m e d i n g s of the 42nd IEEE

Conferenceon Decision and Control

Maui, Hawaii USA,December 2003 FrM14-4

Calculation of Stability Region

Daizhan Cheng Jin Ma

Institute of Systems Science Dept. of Electrical Engineering
Chinese Academy of Sciences Tsinghua University
Beijing 100080, P.R.China Beijing 100084, P.R.China
dchenql?iss03.iss.ac.cn Jinma@ps.tsinqhua.edu.cn

Abslracf-In the stability analysis of power systems it is boundary of the region of attraction under the assumptions
extremely important lo determine the stability @on of a that
working (equilibrium) point. This paper first gives a quadratic (i) the equilibrium points on the stability boundary a A ( z , )
approximation to the boundary sub-manifold of the stability
region, which assures the error he of O(llxIl3). Under certain are hyperbolic;
non-singularity assumption, a precise expression of the suh- (ii) the stable and unstable sub-manifolds of the equilib-
manifold is obtained as a Taylor expansion. The formula is rium points on the stability boundary a A ( x . ) satisfy the
then extended for diKerential-algebraic systems. Its application transversality condition;
to power systems is illustrated via an example. The computation
is based on the lefl semi-tensor product of matrices, which was (iii) every trajectory on the stability boundary a A ( z , )
proposed in [l],[2]. approaches one of the equilibrium points as t + W.
It is well known that the stability boundary is of dimension
1. INTRODUCTION n - 1 [3]. Therefore, the stability boundary is composed
Consider an analytic smooth nonlinear system of the form of the closure of stable sub-manifolds of type-1 equilibrium
points on the boundary. Based on this fundamental fact, it is
j. = f(z), 2: E R", (1) of significant meaning to calculate the stable sub-manifold
of the type-l equilibrium points. Recently, some algorithms
where f(x) is an analytic vector field. Suppose z, is an equi-
have been developed to get an approximation to the stable
librium point of (1). The stable and unstable sub-manifolds
sub-manifold of the type-1 equilibrium points. For instance,
of x, are defined respectively as
in [9] the Taylor expansion was used to get a quadratic
W-'(x,)= { p E R"Ilimt-r x(t. p) + G} approximation, and in [7] another similar quadratic approx-
WU(z.)= { P E R"/1imt--=x(t,p) -+ x , } . (2) imation via normal form of vector fields was presented.
In this paper several formulas for the Taylor expansion of
Suppose zsis a stable equilibrium point of (1). The region the stable sub-manifold are obtained. Its first two terms form
of attraction of xg is defined as the best quadratic approximation. The result is based on a
new matrix product, called the left semi-tensor product of
matrices, which is proposed in [I], [ 2 ] . The paper is self-
The boundary of the region of attraction is denoted by sustain for the required concepts and results about semi-
aA(z s 1. tensor product.
An equilibrium point 2, is hyperbolic if the Jacobian
11. SEMI-TENSOR PRODUCT OF MATRICES
matrix of f at z, denoted by J f ( z , ) , has no eigenvalues
with zero real part. A hyperbolic equilibrium point is said to
be of type-k if Jf(ze)has k positive real part eigenvalues. Definition 2.1 1. Let X be a row vector of dimension np,
In the stability analysis of power systems the region of and Y be a column vector with dimension p. Then we split
attraction and its boundary are fundamental concerns in X into p equal-size blocks as X ' , .. ,XP, which are 1 x n
transient stability computations. The key is whether the rows. define the leji semi-tensor product, denoted by K , as
initial condition of the failure on trajectory lies within a
suitable region of attraction. The problem has been discussed
intensively by many researches. The works of [SI, [IO], [31,
(41 established the theoretical foundations for the geometric
structure of the region of attraction.
Particularly, [IO] and [3] proved that for a stable equilib-
2. Let A E Atmx, and B E M p X pIf either n is a factor
rium point z, the boundary of stability region is composed
of p, say nt = p and denote it as A <t B, or p is afacror of
of the stable sub-manifolds of equilibrium points on the
n, say n = pt and denote it as A +t B, then we define the
*Work was supported by CNNSF 60274010 and National 973 Raject leji semi-tensor product of A and B , denoted b)>C = A K B,
GI998020308 of China. as the following: C consists of m x q blocks as C = (Cl])

0-7803-7924-1/03/$17.00 02003 IEEE 5615

and each block is
C'j = A' D( Bj, i = 1, . . . ,m, j = 1 > . . ., q ;
where A' is i-th mw o f A and Bj is the j-th column of B.
In addition, we can also define the inverse of V, (or V F ) .
Remark. Note t l m when n = p the left semi-tensor product
We define it for only square case. Let X E IR"'. Then
coincidesthe conventional matrix product. Therefore, the left
semi-tensor product is only a generalization of the conven- V;'(X) : = A
tional pmduct. For convenience, we may omit the pmduct
synibol K . is an n x n square matrix, such that V , ( A )= X .
Some fundamental properties of the left semi-tensor prod-
uct are collected in the following: Proposition 2.4 Let A E AL,,,, and B E A'iprs. Then

Proposition 2.2 The left semi-tensor pmduct satisfies (as Ivim,pl

( A @ B)lv[q,nl= B @ A . (12)
long as the related products are well defined)

-
1. (Distributive rule)
For later approach, we need the tensor expression of
A (aB + PC) = a A K B + P A
K K C; polynomials. Let x = ( ~ 1 ,... ,xn)= E IR". Then zk :=
(aB+PC) K A = a B K A + B C K A, a,P E IR. x K . .. D( x,and a k-th degree homogeneous polynomial can
(5)
2. (Associative nile) k
be expressed as F K x k , where the coefficient vector F is a
A K ( B K C)= ( A K B ) K c; (6) 1 x nk row, briefly, F x k := F K xk.
( B K C ) K A = B K (CP( A ) . Next, we consider the differential of matrix of functions.
3. Assume A and B are of the pmper dimensions such
that A K B is well defined. Then Definition 2.5 Let H ( z ) = ( h , J ( x ) )be a p x q matrix
with the entries h t J ( x )as smooth functions of x E R".
( A K B)T = BT K A T . (7) Then the differential of H is defined as an m x nq ma-
4. In addition assume both A and B are im,enible, then trix obtained by replacing h,J by its differential dh,, =
( A K B)-' = B-' K A-'. (8)
(Y,... ,*).
Our goal is to apply it to polynomials. We construct an
nk+' x n k + l matrix @ k as
Next, we define the swap matrix, or permutation ma-
trix [6]. Many properties can be found in [2]. The
swap matrix, W,,,,, is an mn x mn matrix con-
i=O
structed in the following way: label its columns by
( l l , l Z , . . , , I n , . ' . .ml,mZ,. . . ,mn)and its rows by by Then we have the following differential form of x k , which
(11,21,. . . ,m l , . . . , I n , 2n,. . . , mn). Then its element in is fundamental in the sequel.
the position ((I,J),( i , j ) ) is set to
Proposition 2.6
1, I = i and J = j ,
(9)
0, otherwise. D(Zk+') = @k P( Zk. (14)
When m = n we simply denote by IV[nl for W,,,,,.
The following properties [2] can be proved by using
Inductively, we can define
definitions with straightforward computations. .
D k + ' H ( z ) := D [ D k H ( z ) ] , k 21
Proposition 2.3 Let X E R"'and Y E IR" be m o columns.
Then Using above expression, the Taylor series expansion of a
Iv,,,,, K x K Y = Y x, K
(10)
vector field f(z)is expressed as
1v[,,,,,]K Y x = x Y.
K K m _

let A E AI',,,,,,
Panicularl~~, V,(A) and V,(A) be its mw
and column stacking forms. i.e., assume A = (a'j), then
% ( A ) = ( a l l , . . . , a l n , . . . , a m i , . . . ,am,)T, It is exactly the same as the Taylor series expansion of single
V J A ) = ( a l l , . . . ,ami,... ,al,,... variable functions.

5616
 111. QUADRATIC APPROXIMATION Next, we express h(x) by its Taylor expression as
In this section we investigate the quadratic approximation
to the stable sub-manifold of unstable equilibrium points.
+ +
h ( ~=) H ~ x HZX' H3x3 ... + (22)
Comparing with exist quadratic approximations [9], [7], the
advantage of this approach lies on two points: 1. Precise Lemma 3.4 The quadraiic form, H2 = iQ, in the stable
formula is given; 2. The approach is the unique one, which sub-manifold equation (22) satisfies
carrying the error of O(11~11~).
Without loss of generality, we can assume the type-l
equilibrium point concerned is xu = 0.
Using (14), we can express the vector field f in ( I ) by
where p and 7 are defined as in Comllary 3.2 for A = Fl =
JJ. 7, is ihe i-th component of 7. Hess(f,) is the Hessian
f (z) = F,x' = Jjx + Pzx2 + . .. . (16) matrix of the i-rli component f, o f f .
,=I

Where Fl = J f = J,(O), and F, are known II x n' matrix. Lemma 3.5 Equaiion (34)has unique symmetric solution.
We use A-T for the inverse of AT, which is the transpose Summarizing Lemmas 3.2,3.4,3.5, we have the following
of A. quadratic form approximation of the stable sub-manifold.

Lemma 3.1 Lei A be a hyperbolic matrix. Denote by V, and Theorem 3.6 The siable sub-manifold of I,, expressed as
Vu the siable and unstable sub-spaces af A respectively and h ( x ) = 0, can be e.rpressed as
by U, and U, the stable and unstable sub-spaces of A-= 1
respectively. Then h ( z )= H I X + -xT*x
2
+ 0(11~11~), (24)
v6' =U,, v,' =U,. (17) where

The following is an immediate consequence of the above

lemma.

Corollary 3.2 Let A be a ype-1 matrix and assume its only

unstable eigenvalue is fi. Lei 7 be ilie eigenvector of AT wirh where fi and 7 are defined as in Corollary 3.2 far J J = F1,
respect to eigenvalue p. Then 7 is perpendicular to the stable Hess( f") is the Hessian matrix of ihe i-th component f i of
subspace of A. f.
Before investigating the main results the following nota-
tions are required. In fact, they are from definition immedi- IV. HIGHER DEGREE APPROXIMATION
ately. In this section we consider the whole Taylor expansion of
I". := 1, @o :=I,. the stable sub-manifold. In later calculation of the differen-
tials .we have to calculate @ I ; . For this purpose we need the
Without loss of generality, we assume, through Section
following:
2-Section 4, that x, = 0 is a type-I equilibrium point of
system (1).
Proposition 4.1
The following is fundamental for our approach.

Theorem 3.3 Assume xu = 0 is a Vpe-1 equilibrium point

af sysiem (1).
wlns.nl
= n (m.
8-1

i=O
B wl,,", B In.-.->)
(25)

IV(e,) = {x 1 h ( z )= 0). (18) Next, we will solve Hk from (19)-(21). The problem is x k
is a redundant basis of the k-th degree homogeneous poly-
Then h(x) is uniquely determined by thefollowing conditions
nomials, so from (19)-(21) we can not get unique solution.
~191421J.
To overcome this obstacle we consider the natural basis of
the k-th homogeneous polynomials. Let S E Z?. The natural
basis is defined as
B: = { xs/S E 22,IS1 = k}.
Now we mange the elements in in alphabetic order. That
is, for S' = (si;.. ,SA)
and Sz = (s:,... .si), we sign

5617
the order as zs' 4 xs2 if there exists a t, 1 5 t 5 n - 1,
such that
s1 - 2
1 - sl. ' . . , st' = ,
:s 1
St+l > S 2t + 2 .
k 2 3
Then we arrange the elements in E32 as a column and denote
it as q k ) . are non-singular; then
Next, we consider the dimension of k-th order homoge-
neous polynomials. First of all, by mathematical induction
we can get the following formula for dimensions.

Proposition 4.2 The size of the basis Bk is

( n + k - l)!.
lB,!l := d = k 2 0, n 2 1. (26) Remark. 1. So far; we still don't know when Gk is nun-
k ! ( n- l)!
singular: But we do have example to show that unlike
Two matrices T ~ ( n , kE) A f , r x d and T ~ ( n , kE) Af,,, quadratic case, it is possible for Gk, k 2 3 to be singular.
are define as the linear mappings between two bases. (We 2. In case of singular GI;. we can, instead of precise
refer to [2] for detailed construction.) That is, solution, solve a set of linear algebraic equations to get the
lease square solution, which also provides an approximation
zk = TN(n,k)Z(k), z ( k ) = T B ( n , k ) z k . of h b ) .
It follows that V. FOR DIFFERENTIAL-ALGEBRAIC SYSTEM
TB(n,k)TN(% k ) Id.
Since real power systems are differential-algebraic sys-
tems. We consider the following system
Instead of solving Hk, we may try to solve Gk, where
j.= f(x.y), XER", y E R m
HkXk = G k X ( k ) . (31)
t h y ) = 0, E ( + > Y ) E Rm,
9is said to be a symmetric coefficient set, provided that where (o, o) = o, E(o,o) = o, Moreover, we the
if two elements of x k are the same then their coefficients
equation is well posed. Then we should have
are also the same. A straiehtfonvard comoutation shows the
I

following:
(32)
Proposition 4.3 m e s)mmetric coeflcienr set, Hk is unique. we look for the quadratic approximation to the stable sub-
Moreover; manifold of the type-1 unstable equilibrium. Using Chain
rule. we have
(33)
ax
Now we consider the higher degree terms of the equation
h ( z ) of the stable sub-manifold. Denote
aazf
J = -(O,O) - af(0,O) (~(0,0))-13(0,0)
ay ax . (34)
f ( x ) = FIX + F2x2 + ,.. ;
To use Theorem 3.6, we have to find the Hessian matrix
h ( z )= H1z + HZX' + , , . . offi, denoted by Hi,which are expressed as
Note that Fl = J f ( 0 ) = J , H I = qT, and Hz is uniquely Ht = D(V.fz)l(o,~)~i = I,... , 1 ~ . (35)
determined by (24).
Denote by
Proposition 4.4 The coefficients, Hk, k 2 2, of h(z) sarisfy T
the following equation.
(36)

1
k
Ht@t-i((Ina-i @Fk-t+l) - PHk Xk, k 2 2,
(28) T

The following result, which is a summary of the above (37)

arguments, is generically applicable.

5618
 The following formula is fundamental. generators, its frequency fluctuation takes much longer time
scale than that of the other two machines. Consequently, it
Proposition 5.1 Ler A(z) orid B ( x ) be p x q and q x T can be neglected here. w1 and w2 are generator rotor speed
functional mamrices respectively. Then deviations from the synchronous rotation axis. Pml , P m 2
and Pel, Pe2 denote the mechanic power output as well
+
D A ( z ) B ( z )= D A ( z ) D( B ( z ) A(z)DB(z). (38) as the electrical magnetic power of respective generators.
El, E2. E3 are equivalent transient intemal electrical forces
Using Proposition 5.1 and the Chain rule, we can prove of generators. G and B are respective element of nodal
the following formula. conductance matrix that is formed by network reduction. d l
and dz are individual damping coefficient for each machine.
The numerical value of these system parameters can be found
D,~(GY(~ =)D,A(s,y)
)
+ D v 4 z . y ) (I" @ e) (39) in 151.
On the on-line operation of power system, engineers have
Here we use D, for the partial differential with respect to x always been interested in describing the region of attraction.
only. If the region of attraction could be available, the security
Using (38) and (39), we can prove the following formula level of the current operation point is known and preventive
for Hi: scheduling may be made if necessary. However, it's not an
easy task to describe the region of attraction. The only way
available is to integrate the power system backward step by
step along the stable manifold of the unstable equilibrium
point, which is not only time consuming and computation
intensive, but also, actually, can't be realized practically.
Fig.7 gives both the accurate and approximated region of
attraction for this 3-machine system. In this figure, SEP
where X , Y and 2 are defined in (36), (37) and (33) stands for the stable equilibrium, while UEP stands for the
respectively. unstable equilibrium. The number in the parentheses is either
VI. APPLICATION TO POWER SYSTEMS 1 or 2, which shows the type of equilibrium. It's a well
known fact that for an n-dimensional dynamic system, if the
The following example is classical for investigating power assumption of (i)-(iii) in introduction section of this paper
systems. We use it to demonstrate the algorithms developed are satisfied, and if the region of attraction is bounded, the
in the paper. boundary of region of attraction must contain type-n equi-
librium point, which is also called source [3]. Fig.7 clearly
Example 6.1 [5] Consider a three-machine system with shows this. It also can he seen that the approximated region
machine number 3 as the reference machine, the respective of attraction has very high accuracy. Thus, the proposed
dynamic equations are described by the following differential algorithm provides a very good approach for on-line security
algebraic equations: analysis.
61 = w1 In the above example the computer programmed computer
routine are used for the numerical results. In fact, the
(2H& = P1
, - P,] - dlwl
software created by us can be used for power systems with
62 = w2
large number of machines.
(2H2)hz Pm2 - Pe2 - d 2 ~ 2
(41)
Pel - EfGll - E1EzB12 sin(& - 62)
- E1E3B13 sin(& - 63) = 0
Pe2 - E2G22 - EzE1B21 sin(& - 61)
- E2E3B23 sin(& - 63) = 0,
where 61 and 62 are relative angles of generator rotor 1
and 2 with respect to that of generator 3. Fig.] clearly
shows this. It also can be seen that the approximated region
of attraction has very high accuracy. Thus, the proposed
algorithm provides a very good approach for on-line security
analysis.
Since in this 3-machine system, the reference machine
has much greater inertial constant than that of the other

5619
 Fig1 Accurate and approximated boundary of region of VIII. REFERENCES
attraction
[I] D. Cheng, Semi-tensor product of matrices and its ap-
plication to Morgen’s problem, Chinese Science, series
F, Vol. 44, No. 3, 2001, 195-212.
[2] D. Cheng, Matrix and Polynomial Approach to Dy-
namic Control Systems, Science Press, 2002.
[3] H.D. Chiang, M. Hirsch and F. Wu, Stability regions of
nonlinear autonomous dynamical systems, IEEE Trans.
Automat. Conn, Vol. 33, No. I , pp 16-27, 1988.
[4] H.D. Chiang, F.F. Wu, Foundations of the potential
energy boundary surface method for power system
transient stability analysis, I€E€ Trans. Circ. S)% Vol.
35, No. 6, pp 712-728, 1988.
[5] H.D. Chiang, C.C. Chu, G. Cauley, Direct stability anal-
ysis of electric power systems using energy functions:
theory, applications, and perspective, Proceedings of
lEEE Vol. 83, No. 11, pp 1497-1529, 1995.
[6] J.R. Magnus, H. Neudecker, Morrir Diflerenriol Cal-
culus with Applicorions in Sraristics ond Econometrics,
.5 Revised Ed., John Wiley & Sons, 1999.
4 . 3 - 2 - 1 0 1 2 3 4 [7] S . Saha, A.A. Fouad, W.H. Kliemamm, V. Vittal, Stabil-
Rotarangleb, ity boundary approximation of a power system using the
real normal form of vector fields, IEEE Trans, Power
Sys., Vol. 12, No. 2, 797-802, 1997.
VII. CONCLUSION [8] P.V. Varaiya, F.F. Wu and R. Chen, Direct methods
for transient stability analysis of power systems: cecent
results, Proc. IEEE, Vol. 73, No. 12, 1985, pp 1703-
In this paper we provide a series of formulas to calculate 1715.
the coefficients of the stable sub-manifold of the type-l [9] V. Venkatasubramanian, and W. Ji, Numerical approxi-
unstable equilibrium points. The advantage of this approach mation of (n- 1)-dimensional stable manifolds in large
lies on the following facts: systems such as the power system, Auromorico, Vol. 33,
I . Unlike the existing approximations, where only the No. IO, 1997, 1877-1883.
quadratic terms can be approximated. For new approach [IO] J.G. Zaborszky, J.G. Huang, B. Zheng and T.C. Leung,
proposed in the paper, the sub-manifold can be approximated On the phase protraits of a class of large nonlinear
to any order of accuracy as required. dynamic systems such as the power systems, ZEEE
2. Unlike the existing quadratic approximations, the Jordan Trans. Autom. Conrr, Vol. 33, No. I, pp 4-15, 1988.
canonical form expression of the linear part o f f is required.
Only the largest real eigenvalue and the corresponding
eigenvector are need for the new approach. In very high
dimensional case, get the Jordan canonical form is extremely
diflicult.
3. No coordinate transformation is required. In the existing
approximations, the coordinate transformation is required.
Then after the approximated sub-manifold is obtained, it has
to be covert hack to the original (real world) coordinate
frame. It. is not only complicated in computation but also
some deformation happened.
4. The formula can be extended to differential-algebraic
systems.
Simulation for the power systems shows the powerfulness
of the new approach. It provides a more accurate way to
solve the energy-based transieni stability control of power
systems.

5620