IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 30, NO.

3, MAY 2015 1131

Computation of Dominant Poles and Residue

Matrices for Multivariable Transfer Functions
of Infinite Power System Models
Sergio Luis Varricchio , Senior Member, IEEE, Francisco Damasceno Freitas , Senior Member, IEEE,
Nelson Martins , Fellow, IEEE, and Franklin Clement Véliz

Abstract—This paper describes the first reliable Newton algo- the system is of interest, the number of RLC-PI circuits that are
rithm for the sequential computation of the set of dominant poles needed in the descriptor system model, to reproduce the infinite
of scalar and multivariable transfer functions of infinite systems. system behavior, becomes very large, rendering this approach
This dominant pole algorithm incorporates a deflation procedure,
which is derived from the partial fraction expansion concept of an- computationally less attractive than the approach of this
alytical functions of the complex frequency s and prevents the re- paper.
peated convergence to previously found poles. The pole residues The accurate computation of multivariable (multi-input
(scalars or matrices), which are needed in this expansion, are ac- multi-output—MIMO) transfer function (TF) dominant poles
curately computed by a Legendre-Gauss integral solver scheme for and their residue matrices are fundamental to modal analysis,
both scalar and multivariable systems. This algorithm is effectively
applied to the modal model reduction of multivariable transfer modal sensitivity, model order reduction and several other
functions for two test systems of considerable complexity and con- applications. An efficient algorithm to compute the dominant
taining many distributed parameter transmission lines. poles and associated residues of a high-order MIMO TF was
Index Terms—Distributed parameter transmission lines, dom- described in [4], but it is only applicable to finite, linear
inant poles, infinite systems, Legendre-Gauss integral solver descriptor systems. Regarding the eigensolution of infinite
scheme, modal analysis, model order reduction, multivariable system models, [5] presented a Newton algorithm for the
systems, network dynamics, Newton algorithm, transfer function sequential computation of dominant poles and their residues
residues.
for single-input single-output (SISO) TFs. The algorithm in
[5] computes the pole residues using a formula derived from
I. INTRODUCTION the limit definition of a pole-residue. The computation derived
from this formula is here shown to be subjected to severe

D ISTRIBUTED parameter transmission lines (TLs) can be round-off errors, a result which is similar to those reported
accurately modeled in the complex frequency domain in [6]. Thus, its use frequently renders useless the embedded
by transcendental functions having an infinite number of poles. eigenvalue deflation procedure of the algorithm in [5] and its
Therefore, electric power system network models containing sequential pole computation.
distributed parameter TLs are one of the many types of infinite The first major contribution of this paper is the use of a Le-
systems (having an infinite number of poles) that occur in prac- gendre-Gauss integral solver scheme for the accurate computa-
tice. Such systems are linear time invariant (LTI) and can be tion of the pole residues. The accuracy gained in the pole de-
adequately represented in the s-domain by a matrix [1], flation procedure, confers numerical robustness to the Newton
[2], whose entries are nonlinear functions of . A linear (finite) algorithm of [5] allowing the reliable sequential computation of
approximation model of these electric network models, rigor- the set of dominant poles for SISO TFs of infinite systems.
ously represented by infinite systems, can be obtained using, for The second major contribution of this paper is the general-
instance, a descriptor system approach, where each TL is repre- ization of the Newton algorithm in [5] for SISO TFs to mul-
sented by cascaded RLC-PI circuits [3], each PI-circuit having tivariable (MIMO) TFs of infinite systems (Sequential MIMO
a 3-state representation. When the high-frequency dynamics of Dominant Pole Algorithm—SMDPA). Once again, the set of ac-
curate poles can only be sequentially computed due to the accu-
rate computation of their associated residue matrices through a
Manuscript received August 17, 2013; revised December 05, 2013, February
22, 2014, May 27, 2014, and June 25, 2014; accepted June 26, 2014. Date of Legendre-Gauss integral solver scheme. Only after the residue
publication August 14, 2014; date of current version April 16, 2015. Paper no. matrix and pole are accurately computed, can this pole be ef-
TPWRS-01011-2013. fectively deflated from the MIMO TF. The Newton algorithm
S. L. Varricchio and N. Martins are with CEPEL, Rio de Janeiro, RJ,
CEP:20001-970, Brazil (e-mail: [email protected]; [email protected]). is actually applied to the accurately deflated MIMO TF, over-
F. D. Freitas is with the Department of Electrical Engineering, University of coming the severe limitations of the (SISO) algorithm in [5] and
Brasilia, Brasilia, DF, CEP:70910-900, Brazil (e-mail: [email protected]). allowing the smooth sequential convergence to a complete set
F. C. Véliz is with the Pontific Catholic University (PUC), Rio de Janeiro,
RJ, Brazil (e-mail: [email protected]). of dominant poles.
Color versions of one or more of the figures in this paper are available online The proposed algorithms were successfully applied to
at http://ieeexplore.ieee.org. the computation of high-fidelity (over a specified frequency
Digital Object Identifier 10.1109/TPWRS.2014.2336243

0885-8950 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1132 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 30, NO. 3, MAY 2015

window) reduced order models (ROMs) for MIMO TFs of two The subscript “ ” should rigorously be adopted in the notation
test systems, one having 34 buses and the other being the per- for the implicit matrix , but it was left out since there is no
vasive IEEE 118-bus system. The full data for these systems, explicit matrix in the of (4).
having 25 and 177 distributed parameter TLs, is available in The total direct matrix can be defined as
previous publications.
A third contribution of this paper is a comparison of the (6)
merits of a ROM directly derived from an infinite test system
The values for all entries of matrices and can be com-
with those from 4 linear approximation models (LAMs) of the
puted from the equations
same infinite system. One of these LAMs was then selected
for model order reduction through the application of two (7)
other algorithms (Subspace Accelerated Multivariable Dom-
inant Pole—SAMDP [4] and the Sparse Low-Rank Choleski (8)
Factor—SLRCF [7]) developed for descriptor system models.
The application of these methods to the chosen LAM, yielded by assigning a sufficiently large (i.e., or ) value to .
two other ROMs and allowed a CPU time performance com- Let be the strictly proper part of
parison with the proposed SMDPA algorithm.
The proposed SMDPA algorithm have important efficiency (9)
and robustness characteristics such as quadratic local con-
vergence, accurate eigenvalue deflation procedure avoiding Given a pole with associated residue matrix
repeated convergence to previously found TF dominant poles, , one has
ability to produce high-fidelity multivariable ROMs for infinite
systems. These small to moderate-size modal ROMs, which are (10)
directly derived from the infinite system models, have higher
modeling fidelity, more practical use and are obtained in less
CPU time than ROMs derived from large-scale LAMs and A pole associated with a large value of
well-known finite system methods such as SAMDP [4] and is called a dominant pole. From (10) it can be seen that the
SLRCF [7]. dominant poles will cause peaks in the frequency response
of the maximum singular values of -plot of
II. MULTIVARIABLE TRANSFER FUNCTIONS, THEIR or -plot) at frequency values that are
DOMINANT POLES, AND REDUCED MODELS close to the poles' imaginary parts. Obviously this is also true
The dynamic behavior of a linear infinite multivariable for the -plot of for frequencies close to where
system in the s-domain is described by .
A MIMO TF can be approximated by a ROM ,
(1) comprised of a set of dominant poles and associated
(2) residue matrices:
where
and . The subscript of (11)
the direct matrix stands for explicit and the superscript
denotes matrix or vector transpose. Solving (1) for and
The MIMO TF deviation, , is defined as the difference
substituting in (2), yields
between and :
(3)

The expression relating the output vector with the input

vector is defined as the transfer function matrix
: (12)

(4) The adopted measure for the MIMO ROM deviation, ,

is the -plot of computed at each frequency over an
Considering that all poles of are distinct, (4) can be specified frequency interval:
expanded as follows:
(13)
(5)
III. SEQUENTIAL MIMO DOMINANT POLE ALGORITHM
where is a pole of the multivariable and is its as- The dominant poles for SISO and MIMO TFs of infinite
sociated residue matrix. The subscript of matrix stands for systems, over a defined frequency window, can be computed
implicit in order to distinguish it from the explicit matrix and by Newton type, one-at-a-time eigenvalue finding algorithms.
because this matrix lies hidden inside the term . However, the complete set of dominant poles can only be
VARRICCHIO et al.: COMPUTATION OF DOMINANT POLES AND RESIDUE MATRICES 1133

efficiently computed by fully eliminating, in all subsequent where and are the left and right eigenvectors asso-
computations, the effect of the previously found poles of ciated with . The derivative of can be analytically
. This elimination, which is known as pole deflation [4], is determined by differentiating (12) with respect to , as follows:
essential to prevent the algorithm from repeatedly converging
to the already found poles due to their much larger convergence
attraction region [8]. As the presence of matrices and (26)
deteriorates the convergence of the algorithm, their effects must
also be eliminated [5]. The Newton method must therefore be with
applied to the MIMO TF deviation in (12), , whose poles
are the “not-yet-computed” sub-set of the poles of . (27)
The poles of the MIMO transfer function are those
for which
The derivative of , defined in (4), is given by
(14)
(28)
For square TFs , there is an equivalent criterion: the
poles are those for which
Similarly to (20) one can write
(15)
(29)
where is the minimum eigenvalue of .
For any , let be an eigentriplet of Substituting (29) into (28), one obtains
, i.e.:
(30)
(16)
(17) where

with (31)
(32)
(18)

where the superscript * denotes matrix or vector complex con-

IV. NUMERICAL BEHAVIOR OF THE RESIDUE FORMULA
jugate transpose.
DERIVED FROM THE CLASSICAL LIMIT DEFINITION
The derivative of is given by [9]
This section analyses the poor numerical stability of the
(19) residue computation by the formula presented in [2], [5], and
[10]. Computations employing this formula are shown here
where to lead to severe numerical errors that depend on the system
and the TF studied. These findings motivated the development
(20) of the numerically robust residue computation method of
Section V.
Equation (20), which holds for any square matrix, is then sub-
stituted into (19): A. Residue Formula
The limit definition of a residue associated with a pole is
(21) given by
which, by using (16) and (17) can be simplified to (33)

(22) where is a SISO TF defined similarly to its MIMO counter-

part :
Applying the Newton method to the solution of
(34)
(23)
If is considered to be an entry of then the vectors and
yields
are columns of the matrices and , respectively, and is
(24) an entry of .
with The limit definition in (33) yields the residue computational
formula for infinite systems [2]:

(35)
(25)
1134 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 30, NO. 3, MAY 2015

Fig. 1. Simple infinite system: the distributed parameter transmission line.

where the vectors and are the right and left eigenvectors
of associated with , respectively, and scaled such that
. Fig. 2. Percent error in the real part of the residue as a function of pole
accuracy (single transmission line example).
B. Distributed Parameter Transmission Line Model
The infinite system here considered to verify the numerical
performance of (35) consists exclusively of an isolated single the purposes of this numerical performance test, the difference
300 km, 500 kV transmission line, whose terminals are identi- between and can be written as
fied as buses 1 and 2, and whose positive sequence parameters
(41)
per unit length (longitudinal impedance) and (transversal
admittance) are depicted in Fig. 1. so that
Assuming there are injected currents and at both line
terminals, the nodal voltages at these terminals and will (42)
exist and the associated 2 2 nodal admittance matrix can
where . Note that the relative error in the computation
be defined as in [6].
of given by (42) is
The TF chosen to verify the numerical performance of (35) is
the transfer impedance between the two buses, , defined by
(43)
(36)

where and . The symbols and de- Substituting (42) in (35), yields
note the length, the propagation constant and the characteristic
admittance of the TL, respectively [6].

C. Analytical Expressions for the Line Poles and Residues (44)

The system poles and associated residues of can The detection of numerical errors from the use of (35) (or
be analytically determined as equivalently (44)) involved varying the exponent and calcu-
lating the percent error between (35) and the analytical solution
(37) (exact value) given by (38). The exact value for considering
the line parameters of this section and , is analytically ob-
tained from (37):

(45)
(38)
The exact value for the residue of pole is analytically ob-
where
tained from (38):
(39) (46)

(40) where the power and voltage basis used were 100 MVA and 500
kV, respectively.
The percent error of the real part of the residue
where symbols and denote the resistance, the induc-
produced by using (44), are shown in Fig. 2 as a function of
tance and the capacitance per unit length of the TL, respectively.
the accuracy of pole . The real part percent error is defined
D. Inaccurate Residue Values Computed From (35) by

Let be the exact pole value and the value obtained by %

one of the Newton algorithms described in [2] or in [5] and
used in (35) for the computation of the associated residue. For (47)
VARRICCHIO et al.: COMPUTATION OF DOMINANT POLES AND RESIDUE MATRICES 1135

The imaginary part percent error is defined in a similar

way but replacing the real operator by the imaginary operator
. The behavior of as a function of pole accuracy is very
similar to but is not shown here for the sake of brevity.
Note that , which is the abscissa of the green ver-
tical line in Fig. 2, is a commonly adopted relative error tol-
erance exponent (i.e., ) in numerical linear algebra applica-
tions. As one can observe in Fig. 2, the errors in the computation
of the associated residue are unacceptable. The same verifica-
tion was repeated for several other poles of this distributed pa-
rameter transmission line example, and the numerical errors ob-
served in the values of their associated residues were also found Fig. 3. (a) Pole . (b) Pole enclosed by a square contour.
to be unacceptable.
A numerically robust method for the accurate computation
Dividing the interval in subintervals, one has
of the residue associated with a pole of an infinite system is
presented in the next section.
(54)
V. ACCURATE POLE-RESIDUE COMPUTATION
The proposed method for the accurate computation of the
residue matrix , associated with the pole , is based on the where
evaluation of a contour integral [11]
(55)

(48) Substituting (54) into (53), yields

where is a contour in the s-plane enclosing the pole , which (56)

is pictured in Fig. 3(a).
The first step to the solution of the contour integral given by
(48) is to choose the contour . In this work, a square contour The values of the complex frequency in each subinterval is
enclosing the pole was chosen, as shown in Fig. 3(b). given by the following parametric straight line equation:
The complex numbers , shown in Fig. 3(b),
(57)
can be defined by
where .
(49) Differentiating (57) with respect to , one has

(58)
where is the length of half the square diagonal and is a
complex number of unity modulus given by In accordance with (57) for
and for . Substituting (57) and (58)
(50) into (56), yields
The points , shown in Fig. 3, can be written
as functions of the numbers , as (59)

(51)
where
Substituting (49) into (51), yields

(52) (60)
Having defined the contour as the square with vertices
, as shown in Fig. 3(b), the integral in (48) can The right-hand side integral in (59) can be efficiently and
be written as accurately evaluated using the Legendre-Gauss quadrature
method [12]. An integral over must be changed into
an integral over before applying the Legendre-Gauss
(53) quadrature rule. This change of integration intervals can be
carried out by defining the variable

where . (61)
1136 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 30, NO. 3, MAY 2015

In the case of the right-hand integral in (59), one has and

. Thus (61) is reduced to

(62)

Differentiating (62) with respect to , one has

(63)

Substituting (62) and (63) into (59), yields

(64)

where

(65)
Fig. 4. 34-bus test system with 25 distributed-parameter transmission lines.
Applying the Legendre-Gauss quadrature rule to the right-hand
integral in (64), one obtains error in the residue matrix computation is

(66) (72)

where is the number of points (abscissas) and weights The pseudo codes for the proposed SMDPA and for the
utilized in the summation. In this work, the choice was Legendre-Gauss integral solver scheme are included in
and the corresponding values for and are available Appendix A. Considerations on the computational complexity
in several standard handbooks including [12]. of the method are presented in Appendix B.
Substituting (66) into (64), yields
VI. TEST SYSTEMS

(67) The first test system is the transmission network shown in

Fig. 4 and having 34 buses, 25 distributed-parameter TLs, 12
shunt branches representing capacitor or inductor banks, 16
Substituting (67) into (48), one obtains transformers, 16 loads, and 10 generators. The complete data
for this system is described in [3]. The selected MIMO TF has
(68) the electrical currents simultaneously injected into buses 21
(input 1) and 23 (input 2) (both highlighted in red) as its inputs
and the voltages at these same buses as its outputs (the voltages
According with (68), the residue matrix associated with the
at buses 21 and 23 are output 1 and output 2, respectively).
pole , is equal to the sum of the four terms ,
The frequency window of interest for this case ranges from 0
given by
to 4000 Hz, which was verified to capture 95% of the energy
contained in this MIMO TF, as defined in [13].
(69) The second test system is the well-known IEEE 118-bus
system and the frequency window of interest now ranges from
Thus, the term represents the contribution of the side 0 to 10 kHz. This system has 118 buses, 177 TLs, 14 shunt
of the square contour to the total residue matrix, the term branches representing capacitor or inductor banks, 9 trans-
the contribution of the side and so on. In this work, formers, 91 loads, and 54 generators. The input vector entries
the number of subintervals in each side of the square of the chosen TF are the inject currents into buses 30 (input 1)
, is given by and 38 (input 2) and the voltages at these buses are the entries
of the output vector.
(70) The transformers, branches and loads of these test systems
were modeled by lumped parameters. The generators were mod-
The relative error in the computation of each one of the eled as ideal voltage sources and the TLs by distributed param-
four residue terms can be evaluated by eters using hyperbolic functions [2].

VII. APPLICATION RESULTS ON TWO TEST SYSTEMS

(71)
The proposed algorithm was applied for the calculation of
The iterative process converges when , where is the dominant poles and their residue matrices for MIMO TFs of
the maximum specified relative error. Note that the total relative the infinite systems described in Section VI. The objective was
VARRICCHIO et al.: COMPUTATION OF DOMINANT POLES AND RESIDUE MATRICES 1137

Fig. 5. Converged pole spectrum for the initial set of 37 estimates. Fig. 6. -plots for the Infinite and 151-ROM models.

to verify whether the algorithm would be reliable and efficient also indicated in the figure. Note that, by using the effective pole
in computing high-fidelity rational ROMs, over a pre-defined deflation technique adopted in this paper, all 37 initial estimates
frequency window, from these computed modal results. converged to distinct poles. Note also that once one pole of a
The sequential computation of poles was carried out in two complex-conjugate pair is computed, the complete pole pair is
stages, the first computing the set of most dominant poles and appropriately deflated in all subsequent computations.
the second those less dominant but yet important poles. The set In the second stage another 41 poles and associated
of initial estimates for the first stage are purely imaginary num- residue matrices were computed. Thus, a 151st-order model
bers whose values coincide with the frequencies of the peaks of (151-ROM) was obtained (73 complex conjugate pairs and
the maximum singular value plot, -plot, of . In the 5 real poles). The -plots of the infinite and the
occasional cases where 20 iterations are reached, the Newton reduced models are shown in Fig. 6 and seen to be
iterative process is either clearly divergent or nonconvergent. visually coincident over the 0 to 4 kHz window. The frequency
Once a pole has converged its associated residue matrix is responses (magnitude and phase) of the individual entries of the
computed. Then this converged pole is deflated. After the pole infinite and the reduced matrix models (not
deflation procedure or after the maximum number of iterations shown due to lack of space) were also in excellent agreement
has reached, the current software implementation automatically as one could expect since this fact is a necessary condition for
starts another Newton eigensolution process using the next the good agreement between the -plots.
available initial estimate for a pole. The deviation as a function of frequency given by (13) for
As regards the efficient computation of the set of less domi- the 151-ROM is shown in Fig. 9 (in a larger frequency window)
nant poles of the second stage, the Newton algorithm calls for a and in Fig. 10 (in the frequency window of interest), where it is
more refined strategy to select the initial estimates. This is be- compared with a 167th-order model.
cause these poles have a much smaller region of convergence
attraction [8] for the Newton algorithm. The pole estimate has
its imaginary part equal to the frequency in rad/s cor- B. 118-Bus Test System
responding to the maximum of the given by (13). The majority of the poles of the 151-ROM of the 34-bus test
The real part of this estimate is however made equal to the real system (37 complex conjugate pairs of dominant poles com-
part of the complex frequency where the maximum of puted in the first stage and 36 pairs plus 5 real dominant poles
the graph occurs (note that in in the second stage) had their imaginary parts confined to the 0
this graph is fixed and varies). Then this estimate to 4 kHz frequency window. Only 11 complex conjugate pairs
is fed to the proposed SMDPA algorithm for the pole and as- of dominant poles (14%) had their imaginary parts out of this
sociated residue matrix computations. Next the converged pole window.
is deflated and the above described computational procedure is By contrast, the set of dominant poles for the 216-ROM (108
repeated. pairs of complex conjugate poles) of the 118-bus test system
Convergence tolerances equal to in the relative errors have their imaginary parts spread over the 0 to 20 kHz range,
of the poles and associated residue matrices and for the despite the study focus being limited to the 0 to 10 kHz window.
length of half the square diagonal in Fig. 3(b), were used in all Actually, nearly half of the dominant poles (51 complex conju-
computations. gate pairs) have their imaginary values ranging from 10 to 20
kHz. In the first stage, only two pole estimates out of 104 have
A. 34-Bus Test System
not led to a convergent process. All other 102 estimates have
An initial set of purely imaginary estimates were obtained converged to distinct poles, 87% of those requiring only 4 or 5
in the first computation stage. Fig. 5 shows the pole spectrum iterations for convergence. It is also worth mentioning that the
computed by SMDPA for this initial set of 37 estimates. The 118-bus test system MIMO TF had only 6 other dominant pole
number of iterations required for the convergence of each pole is pairs computed in the second stage, where use was made of the
1138 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 30, NO. 3, MAY 2015

where . Note that (75) is analogous to (4) and

that is the linear descriptor system model approxima-
tion of the infinite system model .
Similarly to the ROM MIMO TF deviation defined in (12),
the difference between and is defined as the linear
approximation MIMO TF deviation:

(76)

The defined in (13) determines the ROM de-

viation as a function of frequency, over the given frequency
window. In this section it is considered a model error measure
that is completely described by a single numerical value. This
Fig. 7. -plots for the Infinite and 216-ROM models (118-bus test system).
measure [14] is the ratio between the area of the ROM deviation
curve and the corresponding area of the infinite model:

% (77)

The corresponding LAM percent error is, obviously,

defined similarly to .
The distributed parameter TLs of the 34-bus test system
shown in Fig. 4 were next modeled by LAMs of increasing
level of detail. Table I details some data about the 4 LAMs
that were generated by discretizing the system TLs from a
minimum of 300 to a maximum of 600 cascaded RLC-PI
sections. In this table , and denote the number of
RLC-PI circuits per line, the number of differential equations
Fig. 8. Deviation of the 216-ROM over the frequency window of interest and the dimensions of the matrices and . The ratio / ,
(118-bus test system).
where is the dimension of the matrix model of
the infinite system (equal to the number of system buses added
by the number of generators), is also included. The number of
more elaborate initialization strategy described in the beginning algebraic equations is the same for the four LAMs
of this section. and equal to 119.
The -plots of the infinite and the reduced The error performances of the four LAMs are compared in
models are compared in Fig. 7 and seen to be visually Table II with that of the 151-ROM obtained by the proposed
coincident over the 0 to 10 kHz window. The deviation plot as SMDPA method (cf. Section VII-A). A LAM with value be-
a function of frequency, defined by (13), is shown in Fig. 8 for tween 400 and 500 is seen to have an error performance equiva-
the 216-ROM. lent to that of the 151-ROM. From Table I the matrix dimension
of this equivalent LAM is between 20182 and 25182.
VIII. INFINITE X FINITE SYSTEMS Utilizing high-fidelity LAMs, with model errors close to that
of the 151-ROM (directly derived from the infinite system), is
A linear (finite) approximation model (LAM) of an infinite therefore not practical for large systems and wide frequency
system can be obtained using, for instance, a descriptor system windows, due to the increasingly ill-conditioning nature of the
model, where each TL is represented by a sequence of RLC-PI PI-cascading process, combined with the usual computational
circuits [3]. difficulties associated with large matrices such as huge memory
The descriptor formulation consists of first order differential storage and high CPU time. By contrast, the accuracy of the pro-
and algebraic equations (DAEs): posed 151-ROM can easily and efficiently be further improved
by just adding a few more poles and associated residue matrices.
(73) The curve of the 151-ROM is plotted in Fig. 9 in
(74) a larger frequency window. The eight higher peaks that appear
in the 4 to 8 kHz range of this plot are due to the existence of
where eight dominant poles (i.e., 8 pairs of complex conjugated poles).
and . In this linear Computing these poles and adding them to the 151-ROM, yields
finite approximation, the TF is defined as a 167-order model that shows a significant accuracy improve-
ment over the entire frequency window of interest. The devia-
(75) tion curves of the two ROMs in the frequency window of in-
VARRICCHIO et al.: COMPUTATION OF DOMINANT POLES AND RESIDUE MATRICES 1139

TABLE I
FINITE SYSTEM INFORMATION (34-BUS TEST SYSTEM)

TABLE II
COMPARING MODEL PERFORMANCES (LAM 151-ROM)

Fig. 9. Deviation of the 151-ROM over a larger frequency window.

of Table I (34-bus test system with the TLs modeled by 400

RLC-PI circuits). All three methods were run to compute a
set of the 37 most dominant poles and their associated residue
matrices. The resulting pole spectra computed by the three
methods (SMDPA, SAMDP, and SLRCF) are depicted in
Fig. 11.
Note that out of the 37 infinite-system poles computed by the
SMDPA, 29 have been approximated by the poles obtained by
Fig. 10. Deviations of the 151-ROM and 167-ROM over the frequency the SAMDP for the 20063-order finite system. Four of the poles
window of interest. computed by the SAMDP (
and ) and
one computed by the SLRCF cannot be
shown in Fig. 11 since the horizontal scale adopted for this plot
was optimized for a visual comparison of the matching between
the infinite-system and finite-system poles. One should note that
the poles computed by the SMDPA and SAMDP are true system
poles while the poles computed by the SLRCF are just values
that yield a good match between the responses of the complete
and reduced modes. The software version was Matlab R2010a
and the PC processor was an AMD Phenom 3.20 GHz with 4
GB RAM, 32 bits. The CPU times required by these methods
to compute these sets of 37 poles and their associated residue
matrices are presented in Table III. From this table one can see
that the proposed SMDPA was approximately four and twelve
Fig. 11. Pole spectra obtained by the SMDPA, SAMDP and SLRCF when con- times faster than the SAMDP and SLRCF, respectively, despite
sidering the first 37 dominant poles (infinite model and 20063-order LAM for
the 34-bus test system).
the fact that the two latter methods benefit from much more
mature computer codes.

IX. CONCLUSIONS
terest are shown in Fig. 10. The of the 167-ROM is equal
to %, which is smaller than the of the LAM This paper described the first Newton algorithm for the
with 600 RLC-PI circuits per transmission line (cf. Table II). reliable computation of scalar and MIMO TF dominant poles
It is impossible to rigorously compare the proposed method's (SMDPA) of infinite systems. Its reliable numerical perfor-
computational efficiency against other methods, since there mance is ensured by an efficient deflation procedure [5] which
are no other eigensolution methods that effectively compute prevents repeated convergence to the previously found transfer
the actual poles of an infinite system. The chosen alternative function dominant poles. The numerically efficient and robust
was therefore to compare the SMDPA method operating on implementation of this deflation procedure can only be achieved
the infinite system with the SAMDP [4] and the SLRCF [7], with the accurate computation of the transfer function residue
with the latter two methods operating on the 20063-order LAM matrices for the converged poles. As proposed in this paper, the
1140 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 30, NO. 3, MAY 2015

TABLE III APPENDIX A

CPU TIME PERFORMANCES PSEUDO-CODES FOR THE ALGORITHMS
(INFINITE AND 20063-ORDER MODELS
FOR THE 34-BUS TEST SYSTEM)
Algorithm 1: Sequential MIMO Dominant Pole Algorithm

INPUT: Complete data for , initial pole

estimates and (maximum
number of iterations).
OUTPUT: Dominant poles and their associated residue
residue matrices should be computed by robust integral solver matrices.
schemes such as the Legendre-Gauss quadrature. The proposed 1) Assign a very large value to
method was successfully applied to MIMO TF's model of two 2) Compute and its derivative
test systems, the 34-bus and IEEE 118-bus systems, having 25 3) Compute using (4)
and 177 transmission lines (TLs) with distributed parameters, 4) Solve (31) and (32) for and , respectively
respectively. 5) Compute using (30)
The multivariable SMDPA defaults to the SDPA [5] for scalar 6) Compute and using (7) and next (8) and (6)
transfer functions of infinite systems. The computation of accu- 7) Set
rate residues by the Legendre-Gauss integral solver scheme, the 8) for to do
first main contribution of this paper, has therefore conferred re- 9) Set
liability to the original SDPA [5], lending it suitable for practical 10) while ( and ( do
engineering applications. 11) Compute and its derivative using (11)
The CPU time taken to compute a set of a MIMO TF and (27), respectively, with poles and
dominant poles (37 pairs of complex conjugate poles) and associated residue matrices and complex conjugate
associated residue matrices of the 34-bus infinite test system ones.
is approximately four and twelve times smaller than the corre- 12) Compute and its derivative
sponding times taken by the widely used Subspace Accelerated 13) Compute using (4)
MIMO Dominant Pole Algorithm (SAMDP) [4] and Sparse 14) Compute using (12) and its inverse
Low Rank Choleski Factor Method (SLRCF) [7], respectively, 15) Compute the eigentriplet
when operating on a 20063-order LAM of roughly equivalent of using
fidelity. the QR algorithm
The described results for infinite systems were mostly fo- 16) Solve (31) and (32) for and ,
cused on the high-quality MIMO ROMs that can be produced respectively
for power transmission networks by the proposed SMDPA. 17) Compute using (30)
Such model order reduction ability in a sparse eigensolu- 18) Compute using (26)
tion method that directly operates on infinite systems is of 19) Compute the pole correction using (25)
much practical value to engineering studies. There are very 20) Compute the new pole estimate using
successful algorithms for finding approximations to the dom- (24)
inant poles and associated residue matrices from fitting fre- 21) Compute the relative error
quency response data [15], [16], obtained either from model 22) Set
simulations or field measurements. Nevertheless, the SMDPA 23) end while
algorithm has fundamental differences from these fitted re- 24) if then
sponse methods, as it computes the true poles and residues 25) Set
directly from the infinite system model, rather than approxi- 26) Store the converged pole value and calculate
mated poles from fitted frequency response data. This opens the associated residue matrix {Alg. 2}
the possibility for additional uses of the SMDPA algorithm, 27) end if
such as modal sensitivity studies and root locus analysis of 28) end for
infinite system models.
A large set of TF dominant poles was computed for the Algorithm 2: Residue Matrix Calculation
two test systems and the -plot responses for the resulting
ROMs were visually coincident with those of the infinite INPUT: Converged pole value , half diagonal , the
models over the frequency windows of interest. The SMDPA points and weights for the application
method can be applied to other areas of engineering, physics of the Legendre-Gauss rule, and (maximum
and mathematics, since its mathematical development does number of subintervals).
not exploit any characteristic that is specific to electrical net- OUTPUT: Residue matrix associated with .
works or power systems. The SMDPA method is expected to 1) Compute the complex number using (50)
advance the research and applications of eigenanalysis for in- 2) Compute the square vertices using
finite systems. (52) and set
VARRICCHIO et al.: COMPUTATION OF DOMINANT POLES AND RESIDUE MATRICES 1141

3) Compute using (62) for TABLE IV

4) for to 4 NUMERICAL EXPERIMENTS FOR COMPLEXITY ANALYSIS

5) Set
6) while ( and do
7) Set
8) Compute using (55)
9) for to do
10) Set
11) Compute the complex frequency values over
the straight line connecting the points
and as
12) Compute matrix
13) Compute using (4)
14) Compute the integral of over the
subinterval defined by the extreme points spends only 3.7% of the total CPU time to compute a set of 102
and using the Legendre-Gauss dominant poles. Most of the total CPU time, in all test systems,
is therefore related to the execution of Alg. 2, for the accurate
quadrature rule:
computation of the residue matrices for the MIMO TF.
15) Set Getting into the Alg. 2, the most CPU time expensive opera-
16) end for tion is the assembling of the matrix following by the com-
17) Compute term of the residue matrix: putation of its factors and solver. The assembling of the
matrix for infinite systems requires a significant amount of CPU
time, mainly due to the need to compute hyperbolic functions
for many complex argument values. This CPU time, , is
18) if then shown in Table IV for several electrical networks whose di-
19) Compute mensions are denoted by . The percentage of CPU time re-
20) end if quired for the LU factorization (Fat) of the matrix and
21) Set solver in relation to and the percentage of nonzeros in
22) end while (nn) are also presented in this table. The symbol pre-
23) if then sented in the last column of this table relates to the computa-
24) Store the converged residue matrix term tional complexity associated with computations, as ex-
25) else plained in the following.
26) Increase the value of the half diagonal and restart Assuming the CPU time to be a power function of the
the algorithm. Alternatively, the number of points problem dimension
of the Legendre-Gauss method may be changed.
(78)
27) end if
28) end for the computational complexity [17] for a particular task or
29) Compute the sum of the four terms and store method of any dimension, can be estimated once the exponent
the residue matrix. is known. This assumption is used to derive that this exponent
is approximately equal to
APPENDIX B
CONSIDERATIONS ON THE SMDPA COMPUTATIONAL (79)
COMPLEXITY
The proposed SMDPA involves the recursive and interwoven This computation should be carried out for several matrix ex-
executions of the two algorithms described in Appendix A: the amples of different dimensions to verify whether the obtained
one-pole-at-a-time Newton algorithm (Alg. 1) run, with its em- value of is approximately constant. The value of
bedded pole deflation procedure, is immediately followed by the should be significantly larger than in order to ensure that
run of Alg. 2 to compute the TF residue matrix for this pole by even for large variations in the variable the exponent will
the Legendre-Gauss integral solver scheme. At the end of every remain constant. In [17], it is suggested that, at least, one should
recursive step, the newly computed pole and associated MIMO have . The electrical networks in Table IV cor-
TF matrix residue are incorporated into the set of results for use respond to actual areas of the Brazilian electrical power system
in all subsequent steps of the recursive process so that pole de- (the last one is the full-size Brazilian system model), which ap-
flation is effectively implemented. proximately abide to the recommendations of [17], with
For the 34-bus test system, Alg. 1 spends only 7.0% of the in all experiments. From the results shown in
total CPU time to compute a set of 37 most dominant poles, Table IV, the exponent is seen to be roughly equal to 1. Thus
where convergence to most of these poles occurred within 7 the computational complexity of assembling the matrix is
iterations (cf. Fig. 5). For the IEEE 118-bus test system, Alg. 1 .
1142 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 30, NO. 3, MAY 2015

REFERENCES [16] T. Noda, “Identification of a multiphase network equivalent for electro-

[1] A. Semlyen, “S-domain methodology for assessing the small signal sta- magnetic transient calculations using partitioned frequency response,”
bility of complex systems in non-sinusoidal steady state,” IEEE Trans. IEEE Trans. Power Del., vol. 20, no. 2, pp. 1134–1142, Apr. 2005.
Power Syst., vol. 6, no. 1, pp. 132–137, Feb. 1999. [17] R. J. Lipton, The Question and Gödel's Lost Letter. New
[2] S. Gomes, Jr., N. Martins, and C. Portela, “Modal analysis applied to York, NY, USA: Springer, 2010.
s-domain models of ac networks,” in Proc. IEEE PES Winter Meeting,
Columbus, OH, USA, Jan. 2001.
[3] F. D. Freitas, N. Martins, S. L. Varricchio, J. Rommes, and F. C.
Véliz, “Reduced-order transfer matrices from RLC network descriptor
models of electric power grids,” IEEE Trans. Power Syst., vol. 26, no. Sergio Luis Varricchio (SM'06) received the B.Sc. degree from the Catholic
4, pp. 1905–1916, Nov. 2011. University of Petrópolis (UCP), Petrópolis, Brazil, in 1987 and the M.Sc. degree
[4] J. Rommes and N. Martins, “Efficient computation of multivariable from the Federal University of Rio de Janeiro (UFRJ), Rio de Janeiro, Brazil, in
transfer function dominant poles using subspace acceleration,” IEEE 1994, both in electrical engineering. Currently he is also pursuing the Ph.D. de-
Trans. Power Syst., vol. 21, no. 4, pp. 1471–1483, Nov. 2006. gree in electrical engineering at the University of Brası́lia, Brası́lia, DF, Brazil,
[5] S. Gomes, Jr., N. Martins, and C. Portela, “Sequential computation of in the area of model order reduction.
transfer function dominant poles of s-domain system models,” IEEE Since 1989, he has worked for CEPEL, Rio de Janeiro, as a project manager
Trans. Power Syst., vol. 24, no. 2, pp. 776–784, May 2009. and as a researcher in the areas of modal analysis, power quality, and electro-
[6] A. Semlyen and M. H. Abdel-Rahman, “Transmission line modelling magnetic transients.
by rational transfer functions,” IEEE Trans. Power App. Syst., vol.
PAS-101, no. 9, pp. 3576–3584, Sep. 1982.
[7] F. D. Freitas, J. Rommes, and N. Martins, “Gramian-based reduction
method applied to large sparse power system descriptor models,” IEEE Francisco Damasceno Freitas (SM'08) received the B.Sc. and M.Sc. degrees
Trans. Power Syst., vol. 23, no. 3, pp. 1258–1270, Aug. 2008. from the University of Brasilia, Brasilia, DF, Brazil, in 1985 and 1987, respec-
[8] J. Rommes and G. L. G. Sleijpen, “Convergence of the dominant pole tively, and the Ph.D. degree from the Federal University of Santa Catarina, Flo-
algorithm and Rayleigh quotient iteration,” SIAM J. Matrix Anal. Appl., rianopolis, SC, Brazil, in 1995, all in electrical engineering.
vol. 30, no. 1, pp. 346–363, 2008. Since 1986, he has worked for the University of Brasilia, where he is an As-
[9] D. V. Murthy and R. T. Haftka, “Derivatives of eigenvalues and eigen- sociate Professor. His area of interest is power system dynamics.
vectors of a general complex matrix,” Int. J. Numer. Meth. Eng., vol.
26, pp. 293–311, 1988.
[10] B. Pascual and S. Adhikari, “Dynamic response of structures
with frequency dependent damping models,” in Proc. 49th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Nelson Martins (F'98) received the B.Sc. degree from the University of
Materials Conf., Schaumburg, IL, USA, 2008. Brasilia, Brasilia, DF, Brazil, in 1972 and the M.Sc. and Ph.D. degrees from the
[11] E. Kreyszig, Advanced Engineering Mathematics. Hoboken, NJ, University of Manchester, Manchester, U. K., in 1974 and 1978, respectively,
USA: Wiley, 2011. all in electrical engineering.
[12] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Func- He has worked for CEPEL, the Brazilian Electrical Energy Research Center,
tions: With Formulas, Graphs, and Mathematical Tables. New York, Rio de Janeiro, Brazil, since 1978, developing methods and tools for power
NY, USA: Dover Books on Mathematics, 1970. system dynamics and control.
[13] S. L. Varricchio, F. D. Freitas, and N. Martins, “Spectral energy indices
for model order reduction,” in Proc. IEEE PES General Meeting, Van-
couver, BC, Canada, Jul. 21–25, 2013.
[14] S. Boyd and C. Barratt, Linear Controller Design: Limits of Perfor- Franklin Clement Véliz received the B.Sc. and M.Sc. degrees in electrical en-
mance. Englewood Cliffs, NJ, USA: Prentice-Hall, 1991. gineering from the Federal University of Rio de Janeiro (UFRJ), Rio de Janeiro,
[15] B. Gustavsen and A. Semlyen, “Rational approximation of frequency Brazil, in 2001 and 2005, respectively.
domain responses by Vector Fitting,” IEEE Trans. Power Del., vol. 14, He works for Pontific Catholic University (PUC), Rio de Janeiro, developing
no. 3, pp. 1052–1061, Jul. 1999. methods and tools for power system analysis.