrix pencil is proposed to further smooth model errors.
2) Identification and recovery methods of bad data are proposed. The significance of bad data is evaluated by utilizing the difference in temporal continuity of the synchrophasors before and after noise reduction. Based on the significance level of the bad data, whether to perform noise reduction or recovery of bad data can be determined, thus reducing the impact of bad data on the performance of the MPM.
Finally, the proposed improved MPM is verified with synthetic, simulated, and actual measured PMU data. Also, it is compared with the existing MPM in [
This paper is organized as follows. Section II studies the synchrophasors model and principle of MPM. Section III introduces SSO parameter identification based on improved MPM. Section IV proposes the identification and recovery of improved MPM of the bad data. The performance of the proposed method is evaluated using synthetic, simulated, and actual measured PMU data in Section V. The conclusions are illustrated in Section VI.
The instantaneous data during SSO consists of fundamental, subsynchronous, and supersynchronous components, i.e.,
(1) |
where , , and are the frequencies, amplitudes, and phases of the fundamental, subsynchronous, and supersynchronous components of SSO, respectively. The frequencies of subsynchronous and supersynchronous components satisfy , and is the nominal frequency.
Synchrophasors are obtained by applying the DFT synchrophasor algorithm [
(2) |
(3) |
(4) |
where the subscripts “+” and “” represent the positive and negative spectra, respectively; the superscript “” represents the conjugation; and are the angular frequencies of and as shown in (5), respectively; and is illustrated in (6).
(5) |
(6) |
where is the number of instantaneous data in the DFT algorithm; and is the reporting rate of synchrophasor, and .
Due to the coupling frequency of the subsynchronous components and the supersynchronous components, the frequencies of positive and negative spectra of and are the same. Thus, (3) and (4) can be sumed to positive and negative spectra of oscillation component, respectively, as shown in (7).
(7) |
where and are the positive and negative spectra of oscillation components, respectively.
Consequently, the synchrophasor consists of four parts, as expressed in (8): the positive and negative spectra of fundamental components vary at angular frequencies and , and the positive and negative spectra of oscillation components vary at angular frequencies and .
(8) |
Based on the MPE algorithm, can be transformed into a modal model consisting of four modes as shown in (9)-(11).
(9) |
(10) |
(11) |
where and are the amplitude and angular frequency of mode m, respectively; and and correspond to fundamental modes, and and correspond to oscillation modes.
Note that only is a variable in (9) representing the features of angular frequency changes with time. As shown in

Fig. 1 Single-phase current synchrophasors.
The basic idea of MPM is shown in

Fig. 2 Basic idea of MPM.
The first step is to construct a Hankel matrix based on synchrophasors as:
(12) |
where is the number of synchrophasors; is the number of columns of the Hankel matrix; and can be rewritten as:
(13) |
(14) |
(15) |
(16) |
where matrix represents the time continuity of each column in ; matrix represents the time continuity of each row in ; and matrix represents the amplitude of the four modes. Then, can be decomposed into two submatrices and as:
(17) |
As illustrated in
(18) |
where ; and is a unit matrix.
When is equal to any one of the exponential of the angular frequency among the four modes, the order of can be reduced, and the characteristic polynomial of is equal to 0 as:
(19) |
Finally, by solving (19), the generalized eigenvalues of , i.e., can be obtained, and can be further calculated based on and (9).
In [
Furthermore, as is close to nominal frequency, the value of is quietly small, and even tends to zero when tends to 50 Hz, as shown in

Fig. 3 and varing with frequency.
The fitness of MPM is checked by the coefficient of determination in [
The basic idea of the proposed improved MPM is to establish the angular frequency fitting equations corresponding to fundamental and oscillation frequencies based on the characteristic polynomial coefficients of the matrix pencil, in which the angular frequency conjugation constraints of the two fundamental modes and two oscillation modes are considered, thus guaranteeing the accuracy of calculated and .
Based on the properties of the matrix characteristic polynomial, (19) can be derived as:
(20) |
where is the characteristic polynomial coefficient of the matrix pencil.
Since can be obtained based on the eigenvalues of the matrix pencil, the characteristic polynomial coefficients are equivalent to eigenvalues. In the presence of model errors, calculated by suffers from calculation errors; the value on the left side of (20) is no longer zero when the correct or is substituted, and residuals appear instead. By searching or that minimizes the residuals, or can be solved exactly.
Therefore, after constructing the matrix pencil, the characteristic polynomial coefficients of the matrix pencil are calculated first, and then the fitting equations of and are established based on the characteristic polynomial coefficients of the matrix pencil as:
(21) |
(22) |
where , , , and are the residuals of the fitting equations.
On the one hand, the fitting equations of and replace the eigenvalues of the matrix pencil with the characteristic polynomial coefficients, which can approximately characterize the features of the angular frequencies of the four modes. On the other hand, the conjugate constraints of and are considered. Solving the minimum residuals corresponding to and will eliminate and with more significant errors. The obtained and can satisfy the conjugate constraints of and , and they are closest to the eigenvalues of the matrix pencil.
The numerical solution of (21) and (22) can be obtained as and by minimizing the residuals and , respectively. Then, the amplitude of each mode can be obtained based on (9), and the parameters of the fundamental, subsynchronous, and supersynchronous components can be obtained based on (10).
In the existing studies, Hankel matrix is decomposed into two submatrices to construct the matrix pencil, which contains model errors corresponding to phasors in solving the eigenvalues of the matrix pencil. The multi-order matrix pencil is proposed to reduce the impact of the randomness of model errors in this subsection. is decomposed into multiple submatrices so that the matrix pencil can be extended to multiple orders. Based on the mean value of the characteristic polynomial coefficients of the multi-order matrix pencil, the angle frequency fitting equations of and can be constructed. Hence, the model errors are smoothed to further improve the accuracy of and .
First, is decomposed into multiple submatrices as:
(23) |
where is the number of columns of the submatrix and ; and is the number of submatrices, .
The multi-order matrix pencil is constructed as:
(24) |
where the superscript indicates the matrix pencil, .
Multiple sets of characteristic polynomial equations can be obtained by substituting - into (20). Since and are solutions of all characteristic polynomial equations, by superimposing these characteristic equations directly, the angle frequency fitting equations of and based on the multi-order matrix pencil can be constructed as:
(25) |
(26) |
The solutions of (25) and (26) are the same as that of (21) and (22), respectively.
For identification of SSO based on the synchrophasors, the shorter the data window is, the less low-frequency information in the synchrophasors is, and it is more difficult to identify SSO when the subsynchronous frequency is close to the fundamental frequency .
Since the DFT-based algorithm performs an approximation calculation in the process of identification, the errors of identification results are still significant when is close to , even with a 2-s data window [
The existing studies of MPM ignore the angular frequency conjugation constraints of the two fundamental modes and two oscillation modes when solving the eigenvalues of the matrix pencil directly, so it has to use a 1-s data window to ensure the accuracy of parameter identification. It is precise because the proposed improved MPM compensates for this deficiency that the advantage of not performing the approximation calculation of the MPM can be fully utilized, and the data window for identifying SSO can be shortened to 200 ms.
Noise is one of the factors leading to model errors and will affect the accuracy of parameter identification based on MPM. Before implementing the MPM, singular value decomposition (SVD) based rank reduction is usually applied to the Hankel matrix [
(27) |
where is a diagonal matrix with singular values of ; and and are two unitary matrices. As the first four singular values in correspond to each of the four modes, by intercepting the first four singular values as and the corresponding singular phase quantities as and , the noise that is less significant than the four modes can be removed. In turn, the noise reduction is realized, and the Hankel matrix after noise reduction is given as:
(28) |
As only the first four singular values are retained after the noise reduction, the larger the dimension of the matrix is, the better the noise reduction performance is. For a fixed number of synchrophasors, the dimension of is the largest when the Hankel matrix is a square matrix, so the Hankel matrix will be constructed as a square matrix in this paper to maximize the noise elimination.
The reason why SVD can still retain the features of four modes despite the error of is significant is that the amplitude corresponding to is small enough, thus weakening the effect of on the whole fundamental mode . As long as the noise is less significant to the four modes, noise can be reduced to some extent.
Although noise will cause deviations in all of the synchrophasors, and bad data will cause deviation in a single data point, the effects of noise and bad data on the performance of the MPM are not substantially different, as these deviations are averaged to each row and column of the Hankel matrix during the feature extraction process. When the bad data deviates seriously, the temporal continuity of synchrophasors in the Hankel matrix may be broken. Hence, the parameters obtained by the MPM will be totally wrong, and it is difficult to eliminate the impact of bad data through noise reduction.
The process of noise reduction with single bad data is illustrated in

Fig. 4 Process of noise reduction with single bad data.
By defining , the points in are as:
(29) |
where denotes the difference of and ; is the point of the row and column in ; and and are the points in and , respectively.
The difference of the synchrophasors before and after taking the noise reduction is calculated as:
(30) |
Then, the ratio of the maximum difference to the average difference of the synchrophasors before and after taking noise reduction is:
(31) |
The value of reflects the prominence of the difference in bad data compared with the average difference. Hence, the significance of bad data can be evaluated by comparing the calculated with a threshold value. When the calculated is less than the threshold value, the significance of bad data is relatively low, and the impact of bad data is minimal, which can be treated as noise. The calculated error of parameter identification can be decreased to an acceptable range by noise reduction. In contrast, when the calculated is greater than the threshold value, the difference in bad data is significant enough to be identified by locating . In addition, the presence of bad data may yield obvious errors for parameter identification, the effect of which is difficult to eliminate by noise reduction, so the recovery of the bad data is necessary.
Note that the threshold value of proposed in this paper is not an indicator to distinguish bad data and non-bad data, but a robustness indicator to reduce the impact of bad data on the performance of the MPM. From this perspective, the threshold value does not need to be solved precisely, while it is necessary to ensure that the impact of bad data is minimal enough when the calculated is close to the threshold value. Although the misclassification or omission of bad data may occur under these circumstances, performing either noise reduction or the recovery of bad data can eliminate the impact of bad data.
The basic idea of the recovery of bad data is to correct bad data by performing iterative SVD-based rank reduction of , and replacing the points related to bad data in with the points related to bad data in until the difference of bad data satisfies the recovery accuracy criterion. The process of the recovery of bad data is as follows.
Step 1: begin with the
Step 2: calculate the difference between and , obtain , and then calculate the difference of bad data based on as:
(32) |
where is the position of bad data, and it is obtained previously as introduced in Section IV-B.
Step 3: determine if meets the recovery accuracy criterion. If is less than the recovery accuracy criterion, the current is sufficiently accurate, and the recovery of bad data is unnecessary. If not, obtain the position of the point with the minimum difference of bad data as , and then search the point corresponding to the bad data with the minimum difference in the same position from . Set the existing Hankel matrix in the next iteration as , and replace the point corresponding to bad data in with as:
(33) |
Finally, continue to the next iteration with , until satisfies the recovery accuracy criterion.
Actual measurements may have missing data in addition to bad data. Since the synchrophasors provided by PMU have unified time stamps, if the synchrophasor corresponds to any time is missing, the position of missing data can be determined based on the unified time stamps.
After determining the position of missing data, when assigning an extremely large or small value compared to other synchrophasors to the position of missing data, it is equivalent to the presence of bad data in the synchrophasors and the recovery of the missing data can be addressed based on the identification and recovery of the bad data, as shown in
The overall characteristics of the proposed improved MPM is verified with synthetic and simulated PMU data in Section V-A and V-B, respectively. Meanwhile, the performance of the proposed identification and recovery of bad data are verified based on the synthetic PMU data in Section V-C.
The synthetic PMU data are modeled as (1), and the fundamental frequency is set as Hz. Other parameters of fundamental, subsynchronous, and supersynchronous components are set as , , and , respectively.
The sampling frequency of the instantaneous data is 1.6 kHz, and the reporting rate of PMU is 100 Hz. Performing the DFT algorithm on the instantaneous data can obtain the synthetic PMU data. Then, the improved SSO parameter identification based on MPM is implemented with a 200-ms data window, i.e., 21 synchrophasor data.
The primary research [
As the measurement signal to noise ratio (SNR) of PMUs is around 45 dB [
Test set | Method | SNR (dB) | |||||||
---|---|---|---|---|---|---|---|---|---|
IMPM (200 ms) | |||||||||
40 | |||||||||
30 | |||||||||
MPM (200 ms) | |||||||||
40 | |||||||||
30 | |||||||||
MPM (1 s) | |||||||||
40 | |||||||||
30 | |||||||||
IMPM (200 ms) | |||||||||
40 | |||||||||
30 | |||||||||
MPM (200 ms) | |||||||||
40 | |||||||||
30 | |||||||||
MPM (1 s) | |||||||||
40 | |||||||||
30 | .0 |
Note: ① the bolded number means the relative errors are beyond . ② means the ideal condition. ③ The relative errors of and are similar to and , respectively, so both of them are not listed here.
The relative errors of parameters of each component in SSO obtained by the improved MPM are on the order of , which is closer to the accuracy of the existing MPM with the 1-s data window and superior to that of the existing MPM with the 200-ms data window.
Whether using a 200-ms data window or a 1-s data window, the relative errors of and obtained by the existing MPM are approximately , and the main reason is stated in Section III-A. As the conjugate constraints of are considered in the fitting equations of , the relative errors of the fundamental components obtained by the improved MPM are below the order of .
According to the theory of spectral analysis, the shorter the data window is, the lower the spectrum resolution ratio is, and the interaction between the fundamental component and the oscillation component is severe. Therefore, under the noise condition, corrupted by the low identification accuracy of the fundamental components, the relative error of the oscillation components obtained by the existing MPM with the 200-ms data window is significant, among which the relative error of the amplitude and phase reaches and , respectively. In contrast, as the 1-s data window is long enough, the interaction between the fundamental component and the oscillation component is insignificant, and the relative error of the oscillation components obtained by the existing MPM with the 1-s data window is approximately . Since the improved MPM can identify the fundamental component parameters with high accuracy, the relative error of the oscillation components obtained by the improved MPM with the 200-ms data window is still below .
Consequently, compared with the existing MPM, the proposed improved MPM can accurately identify the parameters for the fundamental, subsynchronous, and supersynchronous components with a 200-ms data window under both frequency-varying conditions and noisy conditions, which significantly improves the real-time performance.
To further verify the effectiveness of the proposed improved MPM, cases with both simulated PMU data and actual measured PMU data are conducted. The simulated PMU data are generated in the PSCAD platform based on the model in [

Fig. 5 Diagram of simulation model.

Fig. 6 Instantaneous data and synchrophasors for simulated PMU data case. (a) Instantaneous data. (b) Amplitude of synchrophasor.

Fig. 7 Instantaneous data and synchrophasors for actual measured PMU data case. (a) Instantaneous data. (b) Amplitude of synchrophasor.
The simulated PMU data with a 6-s data window are analyzed as in

Fig. 8 Identified results of frequency and amplitude for simulated PMU data case. (a) Fundamental component. (b) Subsynchronous component.

Fig. 9 Identified results of frequency and amplitude for actual measured PMU data case. (a) Fundamental component. (b) Subsynchronous component.
In the simulated PMU data case, the results of obtained by the improved MPM fluctuate at approximately 49.47 Hz to 49.57 Hz, and those of interpolated DFT are around 49.50 Hz. However, in the
The results of the supersynchronous component have similar characteristics to the subsynchronous one, so only the results of the subsynchronous component are analyzed here. In the simulated PMU data case, the results of obtained by the improved MPM fluctuate at approximately 8.13 Hz to 8.25 Hz, and those obtained by the existing MPM and interpolated DFT are around 8.19 Hz; the results of obtained by the improved MPM fluctuate at approximately 2.95 p.u. to 3.04 p.u., and those obtained by the existing MPM and IDF are around 3.0 p.u.. The proposed improved MPM can identify the frequency and amplitude of oscillation component accurately.
In the actual measured PMU data case, the results of obtained by the improved MPM fluctuate between approximately 8.14 Hz to 8.52 Hz, and those obtained by the existing MPM and interpolated DFT fluctuate around 8.24 Hz to 8.28 Hz; the variation range of the results of obtained by the improved MPM are similar to those obtained by the existing MPM and interpolated DFT, but and obtained by improved MPM change rapidly with time. Since the identified results of the existing MPM and interpolated DFT are the average of dynamics in 1-s and 2-s data windows and cannot reflect the actual characteristics of SSO at a particular moment, especially during the fast-changing stage of SSO, owing to the 200-ms data window, the improved MPM can intuitively monitor the dynamics of the frequency and amplitude of each component in the process of SSO.
This subsection verifies the performance of identification and recovery of bad data with synthetic PMU data. The parameters of fundamental, subsynchronous, and supersynchronous components are , , and , respectively; the other parameters are the same as those in Section V-A; and the noise with 45 dB SNR is added.
As the improved MPM is performed on 200-ms data windows and contains 21 synchrophasor data for each calculation, the constructed Hankel matrix is an symmetric matrix in this paper, and the
To analyze the effect of bad data, the total vector error (TVE) defined by IEEE standard [
(34) |
where and are the bad data and initial phasor data, respectively.
The ratio of the maximum difference to the average difference with bad data under different is illustrated in

Fig. 10 Value of and identified position of bad data under different . (a) . (b) .
The position of bad data identified by locating with bad data under different is also shown in
The maximum relative errors of the identified results based on the improved MPM before and after the recovery of bad data are illustrated in

Fig. 11 The maximum relative errors before and after recovery of bad data. (a) . (b) .
The characteristics of may vary slightly in other scenarios such as different parameters of PMU data or different positions of bad data, in which misclassification or omission of bad data may occur when is close to the threshold value. However, as stated in Section V-B, the threshold value does not need to be solved exactly. The threshold value of 2.5 is robust enough to ensure that the bad data can be solved by either noise reduction or recovery when is close to 2.5 and thus can also be utilized in other scenarios.
Consequently, the proposed identification and recovery of bad data can significantly reduce the impact of bad data and guarantee the performance of the MPM effectively.
An improved SSO parameter identification based on MPM is proposed by utilizing the high computing efficiency and noise-resilient features of MPM to realize the dynamic monitoring of SSO. The MPM is enhanced by the angular frequency fitting equations based on the characteristic polynomial coefficients of the matrix pencil to ensure the accuracy of the identified parameters, since the existing eigenvalue estimation of MPM ignores the angular frequency conjugation constraints of the two fundamental modes and two oscillation modes. Then, the MPM is enhanced by the identification and recovery of bad data by utilizing the difference in temporal continuity of the synchrophasors before and after noise reduction. The proposed improved MPM is verified with synthetic, simulated, and actual measured PMU data. Compared with the existing MPM, the improved MPM achieves better accuracy for parameter identification of fundamental, subsynchronous, and supersynchronous components with a 200-ms data window. The improvements in the identification and recovery of bad data can significantly reduce the impact of bad data, thus enhancing the practicability of MPM.
References
L. Wang, X. Xie, Q. Jiang et al., “Investigation of SSR in practical DFIG-based wind farms connected to a series-compensated power system,” IEEE Transactions on Power Systems, vol. 30, no. 5, pp. 2772-2779, Feb. 2015. [Baidu Scholar]
C. Lin, “Simplified modelling of oscillation mode for wind power systems,” Journal of Modern Power Systems and Clean Energy, vol. 9, no. 1, pp. 56-67, Jan. 2021. [Baidu Scholar]
H. Liu, T. Bi, X. Chang et al., “Impacts of subsynchronous and supersynchronous frequency components on synchrophasor measurements,” Journal of Modern Power Systems and Clean Energy, vol. 4, no. 3, pp. 362-369, Jul. 2016. [Baidu Scholar]
L. Kilgore, D. Ramey, and M. Hall, “Simplified transmission and generation system analysis procedures for subsynchronous resonance problems,” IEEE Transactions on Power Apparatus and systems, vol. 96, no. 6, pp. 1840-1846, Nov. 1977. [Baidu Scholar]
Y. Li, L. Fan, and Z. Miao, “Replicating real-world wind farm SSR events,” IEEE Transactions on Power Delivery, vol. 35, no. 1, pp. 339-348, Feb. 2020. [Baidu Scholar]
X. Xie, Y. Zhan, H. Liu et al., “Wide-area monitoring and early-warning of subsynchronous oscillation in power systems with high-penetration of renewables,” International Journal of Electrical Power & Energy Systems, vol. 108, pp. 31-39, Jun. 2019. [Baidu Scholar]
J. Shair, X. Xie, and G. Yan, “Mitigating subsynchronous control interaction in wind power systems: existing techniques and open challenges,” Renewable and Sustainable Energy Reviews, vol. 108, pp. 330-346, Apr. 2019. [Baidu Scholar]
J. Shair, X. Xie, J. Yang et al., “Adaptive damping control of subsynchronous oscillation in DFIG-based wind farms connected to series-compensated network,” IEEE Transactions on Power Delivery, vol. 37, no. 2, pp. 1036-1049, Apr. 2022. [Baidu Scholar]
H. Khalilinia and V. Venkatasubramanian, “Subsynchronous resonance monitoring using ambient high speed sensor data,” IEEE Transactions on Power Systems, vol. 31, no. 2, pp. 1073-1083, Mar. 2016. [Baidu Scholar]
X. Xie, H. Liu, Y. Wang et al., “Measurement of sub-and supersynchronous phasors in power systems with high penetration of renewables,” in Proceedings of 2016 IEEE Power & Energy Society Innovative Smart Grid Technologies Conference (ISGT), Minneapolis, USA, Dec. 2016 , pp. 1-5. [Baidu Scholar]
H. Lin, “Power harmonics and interharmonics measurement using recursive group-harmonic power minimizing algorithm,” IEEE Transactions on Industrial Electronics, vol. 59, no. 2, pp. 1184-1193, Feb. 2012. [Baidu Scholar]
N. Zhou, D. J. Trudnowski, J. W. Pierre et al., “Electromechanical mode online estimation using regularized robust RLS methods,” IEEE Transactions on Power Systems, vol. 23, no. 4, pp. 1670-1680, Nov. 2008. [Baidu Scholar]
T. Rauhala, A. M. Gole, and P. Järventausta, “Detection of subsynchronous torsional oscillation frequencies using phasor measurement,” IEEE Transactions on Power Delivery, vol. 31, no. 1, pp. 11-19, Feb. 2016. [Baidu Scholar]
P. Wall, P. Dattaray, Z. Jin et al., “Deployment and demonstration of wide area monitoring system in power system of Great Britain,” Journal of Modern Power Systems and Clean Energy, vol. 4, no. 3, pp. 506-518, Jul. 2016. [Baidu Scholar]
F. Zhang, L. Cheng, W. Gao et al., “Synchrophasors-based identification for subsynchronous oscillations in power systems,” IEEE Transactions on Smart Grid, vol. 10, no. 2, pp. 2224-2233, Mar. 2019. [Baidu Scholar]
X. Yang, J. Zhang, X. Xie et al., “Interpolated DFT-based identification of sub-synchronous oscillation parameters using synchrophasor data,” IEEE Transactions on Smart Grid, vol. 11, no. 3, pp. 2662-2675, May 2020. [Baidu Scholar]
J. F. Hauer, C. Demeure, and L. Scharf, “Initial results in prony analysis of power system response signals,” IEEE Transactions on Power Systems, vol. 5, no. 1, pp. 80-89, Feb. 1990. [Baidu Scholar]
M. Netto and L. Mili, “A robust prony method for power system electromechanical modes identification,” in Proceedings of 2017 IEEE PES General Meeting, Chicago, USA, Jul. 2017, pp. 1-5. [Baidu Scholar]
M. L. Crow and A. Singh, “The matrix pencil for power system modal extraction,” IEEE Transactions on Power Systems, vol. 20, no. 1, pp. 501-502, Feb. 2005. [Baidu Scholar]
Y. Wang, X. Jiang, X. Xie et al., “Identifying sources of subsynchronous resonance using wide-area phasor measurements,” IEEE Transactions on Power Delivery, vol. 36, no. 5, pp. 3242-3254, Oct. 2021. [Baidu Scholar]
H. Liu, Y. Qi, J. Zhao et al., “Data-driven subsynchronous oscillation identification using field synchrophasor measurements,” IEEE Transactions on Power Delivery, vol. 37, no. 1, pp. 165-175, Feb. 2022. [Baidu Scholar]
Y. Zhu, C. Liu, and K. Sun, “Image embedding of pmu data for deep learning towards transient disturbance classification,” in Proceedings of 2018 IEEE International Conference on Energy Internet (ICEI), Beijing, China, May 2018, pp. 169-174. [Baidu Scholar]
F. Zhang, J. Li, J. Liu et al., “An improved interpolated dft-based parameter identification for sub-/super-synchronous oscillations with synchrophasors,” IEEE Transactions on Power Systems, vol. 38, no. 2, pp. 1741-1727, Mar. 2022. [Baidu Scholar]
IEEE Standard for Synchrophasor Measurements for Power Systems. IEEE Std C37.118.1-2011, Dec. 2011. [Baidu Scholar]