Professional Documents
Culture Documents
3, AUGUST 2006
Abstract—This paper describes a new algorithm to compute the faster, more robust, and more flexible algorithm. To avoid
dominant poles of a high-order scalar transfer function. The al- repeated computation of the same dominant poles, a deflation
gorithm, called the subspace accelerated dominant pole algorithm strategy is used. The SADPA directly operates on implicit
(SADPA), is more robust than existing methods in finding both real
and complex dominant poles and faster because of subspace accel- state–space systems, also known as descriptor systems, which
eration. SADPA is able to compute the full set of dominant poles are very sparse in practical power system applications.
and produce good modal equivalents automatically, without any This paper is organized as follows. Section II summarizes
human interaction. some well-known properties of scalar transfer functions and for-
Index Terms—Dominant pole spectrum, large-scale systems, mulates the problem of computing the dominant poles of a scalar
modal equivalents, model reduction, poorly-damped oscillations, transfer function. Section III discusses the DPSE algorithm [7].
power system dynamics, small-signal stability, sparse eigenanal- In Section IV, the new variant SADPA is described. Extensive
ysis, system poles, transfer function.
numerical results are presented in Section V. Section VI con-
cludes this paper.
I. INTRODUCTION
II. TRANSFER FUNCTIONS AND DOMINANT POLES
Manuscript received November 29, 2005; revised January 12, 2006. Paper no.
TPWRS-00753-2005.
J. Rommes is with Mathematical Institute, Utrecht University, 3508 TA A pole that corresponds to a residue with large
Utrecht, The Netherlands (e-mail: rommes@math.uu.nl). is called a dominant pole, i.e., a pole that is well
N. Martins is with CEPEL, Rio de Janeiro RJ-20001-970, Brazil (e-mail:
nelson@cepel.br). observable and controllable in the transfer function. This can
Digital Object Identifier 10.1109/TPWRS.2006.876671 also be observed from the corresponding Bode magnitude plot
0885-8950/$20.00 © 2006 IEEE
ROMMES AND MARTINS: EFFICIENT COMPUTATION OF TRANSFER FUNCTION DOMINANT POLES USING SUBSPACE ACCELERATION 1219
of , where peaks occur at frequencies close to the imagi- 3) Solve and from
nary parts of the dominant poles of . An approximation of
that consists of terms with above
some value, determines the effective transfer function behavior
[13]
The DPSE [7] is based on the DPA [6] and refactored bi-it- 6) The pole has converged if
eration (RBI) [14]. A scalar transfer function (1) is equal to its
transpose
which can be solved with the method [17]. In fact, the new algorithm SADPA can be seen as a general-
ization of the DPA to compute more than one dominant pole.
Algorithm 2: The Dominant Pole Spectrum Eigensolver First, subspace acceleration, a well-known technique for itera-
tive methods, will be described. Second, a new selection strategy
INPUT: System initial pole estimates will be used to select the most dominant pole approximation
and corresponding right and left eigenvector approximation for
every iteration. Third, deflation will be used to avoid conver-
OUTPUT: dominant pole triplets gence to eigentriplets that are already found.
A. Subspace Acceleration
1) Set A drawback of the DPSE is that information obtained in the
current iteration is discarded at the end of the iteration. The only
2) while not all converged do
information that is preserved is contained in the new pole esti-
3) for do mates . The subspaces and , however, also contain in-
formation about other dominant eigentriplets (i.e., components
4) Solve from
in the direction of the corresponding eigenvectors), and the idea
is now to use this information as well. Reasoning this way leads
to a generalization of the DPA.
A global overview of the SADPA is shown in algorithm 3.
Starting with a single shift , the first iteration is equivalent
5) Solve from to the first iteration of the DPA, but instead of discarding the
corresponding right and left eigenvector approximations
and , they are kept in spaces and . In the next iteration,
these spaces are expanded orthogonally, by using modified
Gram–Schmidt (MGS) [17], with the approximations and
corresponding to the new shift (see Section IV-C). Hence,
6) endfor the spaces grow and will contain better approximations. This
idea is known as subspace acceleration. Approximations of the
7) Set and dominant poles are computed like in the DPSE, but the next
8) Compute and question is to select the most promising approximation as new
shift: unlike the DPSE, only one shift per iteration is used.
9) Compute the new pole estimates
Algorithm 3: Subspace Accelerated DPA
15)
Alg. 6 has the same eigentriplets as but with the found eigenvalues
transformed to zero (see also [19], [20]): let be one of the
16) found right eigenvectors, i.e., . Then it follows
17) Set from the orthogonality relations (see Section II) that
18) end if
19) Set
20) Set the new pole estimate
21) end while
INPUT:
OUTPUT: with and
OUTPUT: with
the pole with largest residue magnitude and
and the corresponding approximate right and
1)
left eigenvectors.
1) Compute residues 2)
TABLE I
RESULTS OF SADPA FOR SIX TRANSFER FUNCTIONS
OF THE BIPS. SHIFT s = 1j
Fig. 2. Bode plot of modal equivalent, complete model, and error for transfer
function W =V ref of BIPS (95 states in the modal equivalent, 1664 in
the complete model).
Fig. 1. Bode plot of modal equivalent, complete model, and error for transfer
function W =V ref of BIPS (102 states in the modal equivalent, 1664
in the complete model).
Fig. 5. Bode plot of modal equivalent, complete model, and error for transfer
Fig. 4. Bode plot of modal equivalent, complete model, and error for transfer function P (s)=B (s) of BIPS (41 states in the modal equivalent, 1664 in the
function W =V ref of BIPS (104 states in the modal equivalent, 1664 complete model).
in the complete model).
TABLE II
FIFTEEN MOST DOMINANT POLES AND CORRESPONDING RESIDUES
OF W =V ref , SORTED IN DECREASING jR j=jRe( )j
ORDER (29-STATE MODAL EQUIVALENT)
Fig. 6. Step responses for transfer function P (s)=B (s) of BIPS, complete
model, and modal equivalent (41 states in the modal equivalent, 1664 in the
complete model, step disturbance of 0:01 p.u.).
TABLE III
COMPUTATIONAL STATISTICS FOR THE SADPA (s = 1j , THE DPSE,
AND THE DPSE WITH DEFLATION (DPSED) (SHIFTS 1j; 1:5j; . . . ; 10:5j ), both responses, and the reduced model nicely captures the
AFTER 100 S COMPUTATIONAL TIME system oscillations. The reduced model (30 poles, 56 states)
was computed by SADPA in 200 s (341 factorizations). This
reduced model was reduced to a 41 state (22 poles) by removing
less dominant contributions. A table with the dominant poles
and corresponding residues can be found in [23].
The sixth transfer function, , relates the ac-
value of the TCSC series admittance. This transfer function tive power deviations of a large hydroelectric plant, located in
was used in the basic design of the POD controllers of the two the Southeast region, to disturbances applied to its speed-gov-
TCSCs, in order to damp the North–South mode [23]–[25]. ernor reference. For this transfer function, it is known from nu-
Modal equivalents of the transfer function for damping the merical experiments that the DPSE has difficulties in producing
North–South mode, whose state–space realization has a direct a good modal equivalent: several DPSE attempts are needed
transmission term , are considered in [23]. to obtain an acceptable modal equivalent. The SADPA is able
Fig. 5 shows the frequency response of the complete model to automatically produce good results for both the frequency
and the reduced model (41 states) together with the error. and step response, as can be observed from Figs. 7 and 8. The
Fig. 6 shows the corresponding step response (step ). The SADPA computed the reduced model (80 poles, 118 states) in
North–South mode ( rad/s Hz is well observable in 450 s (533 factorizations) using just a single initial shift ,
ROMMES AND MARTINS: EFFICIENT COMPUTATION OF TRANSFER FUNCTION DOMINANT POLES USING SUBSPACE ACCELERATION 1225
Fig. 7. Bode plot of modal equivalent, complete model, and error for transfer
function P t ( ) Pref ( )
s= s of BIPS (118 states in the modal equivalent,
Fig. 9. Bode plot of modal equivalent, complete model, and error for transfer
1 1
1664 in the complete model).
function ! = P of the New England test system (five states in the modal
equivalent, 66 in the complete model).
small signal stability, but the SADPA algorithm is general and [9] A. C. Antoulas and D. C. Sorensen, “Approximation of large-scale dy-
could be effectively applied to problems in other engineering namical systems: An overview,” Int. J. Appl. Math. Comput. Sci., vol.
11, no. 5, pp. 1093–1121, 2001.
fields that allow sparse descriptor system formulations. [10] A. C. Antoulas, Approximation of Large-Scale Dynamical Systems.
Currently, a generalization of the algorithm to MIMO systems Philadelphia, PA: SIAM, 2005.
is being developed [8]. [11] M. Green and D. J. N. Limebeer, Linear Robust Control. Englewood
Cliffs, NJ: Prentice-Hall, 1995.
[12] T. Kailath, Linear Systems. Englewood Cliffs, NJ: Prentice-Hall,
1980.
[13] J. R. Smith, J. F. Hauer, D. J. Trudnowski, F. Fatehi, and C. S. Woods,
APPENDIX I “Transfer function identification in power system application,” IEEE
SMALL TEST SYSTEM Trans. Power Syst., vol. 8, no. 3, pp. 1282–1290, Aug. 1993.
[14] J. M. Campagnolo, N. Martins, and D. M. Falcão, “Refactored bi-iter-
For illustrational purposes, the SADPA was applied to a ation: A high performance eigensolution method for large power sys-
tems,” IEEE Trans. Power Syst., vol. 11, no. 3, pp. 1228–1235, Aug.
transfer function of the New England test system. This small 1996.
benchmark system has 66 state variables (for more information, [15] L. H. Bezerra, “Written discussion to [6],” IEEE Trans. Power Syst.,
see [6]). The transfer function was vol. 11, no. 1, p. 168, Feb. 1996.
[16] B. N. Parlett, “The Rayleigh quotient iteration and some generaliza-
being the rotor speed and the applied mechanical tions for nonnormal matrices,” Math. Comp., vol. 28, no. 127, pp.
power deviations to the synchronous generator at bus 36. The 679–693, Jul. 1974.
tolerance used was , and no restarts were used. Every [17] G. H. Golub and C. F. van Loan, Matrix Computations, 3rd ed. Bal-
timore, MD: John Hopkins Univ. Press, 1996.
iteration, the pole approximation with largest [18] Y. Saad, Numerical Methods for Large Eigenvalue Problems: Theory
was selected. Table IV shows the found dominant poles and the and Algorithms. Manchester, U.K.: Manchester Univ. Press, 1992.
iteration number in which the pole converged. Bodeplots of two [19] M. E. Hochstenbach, “Two-sided and alternating Jacobi-Davidson,”
Lin. Alg. Appl., vol. 358, no. 1–3, pp. 145–172, 2003.
modal equivalents are shown in Figs. 9 and 10. The quality of [20] D. R. Fokkema, G. L. G. Sleijpen, and H. A. van der Vorst, “Jacobi-
the modal equivalent increases with the number of found poles. Davidson style QR and QZ algorithms for the reduction of matrix pen-
cils,” SIAM J. Sci. Comp., vol. 20, no. 1, pp. 94–125, 1998.
[21] L. A. Aguirre, “Quantitative measure of modal dominance for contin-
ACKNOWLEDGMENT uous systems,” in Proc. 32nd Conf. Decision Control, Dec. 1993, pp.
2405–2410.
[22] The Mathworks, Inc., Matlab R13 [Online]. Available: http://www.
The first author would like to thank G. Sleijpen for pointing mathworks.com.
to the work of N. Martins. Both authors would like to thank [23] N. Martins, F. G. Silva, and P. C. Pellanda, “Utilizing transfer func-
CEPEL and ELETROBRAS for providing the test systems. tion modal equivalents of low-order for the design of power oscillation
damping controllers in large power systems,” in Proc. IEEE Power Eng.
Soc. General Meeting, June 2005, pp. 2642–2648.
[24] C. Gama, L. Ängquist, G. Ingeström, and M. Noroozian, “Commis-
REFERENCES sioning and operative experience of TCSC for damping power oscil-
lation in the Brazilian north-south interconnection,” in Proc. CIGRÉ
[1] J. J. Sanchez-Gasca and J. H. Chow, “Power system reduction to sim- Session No. 38, Aug. 2000, paper 14-104..
plify the design of damping controllers for interarea oscillations,” IEEE [25] Impact of the Interactions Among Power System Controls, CIGRE,
Trans. Power Syst., vol. 11, no. 3, pp. 1342–1349, Aug. 1996. Tech. Rep. 166, Jul. 2000.
[2] B. C. Pal, A. H. Coonick, and B. J. Cory, “Robust damping of inter-area
oscillations in power systems with superconducting magnetic energy
storage devices,” Proc. Inst. Elect. Eng., Gen., Transm., Distrib., vol.
146, no. 6, pp. 633–639, Nov. 1999.
[3] D. Chaniotis and M. A. Pai, “Model reduction in power systems using Joost Rommes received the M.Sc. degree in computational science in 2002 and
Krylov subspace methods,” IEEE Trans. Power Syst., vol. 20, no. 2, pp. the M.Sc. degree in computer science in 2003 from Utrecht University, Utrecht,
888–894, May 2005. The Netherlands. Currently, he is pursuing the Ph.D. degree at the Mathemat-
[4] A. Ramirez, A. Semlyen, and R. Iravani, “Order reduction of the ical Institute, Utrecht University, and works on model reduction and large scale
dynamic model of a linear weakly periodic system—Part I: General eigenvalue problems.
methodology,” IEEE Trans. Power Syst., vol. 19, no. 2, pp. 857–865,
May 2004.
[5] G. Troullinos, J. Dorsey, H. Wong, and J. Myers, “Reducing the order
of very large power system models,” IEEE Trans. Power Syst., vol. 3, Nelson Martins (SM’91–F’98) received the B.Sc. degree in electrical engi-
no. 1, pp. 127–133, Feb. 1988. neering in 1972 from University of Brasilia, Brasilia, Brazil, and the M.Sc. and
[6] N. Martins, L. T. G. Lima, and H. J. C. P. Pinto, “Computing dominant Ph.D. degrees from the University of Manchester Institute of Science and Tech-
poles of power system transfer functions,” IEEE Trans. Power Syst., nology, Manchester, U.K., in 1974 and 1978, respectively.
vol. 11, no. 1, pp. 162–170, Feb. 1996. He has been with CEPEL, Rio de Janeiro, since 1978 in the development of
[7] N. Martins, “The dominant pole spectrum eigensolver,” IEEE Trans. computer tools for power system dynamics and control.
Power Syst., vol. 12, no. 1, pp. 245–254, Feb. 1997. Dr. Martins is the Convenor of the CIGRE TF 38.02.23 on Coordinated
[8] N. Martins and P. E. M. Quintão, “Computing dominant poles of power Voltage Control in Transmission Networks and the current Chair of the Power
system multivariable transfer functions,” IEEE Trans. Power Syst., vol. System Stability Controls Subcommittee, PSDP Committee, IEEE Power
18, no. 1, pp. 152–159, Feb. 2003. Engineering Society.