3126 IEEE Transactions on Power Apparatus and Systems, Vol. PAS-lOl, No.

9 September 1982

SELECTIVE MODAL ANALYSIS WITH APPLICATIONS TO ELECTRIC POWER SYSTEMS,

PART II: THE DYNAMIC STABILITY PROBLEM

G. C. Verghese 1. J. Perez-Arriaga F. C. Schweppe

IEEE Member IEEE Member IEEE Fellow

Electric Power Systems Engineering Laboratory

Massachusetts Institute of Technology
Cambridge, Massachusetts 02139

Abstract - Selective Modal Analysis (or SMA) is number of troublesome oscillating modes is usually
a physically motivated framework for understanding, a small fraction of the very many modes of behavior
simplifying and analyzing complicated linear time- of the system; often, just one mode is of interest.
invariant (or LTI) models of dynamic systems [1,2,3]. h o
SMA allows one to focus on any prespecified dynamic nw e und oo toe the ex ite ofoflanope
now well understood to be the existence an ope-
pattern of interest
pattrn o in te
intrestin the moel. particular,'
model. Inn paticuar, rain codto thti.ntbentesal
. .
One can eff iciently and accurately compute the ~~~~~~~~rating
condition that Is unstable in the small.
oigenevcanuefficent agendvaccuraofthel
cpute
eigenvalues and eigenvectors of the natural modes tdes The
de onset o ofemt nov power
spontaneous olna oscillations
systemeairi n
of interest and their sensitivities, and also deter- dessnot way, as shown in[ ea smaviprtur-
mine physically meaningful reduced order models con- bation
bation bavo
behavior of thower
of theconditionsyte ite in the vicinity
vicini
tamingt thesefnatural
tainng mode. sMA
teseaturl moes. iclr
MA iis pariculrly of a given operating power system
can therefore be ty
suitable for dealing with compoit models,e. described to first order by a set of LTI differential
model s cons isting of several dynamic subsystems eutin ntesaesaefr
interrelated by static constraints. An introduction q
to the basic concepts of SMA pertinent to the appli- x = Ax + Bu (1)
cations being considered here is presented in the
companion paper [3]. This paper concerns the appli- where the state vector x represents the perturbation
cation of SMA to the Dynamic Stability problem in of the system state from its nominal value at the
electric power systems; it is shown how SMA is well chosen operating condition, and the vector u repre-
suited to meet the demanding requirements of Dynamic sents perturbations of "inputs" such as load power
Stability analysis. This is illustrated in [3] with demands. The numerical values of the constant
examples, including a 60-machine model of a dynamic matrices A and B depend on the operating condition
instability occurrence in an actual power system. as well as on the system parameters. Describing
system dynamics in the vicinity of the specified
1. DYNAMIC STABILITY: INTRODUCTION AND REVIEW operatinc condition now reduces to the conceptually
simple task of evaluating the modal structure of
(1), i.e., the eigenstructure of the matrix A. The
A paradigm for many of the large-scale issues procedure now commonly followed involves computa-
involved in dealing with electric power systems is tion of the complete eigenstructure of A, and of
provided by the so-called Dynamic Stability problen. its variations with changes in selected system
Dynamic Stability concerns the dynamic behavior of conditions and parameters. Studies have repeatedly
the power system under small perturbations about shown the adequacy of model (1) in providing
some operating condition, and more specifically the reasonable qualitative and quantitative understand-
phenomenon of slow and poorly damped, or sustained, ing of observed incidents of dynamic instability;
or even diverging power oscillations. The oscilla- reference [5] is a recent example.
tions may occur between one or a few machines and A consiZderabl e amount of effort has been ex-
the rest of the system ("local oscillations"), or pended in research work and on-site studies for
between large groups of machines ("generalized t
tackl ing the complex problems of tuning voltage
oscillations"); the latter typically occur in theto exciters to
0.1 to 0.5 Hz range, the former in the 1 to 2 Hz stabilize achine systems th efforth
range. An important point to be noted is that the succeeded nustabiliing man Trbsoefpor
succeeded i n stabili|zi|ng many troubl esome-power
system configurations and has also resulted in a
fairly sound body of theoretical and practical
knowledge, see [6]. In theory, the eigenstructure
and sensitivity information can be made the basis
for choosinq proper operating conditions and system
82 WM 071-9 A paper recommended and approved by the parameters, see [7]. In practice, however, most of
IEEE Power System Engineering Commnittee of the IEEE these approaches to power system stabiliTzation via
Power Engineering Society for presentation at the IEEE excitation control, for controllers currently being
PE5v1982 Winter Meeting, New York, New York, January 31- imzplemented or in operation, are l imited to rather
February 5, 1982. Manuscript submitted August 31, 1981; crude models of either the power system external
made available for printing November 4, 19.81. to the power plant being control led (and therefore

0018-9510/82/0900-3126$00.75 ) 1982 IEEE


lack coordination) or of the entire system (thus
lacking accuracy). They also lack adaptability to
changing operati'ng conditions. These shortcomings
stand in the way of a truly comprehensive approach
to the Dynamic Stability problem. This has been
empirically shown by the reoccurrence of dynamic
instability in systems already heavily equipped
with stabilizers, as'documented in [8]. More ad-
vanced approaches are mostly of an academic nature,
see [9].
Any comprehensive practical approach to the
Dynamic Stability problem has to deal with the in-
escapable size and complexity of the system (1) in
any realistic model. An accurate study of Dynamic
Stab i Ility requ i res initial modeli ng of exten-
sive portions of the system being studied (because
generalized oscillations are not physically localized)
and of the many disparate system components and pro-
cesses that appear significantly involved. As a
result, the matrix A in (1) can easily have dimen-
sions in the thousands, and determination of the
eigenstructure, even as an off-line task at the
planning level, involves formidable difficulty and
cost. (The standard eigenanalysis algorithm needs
of the order of n3 operations when applied to an
n x n matrix.)
1.2 Review of Some Prior Work''
A careful analysis of the abundant literature
on this topic reveals that the current state-of-the-
art falls short of providing a comprehensive approach
to the Dynamic Stabi l ity problem. However, two
fairly recent approaches will now be discussed in
more detail since they have some points of contact
with the method to be presented' in-this paper, namely
Selective Modal Analysis (or SMA). The method in [4]
exploits knowledge of the frequency range of interest ~~~~~1.
(0.1 Hz to 2 Hz), unlike. the complete eigenstructure
.
computation. A frequency-doma'in model of the linear-
ized system is used to calculate the speed pertur-
bation phasor at a selected generator (one believed - work for accte
significantly involved in the mode- of interest) when
the generator is excited by an exponential . pertur-
bation torque at. a- complex
. frequency; the torque-
to-speed transfer function is modeled' as a second-
-is
order system whose poles are then chosen to match
the calculated speed response to the torque excita-
tion. The calculated locations may be used to
. . pole
> concizsely:Vrev
update the excitatilon frequency and move closer to .
a resonance condition, thereby iteratively improving
estimates of the modes. The mode patterns or eigen-
vectors are estimated from the relative magnitudes
of system variables at the resonance. The method
is reasonable but quite ad hoc in several respects:
it does not guarantee that the modes of interest
will be identified; its accuracy is uncertain; sen-
sitivity calculations in this framework would appear
to require repeated solution; and the formulation
itself is rather specialized. The emphasis in [4]
on concentrating the analysis on a portion of the
complete dynamic spectrum and on the associated
model structure is however germane to the philosophy
underlying SMA.
In [10] the linearized power system model is
formulated in the frequency domain and rearranged
so that i ts structure is that of the so-cal led
classical electromechanical model (swing eqs. and
constant-voltage-behind-reactance models for the
3127
synchronous machines), with functions of the Laplace
operator s appearing in the place of constant co-
efficients. The resul'tant algebraic equation in s
is then solved iteratively to yield the desired
eigenvalue. The approach is interesting, and bears
some relationship to SMA. Several objections, how-
ever, may be raised. The iterative technique can
only deal with one eigenvalue at a timne, which leads
to difficulties in the iterative algorithm when
tracking eigenvalues in clustered groups. The struc-
ture of the reduced model is constrained to be that
of the classical electromechanical model; it will be
seen later that this will not usually be adequate
for dynamically unstable modes, thereby hindering
the convergence of the iterative algorithm. It is
not clear from the paper whether the authors have
an efficient procedure to compute their reduced
operational transfer matrix. If they do not, the
computational savings might be wiped out by the need
for updating this matrix for each eigenvalue and
iteration. No theoretical basis (convergence con-
ditions, etc.) for the method is presented, which
make's this method also appear heuristic and ad hoc.
It is bel ieved that SMA provides a conceptually and
computationally sounder and more comprehensive
framework than the above' mentioned approaches.
1.3 Organization of the Paper
*
The remainder of the paper is organized as fol-
lows. The SMA theory pertinent to Dynamic Stability
analysis is briefly reviewed in Section 2, including
a detailed discussion of composite models. Section 3
concerns the application of SMA to Dynamic Stability.
Reference is made to specific examples, in particular
to the results of the analysis of the realistic
177th order model described more fully in the com-
panion paper [31.
II SELECTIVE MODAL ANALYSIS: A BRIEF INTRODUCTION
2.1 Summary of Basic Concepts
S M
efficient and physicall-ased
m aaccurate, eff ici ent and phys ical ly-based
~~~~modeling and anlysis of selected l portions.of
- the
structure and behavior of 1linear time'-invariant
(
~~~~~~~(orLTI) systems. A heuristic introduction to SM1A
presented in the companion paper 13]'. Referenc'es
[1,2] offer Here a rigorous and far material
more complete
in [3] des-
cription.
b c r the pertinent will
~~~~~~~~be iewed and add;it ional1 resul1ts' con-
cerning
c composite models
w wi1l 1be presented.,
b pend
Consider the problem of accurate determination
of the eigenstructure (i.e. eigenvalues, eigenvec-
tors and their sensitivities with respect to system
parameters) of a group of h selected modes (or SM)
of the undriven N-th order system
* A (2)
that are primarily associated with a dynamic pattern
of interest.Number these as the first h modes.
Let the state x in (2), after possible permu-
tation of variables, be partitioned as
F All A1 rr]
t~1A
z = F 2(
A- z
LziL21 22j z
where r is a subset of n state variables (n.>h) ,
the relevant states, that is considered to be
3128

significantly involved in the SM, and z is the re-

maining subset of ir=N-n less relevant states. 9Q;vV ei (
- ~~~~~~r(t)
v2 ,(9)
The association of modes with states may be i-l
heuristically established from a basic understanding
of the system. This relationship, however, can be where vri is the partition of v. correspohding to
stated precisely by means of the participation matrix the r states, and 2T, i=l,..., h are arbitrary
P, see [11,12], defined as constants, can be exactly reproduced by a constant
matrix M that satisfies the linear system of equa-
P
=lPki= {ki vkil (4) tions
where Vki (wk.) is the k-th entry of the i-th right M[vrl ..
Vrh]= [H l)Vrl*H(Xh)VrhI (10)
(left) eigenvector v. (w.) of A, defined by The resulting reduced order model (or ROM)
AvY = v.A. ' v; *O , (5) ir(t) = (A 1 + M) =ARF(t) , (11)

wTA = A.w. , wT0O

'
(6) where r represents an approximation
contains the eigenvalues and right eigenvectors of
to r, exactly

where A. is the i-th eigenvalue; the eigenvectors the hSM. Other properties of the ROM (11) are:
are normalized such that (i) It is not affected by any variable trans-
T formation (of the original system) of the form
W.V. = if i= j ; 0 otherwise. (7) rr
(Superscript T denotes the transpose of the asso- T (12)
ciated vector or matrix.) L LTa Tbj LzJ
The entries of P are termed the participation
factors of the system and are dimensionless (i.e.
where Tbb is nonsingular.
independent of the choice of units);- . the magnitude
. . (I I)0 The model ordet reduction process may be
of p . precisely measures the relative participation carried out so as not to alter any relationship in
of of
tkt k-th state
kth
e variable in
stat vaial in the i-th mode,.....and
th'e i.t mode, and the original model (2) that involves only r states,
thus preserving this part of the original system
vice versa. The participation factors can be used
to define the subset r in (3), as will be explained
sthustres
structure.
shortly. (iii) The system matrix A +M is real if any
Note that (3) defines a distinction' between complex mode in the h SM is accompanied by its con-
relevant dynamics and less relevant dynamics in (1), jugate.
as illustrated in Fig. 1. Assume temporarily that (iv) When h= I an obvious solution to (10) is
the eigenstructure of the SM is known. The follow-
ing procedure singles out the relevant dynamics in M . H(A) (13)
a direct manner, while the less relevant dynamics is yIelding=
collapsed in a way that leaves the eigenstructure r A(A)r = [A
+ H(A)]F ' (14)
of the h SM intact. R 11
which can be alterhatively formulated as

X A A (15)

A I
This corresponds, in a detailed block diagram re-
presentation of the system, to "freezing" the less
RELEVA T DYNAMICS _ relevant portions of the block diagram at their
constant values for s=A, where s is the Laplace
A 12 S A 21 I operator. This property may also allow one to de-
termine A R (X) in symbol ic form based on physical
I i |T grounds, as wi 1 1 be seen later.
When A, vri, i=1,... h in (10) are not
|I known a priori, a natural iterative procedure that
LE$S RELEVANT DYNAMICS I is suggested by (10) can be used to determine the
L._1 ___ _ _ _ _ __ _ __ _ _JSM and (11). At the j-th stage one performs an
eigenanalysis of All + M. where i+lM is a solution
Fig. 1 The Relevant and Less Relevant Dynamics of
First note that the effect of the less rele- M[vrl.. h]-[H(JAl) .vlu.H(A;h) ivrh] (16)
vant dynamics. transfer function an o~0. =,.,hcnb bandfo
H(s 12 22s A A2 (8) eigenanalysis of Al1.
If convergent, this algorithm
must converge to h ei genval1ues/vectors of A.
for input signals containing only the h SM
 3129

For h= 1 and when (13) is chosen, the local where x2., uq9 and y2 are column vectors of dimensions
ratio of convergence p. of the algorithm is given by N2, pg and respectively for Q= 1,2,... ,L.
n
Pki -T
x x XrXp7 =Av +3
X.____ k WV. (7 i-th DYNAMIC SUBSYSTEM
k
lim j+ it Bu
Jx
X ki zinzm y1 :Cj xi + Di uL
k=n+l k
(where the notation should by now be clear). Pi is
called the participation ratio of the r states in the Ye Ut =1,2 ...,L
i-th mode. The algorithm thus converges locally if
and only if IPl l>. STATIC NETWORK: u -
Jy
The role of the participation factors in de-
fining the relevant states r when h= 1 is thus clear:
a good choice of r is one for which IPi is large. Fig. 2 Composite Model
For h> I it has been heuristically inferred from
experience in the application of the algorithm that The composite formulation (20) contains more
the participation ratios still provide useful indi- structural information than the associated global
catiTons of which states a re relevant : good conver- model (2), and SMA can take advantage of this addi-
gence properties are attained when (i) the partici- tional information. Once the choice of relevant
pation ratios for at least the n modes of A that states has been made as usual, (20) becomes
are closest to the eigenstructure of A1 are large 1 A A r B
(|p|>>l); (ii) the participation ratios for the 2 112 122 Q 9
remaining modes are small (|pI<<1); a gap region, = j.
not necessarily strictly empty, between both groups Z A2 A22
of values of IPI should be apparent; (iii) all the _ j
h SM belong to the first group, with none of them
falling in the above gap.
It must be noted that complete eigenanalysis of r
the full system matrix A is required for determining C _
the participation matrix P. Obviously one cannot Y 2 t 9(
always count on a perfect knowledge of P for select- 1 J 2a
ing the relevant states r in practical applications
of SMA. Section 3.2 elaborates on this point and
shows how the selection of the r states can be made where x= z and r2. and z are column vectors of
i n Dynam ic StabilI i ty modelIs. _
dimensions n2. and N2-n2 respectively, for 2= 1,2,..
The sensitivity aTi of the eigenvalue Xi with ..L. Here the distinction between relevant and
respect to a parameter Tn in A can be computed as less relevant states is explicit at subsystem level.
The associated global formuation can be directly
3' _ - P written from (21) as
i an = P; - OTil 8)
r
;
where
- A T -:1
cz. w [A (n)+H(Xi,n)Iv.I
(19)
and has the interpretation of being the sensitivity A +B1 (IJDd)'JCl A12+B1 (JDd 2
computed from the ROM (14) while ignoring the direct d d d d d d
dependence on n of the X in H(A). +B (I-JD 1JJ2I
AA21 ++B 2(I-JDd') 1JC AA22+B2(Jd) Z
Li (22)
2.2 Composite Models d d d dd
Many large physical systems can be formulated
in a natural way as a collection of L dynamic sub- where [ AIA diag{[.] }, rT= [rrT.. rTI
systems (see Fig. 2).
=AX + B u and T [z{.1.zL
2.2. 2.2.1
2.= 1,2, . . . ,L, (20a) The composite formulation can be exploited to
y = C x2 +D u2. ease the computation of H(X) for a given X. This
calculation is repeatedly needed in the iterative
that are interconnected by static constraints algorithm (16). Model order reduction can be per-
formed at subsystem level, yielding the reduced
1t IYll ~~~~~~~composite formulation
uA_( yAJ.120b) jQ.= AR F2. + BR 2.}
IUL YLJ > Q= 1,2, ,L ~~~~~~~~~~~~~~~~~~(23a)
- YQ~~~~~~~~~~~2= CR r2+DR uQ
3130

J= - (23b) are properly used, Approximations to the computa-

where tion of H(X) from (8) or from (24) and (14) can be
A =A +A (At )1A21 (23c) devised bas'ed on the above assumption, see [1].
AR = A 12 II I. APPLICATION OF SMA TO DYNAMIC STABILITY ANALYSIS

BR B1I + A12 (l -A22 )B22 (23d) 3.1 Models

2
Q Q Figure 3 shows one possible block diagram fora
C = C A+ C )-1A
(kI - (23e) power plant model, suitable for Dynamic Stability
Rg 1g 2g 22 21 analysis. Here the variables are the real (P) and
+ C2 (AI -A22 )B2
-l reactive (Q) power out of the generator, its termi-
D- D (23f) nal voltage magnitude (Vt) and angle (a), the angu-
QQ Q Q lar displacement (6) and velocity (w) of the rotor,
and the global ROM (14) where the mechanical power into the shaft (P ), the air-
gap electric power (PE) and some measure of the vol-
AR (A) = AR + BR (I - J DR ) J CR tage applied to the field winding (EFD).
Rd Rd Rd Rd
=A R + B R J(I - D R J) CR (24)
d d d d
The following possibilities for further compu- Machine
tational simplification can be mentioned: p
(i) Consider the case where only Lo out of all[I-G QE[ s
the L dynamic subsystems are significantly involved FD
in the SM. Then the composite formulation allows us P
to completely collapse the dynamics of the less in- Prime WE -ntX
volved subsystems (since none of their state vari- w + I-a E
ables is considered to be relevant), so (23a) ls
become-sRgltr
r = AR r + B u
YQ

Rg 9+ DR B%U}'t
91
=l2,..L (25a) F

0~~~~
YQ GQt Q= LO + 1,.*.,
where C Fig. 3
G9 R ( ) B
and the contribution of the less involved subsystems
can be incorporated into the static constraints (23b).
(ii) If, due to the nature of the system and a
sui table choice of states,
Ba 2 0 and/or C 2 = 0 (26)
then the computation of A (X) in (24) is greatly
simplified. R and J is the so-called Jacobian of the network.
Note also from (22) that (26) implies A22
is block diagonal, with L blocks corresponding
to the L dynamic subsystems. In this case the
direct computation of H(A) from (8) may become
attractive.
(Tii) It may be the case that, even if A22 is
not strictly block diagonal, there is a negl igible
contribution from the off-block-diagonal entries of
A22 to the eigenstructure of the h SM. This will1
occur whenever the direct interaction between 'the
less relevant portions of the dynamic subsystems
(i.e., interaction that takes place directly through
the algebraic constrai'nts and not via the relevant
portions of the subsystems) ois of ittle relevance
to the h SM'. will
Note that the above assumption is a plau-
sible one whenever the selected relevant dynamics
substantially captures the interactions of interest
between the subsystems. This is to be expected in
most practical cases where the principles of SMA'
L (25b)
Power Plant Block Diagram
R ' (25c) The transfer functions GE , GM, GI, GR and Gp for the
different components may contain any desired level
of detail.
The model for a system of power plants of the
above form, interconnected by a transmission net-
work, is naturally formulated in the composite form
(20), where
~~~~~~~~T=[6,,T~T ~T ,T T=
y~[~,x T=[Q
,XE M R ,
Figures in the range of 15 to 20 states' per power
plant may be considered typical of the level of
modeli ng detai'l that may be needed in a Dynamic
Stability'study. (There is still no consensus on
this issue, see [4, 51.) Here it wi ll be assumed
for simplicity that the loads are modeled as constant
impedances, but load dynamic models can be incor-
porated into (20).
Models of widely different orders have been
analyzed for Dynamic Stabilit-yusing the SMA approach,
[1i,3], with the largest being a 60-machine power
see
system ldel that is successful'ly used in [5] to
descrTbe an actual dynamic stability incdent.
Modelse nofwdely dferevnt Stersh
The starting point in the selection of the re~
levant states is the definit'ion of the h selected
natural modes (SM). In Dynamic I Stabi ity studies
we are interested in those ightly damped, oscil la-
tor modes in the frequency range of 0.1 to 2 Hz

that are associated with power oscillations.
A good basis for selecting a suitable set of
relevant states is the participation matrix P. The
usefulness of P in the identification of dynamic
patterns has been tested in numerous examples, see
[3], and does not seem to be affected by the order
of the model. Since any procedure that uses P
requires complete eigenanalysis of the full-order
system matrix, such methods are restricted in prac-
tice to situations where such computational cost is
justifiable. For instance, when we are trying to
understand the structure of a nominal model of a
system that is going to operate for (and to be ana-
lyzed for) a large set of slightly different oper-
ating conditions, one may permit a single computa-
tion of the complete eigenstructure of the nominal
model; structural insights obtained from this case
may then be applied to slightly different operating
conditions. Note also that for purposes of just
selecting relevant state variables it is typically
enough to have an approximate knowledge of P.
A fact that has been observed in all the power
system models studied so far is that the set of re-
levant states r that is suggested by the participa-
tion matrix P always includes the w and 6 states of
all the machines and, sometimes, a few more states.
If rT=[6T,JT] (where now 6 and w are vectors with as
many entries as there are machines), then the eigen-
values of the resulting A1l matrix in general, but
not always, yield good approximations to the fre-
quencies (imaginary parts) of the swing modes. The
term "swing modes" is loosely used here to describe
those natural modes of the system that are heavily
associated with intermachine power oscillations.
One or more of these modes are always involved in
any Dynamic Stability problem. The distinction be-
tween swing and nonswing modes is not always clear,
however, see [3].
It is commonly assumed that an approximation
to the swing modes is provided by the classical
electromechanical model of the power system (the
block diagram in Fig. 3 reduces to the inertia
equations and a constant-voltage-behind-reactance
static electric machine model). There are good
reasons to maintain that the partition of the system
matrix A of a detailed power system model, when
rT=[6T,WT] is chosen, closely approximates the
system matrix of the classical electromechanical
model, see [1]. It has been shown that the distinc-
tion between both matrices is only due to the diffe-
rences between the effective stator winding reac-
tances used in more detailed models and the transient
reactance used in the classical model. Moreover,
the machine stator reactances in (22) occur in com-
bination with the generally larger network reactances
in J. It may then be argued that eigenanalysis of
All will provide a reasonably good approximation to
the frequencies of the swing modes.
T T T
One should not conclude, however, that r =[6,w]
in generanl i s the right choice for analysis of Dy-
namic Stabil ity. A typical feature of dynamic in-
stabi l ity phenomena i s that the "swing modes" of
interest (the unstable and the very l ightly damped
ones) are s igni ficantly affected by system compo-
nents (exc itation control s, governors, etc .) whoseS
associated state variables are not included in the
3131
6,w set. Hence some augmentation of the 6,w set of
relevant states will in general be needed in order
to properly analyze the critical modes. The above
statements have been numerically corroborated in
practice, see [1, 3].
When the participation matrix P is not available,
the cri teri a to be used i n the sel ect.ion of r wi ll
strongly depend on the a priori information about
the system, since knowledge regarding the existence
of critical modes, or their approximate natural
frequencies, or regarding the most involved units
might be available. In the absence of such a priori
information, a sensible criterion for augmenting r
consists of examining the results of a couple of
iterations of the SMA iterative algorithm with
rT=[6T,WT]. Slow convergence to a lightly damped
mode(s) may indicate dynamic instability. The most
involved units can be pinpointed with the matrix P
of the reduced model and the relevant set r can be
augmented with non-6,w states from these units (this
approach has been successfully tested on the 60-
machine model, see [3]). If numerous units appear
to be significantly involved in the critical mode(s)
then eigenvalue sensitivities can be used to deter-
mine the units and components requiring augmentation.
How to efficiently evaluate these sensitivities will
shortly be shown.
It must be noted that the 6,w set will in ge-
neral be a small fraction of the total number of
states in the model. Further reduction of the di-
mension of r might be achieved if coherent groups
of generators for the h SM are identified, see [1].
Then one generator from each group can be chosen as
reference and the remaining 6,w states transformed
into 6k S7ref' Q 17ef'
W The initial relevant set
r now will only consist of the 6,w states for the
reference machines.
The standard SMA algorithm described by (16)
has been successfully appl ied to several power sys-
tem models of widely different orders, see [ ,3].
The results for the dynamically unstable 177th order
model are particularly interesting and have been
reported in detail in the companion paper [31.
Excellent convergence properties of the algorithm
have been obtained whenever the selction of rele-
vant states compl ied with the previously stated
rules.
3.3 Order Reduction at Subsystem Level
It will now be shown how the model reduction
represented by (15) can be in fact carried out by
manipulation of the block diagram of the power plant
(Fig. 3), thus avoiding the computations in (23)
and yielding a particularly meaningful ROM.
Fiprst the overallttransfer function of the
power plant will be written in a way that facilitates
the identification and understanding of the contri-
bution of the different power plant components.
From the prime mover and the swing (inertia) equa-
tions one easily obtains
t = Gxls,s E(7
6 = - ( (28)
3132

whe re G G
G (s) G=- 1s G =_ IDG() D-
(29)( TS+W0 W TS+w D G +G
ER+GVPL P+ (GER2+ GVQL)
T is the machine inertia, Dw represents otherwise (33b)
effects,
unaccounted for torque and w is the no- wwhere aall1 the remai|ni|ng Laplace operators
minal~~~
vauoftemciesed
~ ~ explicitly shown. The state space model
have been
The above equations can be combined with the
synchronous machine model 1 0o 1 0 1FP
-- 1- - -~~
~~P _. +.l
oavtj =GE4
GE GE3 |I |Q |
GE2 GE6 ( 30)O j I IJG) Q[

SEJI
GE5
7 EE8 E9
G0
J tl
OGj ER VPL-Tw
ER3 P VQ
to give L 1 1 L
=-G
sc
P+G E4PP+G E5 Q+G E6EFD+ sw (P E
G -p) (31) G ER6GPP JGE4+G
-L ER6GEI1RGcPL GaQj
c Q
where the contribution to at that results from the (34)
difference between P and P (in a model where the
induced voltages associated with stator flux link- A G
= (35)
ages are ignored the difference is due to the stator is consistent with (33). This is the desired ROM
ohmic losses only) is explicitly shown. After some at subsystem level. Note that in practice 6 and Xj
manipulation (31) becomes are almost identical to 6 and u and also that (34)
o=(s GW+GE4
E4+GER6GElGR
+ER E +GER6GG G
+ )P exactly yields (33). It can also be checked that
w
R~-otPL the subsystem reduction equations (23) directly
+ (G E5+ G ER6GE2 G R+ G aQL ) Q (32a) result in (34).
E5 ER6 E2
atQL''~ The reduced order formulatTon (34) is very
where simple to compute and can be used to efficiently
G ERi =A G Ei /(l- G GRR (33) perform the reduction of the t-th power plant model
E3 at subsystem level whenever rT=[6 It must be ,,W9J.
and GaPL and GonL are small correction terms account- noted that the procedure just described is by no
ing for the difference between P and PE (ohmic means limited to the case where 6 and w are the re-
losses). This equation thus shows the relationship levant states and it has actually been appl ied for
between the inputs P, Q and the output Ol. other choices of r see [2]. It should be remem-
The same procedure can be followed with the
bered that the perftormance of the SMA algorithm
does not depend on whether the reduction is per-
other
oututvaialetobformed at subsystem or global level.
t =(G ERI GER3GGP+ GVPL)P+ (GER2 + G VQL) Q (32b) An important aspect of symbol ic reduced formu-
lations such as (34) is that they greatly facilitate
where G and GvQL account for the small difference the understanding of the model behavior. Reduced
Pas
efore. Note that the global transfer plant models as in (34) together with the network
function equations (32) for the power plant make equations yield a mechanical model of the standard
explicit the contribution from each component type
(namely inertia, prime mover, voltage regulator, [ O I]
stabilizer and synchronous machine). 1 || (36)
Model order reduction at subsystem level is .
easily performed from (32). For instance, let Ls
L KJ K L]
r= [6p, i]. Then from (15) and from the above
procedure for obtaining (32) it is clear that all where K and K Dare the matrices of X-dependent syn-
the component blocks G G G, G, etc. in (32), chronizing (spring) and damping coefficients, res-
except for G , must become Vrozen at their values pectively.
for s= X. TWus (32) becomes It can then be said from inspection of (34),
l+sG ER6 GP
G (36) that the prime mover appears in the role of
l
{ I22 +GDsG +
E4 +GER6GElGR +GaPL p providing "generalized" damping action. As GM is a
complex number, both synchronizing and damping
o O M effects will result, with the relative importance
of each one for a g iven mode being ascerta ined from
+ (GE5+ GER6GE2GR+GGQ) Q (33a) the value of GMGk) itself. The role of the remain-
etQL ~~~~~~ingpower plant components can al so be roughly
assessed in a s imi lar manner.

r =[6T ,wT] y ields a ROM in the form of ( 36) that

Note that the SMA iterative algorithm whenever

exactly contains any desi'red number h. n of swing
modes. In particular, for h= n the uniquely deter-
mined, real, reduced system matrix All exactly
contains all the swing modes of the system.
Sensitivity analysis concerning the influence
of a particular component of a given power plant on
a particular mode is greatly facilitated by the sym-
bolic formulation (34). Note that each power plant
is represented by the numerical value of its trans-
fer function for s= X (for instance, the complex
number Gpz(X) represents the stabilizer of the Q-th
power plant). This simplifies the sensitivity ana-
lysis and the eventual control design effort (which
now does not have to account for the individual
sensitivities with respect to all the internal pa-
rameters of the stabilizer, for instance, but only
for providing the correct gain and phase at the
given frequency A).
Although the above ideas have been illustrated
with a particular choice of state and input/output
power plant variables, extensions to other formu-
lations are conceptually as direct. The influence
of the formulation on the computational efficiency
of the SMA procedures should not however be ignored,
see [1]
3.4 Notes for a Real-Time Control Strategy
for Dynamic Stability
Our conception of a comprehensive approach to
the Dynamic Stability problem requires that it in-
volve at least the following elements: (i) a linear-
ized model of the power system, updated as necessary
and containing all the potentially relevant compo-
nents in detail; (ii) an accurate and efficient
method of assessing the Dynamic Stability status of
the system based upon the above model; (iii) a simple,
reliable, effective and adaptive control strategy
that is based upon the information made available
by the above-mentioned method of analysis.
Broad guidelines for designing a comprehensive
analysis and control scheme for Dynamic Stability
will now be discussed. This scheme corresponds to
a central supervisory type of control, see [11, that
is based upon information that can be made available
by SMA techniques. Essential elements of this
approach are:
(a) Regular (or immediate on entry into an
emergency state) determination of the Dynamic Sta-
bility status of the system. Efficient use of a
priori information (measurements, recurrent critical
modes, etc.). Determination of the level and type
of involvement of each power plant in the critical
mode(s).
(b) An information flow that exploits the
decomposition characteristics of the SMA procedures,
namely model reduction at subsystem level and abili-
ty to compute sensitivities with respect to transfer
functions of entire components (e.g. gain and phase
objecttves for the stabilizers at specified fre-
quencies, for instance, may be assigned to power
plants by a higher level control, but detailed im-
plementation is left to the individual power plants.)
(c) A control scheme designed around the in-
formation contained in the sensitivities and parti-
cipation matrix of the reduced model. The control
problem might be posed as a constrained optimization
problem, for instance, see [13]. The constraints
3133
may include the limits on the available control pa-
rameters and some minimum damping requirement for
any eigenvalue of the system. The objective might
be to somehow minimize the amount of control action.
As indicated previously, the control may be imple-
mented in two or more echelons, if necessary, until
all critical modes are eliminated.
It must be realized that the basic philosophy
behind the proposed approach is also meaningful in
the context of off-line operation and planning
studies, as for instance in assessing the Dynamic
Stability characteristics of new power system con-
figurations resulting from addition of generation
and/or transmission facilities.
IV. CONCLUSIONS
It is our main conclusion that SMA has much
potential to be made the basis for a comprehensive
approach to the Dynamic Stability problem: SMA-based
algorithms can accept very high order models, accu-
rately and efficiently compute the poorly damped
modes and their sensitivities, determine the system
components most involved in these critical modes,
and feedcthis information to an adaptive control
scheme designed around such information. Critical
aspects of SMA have been successfully tested on a
60-machine model of a dynamically unstable condition
in an actual power system.
Realization of this SMA potential still requires
some research in order to assess the computational
efficiency of SMA procedures more precisely. Other
aspects still requiring further consideration in-
clude the choice of the most adequate model formu-
lation and the elaboration and testing of a prac-
tical set of rules for the selection of the rele-
vant states when the participation matrix is not
available.
Research is in progress concerning the appli-
cation of SMA to other power system problems. This
includes using the participation matrix for identi-
fying coherent groups of generators (to be employed
in the elaboration of dynamic equivalents for
Transient Stability studies), and also the appli-
cation of the SMA theory on reduced input/output
characterization of LTI systems, see [1, 2] , for
determining modal dynamic equivalents for different
types of stability analysis.
ACKNOWLEDGEMENTS
Support for this work has come partially fromr
the Electric Power Research Institute, under Contract
RP 1764-8, monitored by Dr. Neal Balu, and, in case
of the second author, from the ITP Foundation in Spain.
Dr. Stephen Umans helped with modeling issues.
REFERENCES
1. Pwrez-Arriaga, I.J., "Selective Modal Analysis
with Appli,cations to Electric Power Systems,"
Ph.D. Thesis, Electrical Engineering, Massachu-
setts Institute of Technology, June 1981.
2. Peez-Arriaga, I.J., Verghese, G.C., and
Schweppe, F.C., "Selective Modal Analysis,"
submitted for publ ication. Preprints avai lable
from the authors.
3. P&rez-Arriaga, I.J., Verghese, G.C,, and
Schweppe, F.C., "Selective Modal Analysis with
3134

Applications to Electric Power Systems, Part I:

Heuristic Introduction," companion paper, these
Transact ions.
4. EPRI, "Frequency Domain Analysis of Low-Frequen-
cy Oscillations in Large Electric Power Systems,"
EPRI EL-726, Project 744-1, Final Report, March
1981.
5. Van Ness, J.E., Brasch, F.M., Jr., Landgren, G.L.,
and Naumann, S.T., "Analytical Investigation of
Dynamic Instability Occurring at Powerton
Station," IEEE Trans., Vol. PAS-99, No. 4, July-
August 1980, pp. 1386-1395.
6. IEEE, "Power System Stabilization via Excitation
Control," IEEE Tutorial Course 81 EHO175-0-PWR,
1980.
7. De Mello, F.P., Nolan, P.J., Laskowski, T.F.,
and Undrill, J.M., "Coordinated Application of
Stabilizers in Multimachine Power Systems,"
IEEE Trans., Vol. PAS-99, No. 3, May-June 1980,
pp. 892-901.
8. Cresap, R.L., and Hauer, J.F., "Emergence of a
New Swing Mode in the Western Power System,"
IEEE Trans., Vol. PAS-100, No. 4, April 1981,
pp. 2037-2045.
9. Yu, Y., and Moussa, H.A.M., "Optimal Stabiliza-
tion of a Multi-Machine System," IEEE Trans.,
Vol. PAS-91, No. 3, May-June 1972, pp. 1174-1182.
10. Obata, Y., Shoichi, T., and Suzuki, H., "An Effi-
cient Eigenvalue Estimation Technique for Multi-
machine Power System Dynamic Stabil ity Analysis'
IEEE Trans., Vol. PAS-100, No. 1, January 1981.
11. Pgrez-Arriaga, I.J., Verghese, G.C., and
Schweppe, F.C., "Determination of Relevant State
Variables for Selective Modal Analysis," Proc.
Joint Automatic Control Conference Charlottes-
ville, Virginia, June 1981.
12. Verghese, G.C., Perez-Arriaga, I.J., and
Schweppe, F.C., "Measuring State Variable Parti-
cipation for Selective Modal Analysis," IFAC
Symposium on Digital Control, New Delhi, India,
January 1982.
13. Dy Liacco, T.E., "Real Time Computer Control of
Power Systems," IEEE Proc., Vol. 62, No. 7,
July 1974, pp. 884-891.