54

Transactions on Power Systems, Vol. 7, No. 1, February 1992

FAST CALCULATION OF A VOLTAGE STABILITY INDEX
P-A L6f T Smed G Andersson
Student Member Student Member Member
D J Hill
Member
Department of Electric Power Systems
Royal I nstitute of Technology
S - 100 44 Stockholm, Sweden
Department of Electrical Engineering
and Computer Science
University of Newcastle
New South Wales 2308. Australia
Abstract
The minimum singular value of the power flow jacobian ma-
trix has been used as a static voltage stability index, indi-
cating the distance between the studied operating point and
the steady state voltage stability limit. Here a fast method
to calculate the minimum singular value and the correspond-
ing (left and right) singular vectors is presented. The main
advantages of the developed algorithm are the small amount
of computation time needed, and that it only requires infor-
mation available from an ordinary program for power flow
calculations. Furthermore, the proposed method fully utilizes
the sparsity of the power flow jacobian matrix and hence the
memory requirements for the computation are low. These ad-
vantages are preserved when applied to various submatrices
of the jacobian matrix, which can be useful in constructing
special voltage stability indices. The developed algorithm was
applied to small test systems as well as to a large (real size)
system with over 1000 nodes, with satisfactory results.
INTRODUCTION
Problems related to voltage instability in power systems
are today, in many countries, one of the major concerns in
power system planning and operation. One of the causes for
this increased interest in voltage stability problems is that
a load growth without a corresponding increase of transmis-
sion capacity has brought many power systems closer to their
voltage stability boundaries.
Throughout the world there have been a number of system
black-outs caused by voltage instabilities, among these occur-
rences of voltage instabilities are the French collapses of 1978
[1,2] and 1987 [3], the black-out in northern Belgium 1982 [4],
and the Swedish collapse of 1983 [5]. Some of these experi-
enced network collapses have been caused by an unexpected
raise in the load level, sometimes in combination with unusual
conditions in the system, or by a network disturbance such as
the loss of an important transmission line, a transformer or a
generator. The Swedish collapse of 1983 can be described as
a voltage instability caused by a severe disturbance which led
91 WM 203-0 PWRS
by the IEEE Power System Engineering Committee of
the IEEE Power Engineering Society for presentation
at the IEEE/PES 1991 Winter Meeting, New York,
New York, February 3-7, 1991. Manuscript submitted
August 31, 1990; made available for printing
January 3, 1991.
A paper recommended and approved
to cascade tripping of main transmission lines and a network
collapse about 50 seconds after the initial fault. Both the
Swedish voltage collapse and other similar instabilities can be
characterized by a voltage drop which at first is slow and then
becomes rapid.
The voltage collapse phenomenon can be related to the ac-
tion of tapchangers on transformers, current limiters of gen-
erators, inadequate reactive power resources (at least locally),
and load characteristics at low voltage magnitudes [ 6] . The
load modelling is important for voltage stability studies and
the composite load model characteristics may for low voltage
magnitudes be different from those models traditionally used
in (transient) stability studies. I t has also been observed that
the voltage magnitude at nodes may not give a good indica-
tion of the proximity to the voltage stability limit, see e.g.
reference [?I.
The study of voltage instability problems has due to the
above network collapses and other disturbances caused by
voltage stability problems become an important and inter-
esting area of research and studies, both for universities and
the power industry. One of the aims of the present research
in this area is to find a voltage stability indicator, available
from normal power flow calculations, which is suited for both
planning and operational purposes. The purpose of a static
voltage stability index is to, in some respect, quantify how
'close' a particular operating point is to the point of voltage
collapse, i.e. to estimate the steady state voltage stability
limit of the power system. One suggestion for a static volt-
age stability index is to use the minimum singular value of
the power flow jacobian matrix. The use of this indicator,
obtained from a (full) singular value decomposition of the
power flow jacobian matrix, has been proposed and studied
by Thomas and Tiranuchit in [8,9,10].
This paper will first discuss and describe some general as-
pects and facts concerning the singular value decomposition
of a matrix, these conclusions will then be applied to the
power flow jacobian matrix. The following section presents
a new fast method to calculate the minimum singular value
and the corresponding singular vectors for a matrix, and
implementable versions of this algorithm are given for two
power flow matrices. Application of the proposed method
to the power flow jacobian matrix and related submatrices
gives rapid computation of static indices for voltage stabil-
ity. In the next section some numerical results obtained from
studies on a number of different networks with the proposed
fast algorithm for calculation of the minimum singular value
are presented and discussed, and finally some conclusions are
given. The method presented in this paper has been tested
on both small test systems and on a real size power system
model (with over 1000 nodes).
0885-8950192$03.006 1992 I EEE
55
SINGULAR VALUE DECOMPOSITION
The singular value decomposition is an important and prac-
t i d y useful orthogonal decomposition method used for ma-
trix computations [ll]. If the matrix A is an n by n quadratic
real matrix, then the singular value decomposition is given by
It
i=l
where U and V are R by i z orthonormal matrices, the singular
vectors ui and vi are the columns of the matrices U and V
respectively, and C is a diagonal matrix with
C(A) =diag{ai(A)} i =1,2,. . . ,n (2)
where U; 2 0 for all i. The diagonal elements in the matrix
2 are usually ordered so that u1 2 uz 2 . . . 2 U,, 2 0.
If the matrix A has rank r ( r 5 n) its singular values are
the square roots of the r positive eigenvalues of ATA, which
also are the r positive eigenvalues of AAT. These square
roots 61, ~ 2 , . . . , U, are the only nonzero entries in the n by
n diagonal matrix E. U and V are orthonormal matrices of
order n, and their columns contain the eigenvectors of AAT
and ATA respectively. If thus U; is the i th singular value of
A, the vector U; is the i th left singular vector and the vector
vi is the i th right singular vector, then the following relations
between the singular values, the left and right singular vectors
and the matrix A can be written
AV, = DIU, (3)
ATut = (Java (4)
The relation between the singular values and the eigenval-
ues of a matrix comes, as has been indicated in the text above,
from the fact that the entries ua in the diagonal matrix E are
the singular values of A, and by construction their squares
are the eigenvalues of ATA (or AAT). This relationship be-
tween the eigenvalues from either one of the matrices ATA
or AAT and the singular values of the matrix A could also
be described by the following relations
A( A~ A) =A( AA~ ) =E~ ( A)
(5)
A(A) =diag{Xi(A)} i = 1,2, ..., n (6)
where
is a diagonal matrix containing the eigenvalues of the inatrix
ATA or AAT and C is the diagonal matrix, defined above,
which contains the singular values of the matrix A. Further,
let A =UZVT be the singular value decomposition of A,
then the eigenvalue decompositions of ATA and AAT are
given by the following two equations
The smallest singular value of a matrix A is a measure of
the distance, in the &-norm, between A and the set of all
rank-deficient matrices [12]. Furthermore the singular values
are fairly insensitive to perturbations in the matrix elements,
and thus the singular value decomposition is well-conditioned
[13]. To use the above theory on power systems a linearized
relation between the active and reactive powers at nodes ver-
sus the voltage magnitudes and node angles has to be found,
which is established by the power flow jacobian matrix.
The jacobian matrix, J , in the above equation thus contains
the first derivatives of the active power part, F, and the reac-
tive power part, G, of the power flow equations with respect
to voltage magnitudes V and node angles 0.
If the singular value decomposition is applied to the power
flow jacobian matrix, J , then the so obtained matrix decom-
position can be written as
J =UCVT (10)
The minimum singular value, u,&(J), is a measure of how close
to singularity the power flow jacobian matrix is. If the mini-
mum singular value is equal to zero, then the studied matrix
is singular and no power flow solution can be obtained. The
singularity of the jacobian matrix corresponds to that the in-
verse of the matrix does not exist. This can be interpreted
as an infinite sensitivity of the power flow solution to small
perturbations in the parameter values. At the point where
u,,(J) =0 several branches of equilibria may come together
and the studied system will experience a qualitative change in
the structure of the solutions due to a small change in the pa-
rameter values. This point is called a static bifurcation point
of the power system; see e.g. [14].
The effect on the [ A0 AVIT vector of a small change in
the active and reactive power injections can, according to the
above theory of singular value decomposition of matrices, be
computed as
[ 2; ] =VC-'UT [ 2; ]
(11)
The inverse of the minimum singular value, U;', will thus,
from a small disturbance point of view, indicate the largest
change in the state variables. Let
[ ;; ] =U,,
where U,, is the last column of U, then
where vIL is the last column of V. From the above formulas
and discussion about the singular value decomposition of the
power flow jacobian matrix the following interpretations can
be made for the minimum singular value and the correspond-
ing left and right singular vectors
1. The smallest singular value, U,,, is an indicator of the
proximity to the steady state stability limit;
2. The right singular vector, v , ~ , corresponding to U,, in-
dicates sensitive voltages (and angles);
3. The left singular vector, U,,, corresponding to U,& indi-
cates the most sensitive direction for changes of active
and reactive power injections.
Another property of the singular value decomposition
which could be worth noticing is that by adding a column
to the studied matrix the largest singular value will increase
and the smallest singular value will diminish [12]. The size
56
of the power flow jacobian matrix will increase with one row
and one column each time a generator node (PV-node) hits
its limitation for the reactive power capability and changes
into a PQ-node. This change in dimension of the jacobian
matrix will, as described above, reduce the numerical value of
the minimum singular value for the studied matrix.
ALGORITHM
General
In this section the new method for calculation of the min-
imum singular value and the corresponding (left and right)
singular vectors is presented. The algorithm presented in this
section was proposed by T. Smed [15]. In order to be feasi-
ble for practical purposes, the following properties are either
necessary or at least desirable for the computational scheme
1. Sparsity must be exploited;
2. Memory requirement should be as low as possible;
3. The jacobian matrix and its LU-decomposition can
be assumed given after that a power flow calculation
has been performed. Preferably, only this information
should be used in the proposed algorithm - and fur-
thermore it should be used efficiently.
The method proposed here is based on consideration of
amplification in directions defined by the singular vectors un
and v,,. Consider a vector s and the solution r of
JTr =s (14)
I f s is sufficiently rich in the direction of v,, then r will be
rich in the direction of U,,. The reason for this is easy to see
if the singular value decomposition of J =UXVT is used to
analyse the problem. Since the columns of V constitute an
orthonormal basis in R, s can be written as
n
Solving (14) gives
r =UZ-V*S =2 u;a;u;
(16)
i =l
Clearly, s will be amplified mostly in the direction v,, and
r will be correspondingly rich in the direction of U,,. Further-
more, an estimate of U,, can be obtained as
If subsequently the following equation is solved
Js =r (18)
6 can be written as
And a new, improved, estimate for the minimum singular
value of the studied matrix can be obtained from
Computation of c,,(J)
The observations above suggests the following iterative pro-
cedure for finding U,, of the power flow jacobian matrix.
Algorithm 1 Basic algorithm for finding the minimum sin-
gular value and corresponding singular vectors of the large
power flow jacobian matrix, J
1. Solve J~u! . =4.
I Idk II 2
11U!. 11 a
2. Estimate b,, -
3. Solve ~v!, + =U,
4. Estimate b,, - II U, I12
l ~v:, +l l l z
Note that algorithm 1 can, if only every second estimate of
U is used, be interpreted as inverse iteration applied to either
of the matrices JTJ or JJT. As it stands, algorithm 1 only
shows the basic idea of the iterative scheme to be used. A
more implementable version of this basic algorithm for finding
the minimum singular value of the jacobian matrix is given
by algorithm 2 below.
Algorithm 2 Algorithm for finding the minimum singular
value, U,,, and corresponding left and right singular vectors,
U,, and vIL, of the power flow jacobian matrix, J. Let the
LU-decomposition of J be given by J =LU
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
Note
Find a suitable start vector v!:) (see below). Set k =1.
Normalize vl.:
Solve uTz =v!,)
Solve L~ u ! , ~ =z
Estimate b!- =IIU!;~~;
If Ab <tol, stop; else go to 7.
Normalize U!!:
Solve LW =U!,
Solve UV!,+ =w
Estimate b!;k) =~ ~ v ! , ~ ~ ~ ~ ; ~
If Ab <toll stop; else set k =k +1 and go to 2.
that the back and forward substitutions with the two
known matrices L and U are the only major steps of compu-
tational work required in the above algorithm. This together
with the use of sparsity techniques to carry out these com-
putations satisfies the properties stated in the beginning of
this section, and thus makes the above method fast and effi-
cient. I t should also be noted that the proposed method,only
requires informatiofi that is available from an ordinary power
flow program.
In the choice of starting vector, v!:, at step one in the
above algorithm it is of importance to ensvre that the chosen
vector is sufficiently rich in the direction of v,,. One possibil-
ity is to choose v!,! =[0 . . . O 1 . . . 1IT, where the zeros corre-
spond to 0 and the ones to V in the state variable vector of
the power flow equation (9). The motivation for this choice
is that most voltages tend to be either raised or depressed
simultaneously. The direction in which the angles change,
however, is more system dependent, the choice of slack bus
and hence the angle reference has an impact of the sign on
the components in vn corresponding to the angles.
Computation of a;,(Gs)
Under the assumption that detFe #0, which can be inter-
preted as if there are no steady state angle problems in the
studied system, Schur's formula [16] can be used to calculate
the determinant of the power flow jacobian matrix
detJ =detFe detGs (21)
where the submatrix Gs is defined as
Gs :=G v - Ge Fi ' Fv
(22)
Thus, if detFe #0, the matrix J is singular if and only if
the matrix Gs is singular. The above presented matrix Gs
is thus a submatrix, associated with the matrix J, that is
indicative of steady state voltage stability problems [6].
One physical interpretation of the matrix Gs could be
found by setting AP =0 in equation (9)
The quotient between the reactive power part of the power
flow equations and the voltages can then be written as
9 =G v - Ge Fi ' Fv
AV
(24)
where the right hand side of the equation above was defined as
G s in the equation (22). The matrix Gs will thus, for small
disturbances, describe the effect on the voltage magnitudes
of purely reactive power changes in the network. This then
suggests the use of the minimum singular value of the matrix
Gs as a static voltage stability index.
The second term, Ge Fi ' Fv , in the definition of GS (22)
might have a larger influence on the result if the studied sys-
tem is stressed, i.e. high active and reactive power flows in
the network. The importance of active (and reactive) power
stress in the system is one of the factors that has been identi-
fied as important from practical cases of voltage collapses in
different power systems [ 6] . It should also be observed that,
even though the matrix GV is sparse, the matrix G s will be
dense due to the influence of the second term in the above
definition of the matrix Gs .
In order to find a sparse formulation for the computation
of u, , (Gs), note that the following two equations
[ Gv - GeFg' Fv] vC' = U!,"' (25)
(26)
T (ki-1) = (kc)
[Gv - GeFg' Fv] U, vn
correspond to equations (14) and (18) in the general discus-
sion at the beginning of this section. Thus if the the iterative
scheme earlier given for the calculation of on(J) could be mod-
ified to the two equations above the minimum singular value
of GS would be found.
I t is easy to check that v!,k) in (27) also satisfies (25).
Analogously, U!:") in (28) also satisfies (26).
Noting that (27) and (28) can be written as
a modified algorithm for calculation of u, , (Gs) has been for-
mulated; this modified version of algorithm 2 is presented
below. Note that the only modifications needed to the previ-
ous algorithm are the changes made in steps 5 and 11below,
where the elements on the right hand side of the equation
system corresponding either to the active power equations or
to the node angles are set equal to zero.
Algorithm 3 Algorithm for finding the minimum singular
value, U,,, and corresponding singular vectors, U,, and v,,, of
the matrix Gs =GV - Ge Fi ' Fv. Let the LU-decomposition
of the power flow jacobian matrix, be given by J =LU
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
Find a suitable start vector v!:), such that v,,,i =0 for
al l i : s corresponding to angles in v!:'. Set k =1.
Normalize v!,'"':
Solve UTz =vi?
Solve LTU',"' =z
Set u,,,i =0 for all i:s Corresponding
to active power equations.
Estimate d!~' -" =~ ~ u ! ~ ) ~ ~ ~ ~
If Ad <toll stop; else go to 7.
Normalize ua):
Solve LW =U!,"'
Solve UV;~+" =w
Set v,,.; =0 for all i:s corresponding
to node angles.
Estimate d!:" =~ ~ v ~ ~ ~ ' ' ~ ~ ~ '
If Ab <tol, stop; else set k =k +1 and go to 2.
NUMERICAL RESULTS
The algorithm described in this paper was implemented
in a production grade program for power flow calculations,
and results from some of these studies are presented in this
section. The method, presented in the previous section, was
used to calculate the minimum singular value together with
the corresponding singular vectors for either the matrix J or
the matrix G5 for all studied cases for which the power flow
program did converge.
The new fast algorithm was tested on the following systems:
Ward-Hale 6-node system [17], AEP 30- and 57-node systems
[18], as well as a 1033-node model of the Swedish power sys-
tem. The tolerance (tol), which in the above algorithms was
58
used to determine when to stop the iterative procedure, was
chosen to be a relative tolerance and set to either lo-' or
lo-'. The tolerance of was only used during studies
performed on the 1033-node model of the Swedish network.
For all systems, but the largest network, the results obtained
with the above described method was compared with results
from other programs (MATLAB and SLATEC). For all tested
cases the results obtained with the new method have, within
the specified tolerance, agreed with results obtained from a
full singular value decomposition of the studied matrix.
The CPU-time needed to calculate the minimum singular
value has approximately been equal to one to three full (non-
decoupled) power flow iterations with the used program. The
sparsity of the power flow jacobian matrix, J , and the spe-
cial structure of the matrix Gs were fully utilized during the
calculations. The sparse solution method used has been the
same as the one used for solving the power flow equations
in the program. It should also be noted that the above pre-
sented computation time for the proposed new fast algorithm
includes a LU-factorization of the studied matrix.
Hard-Kale +node test system
1.4
1
0 0 1 0 2 0 3 0 4 0 5 0 8 0 7 0 8
a16
Figure 1: a,(J) and a,(Gs) from Ward-Hale system plotted
versus the reactive power load at node number six
Figure 1 shows the minimum singular values of the two
matrices J and G s together with the voltage magnitude at
two nodes plotted as functions of the reactive power load at
node number six. During these simulations, performed on
the Ward-Hale 6-node system [17], the reactive power part of
the load at node six was increased in steps until the power
flow program used no longer converged. No limitations for
reactive power generation at generators were introduced to
the system, and the studied matrices thus did not change
their dimensions.
The second graph, figure 2, shows the minimum singular
value for the two matrices J and Gs together with the voltage
magnitude at two nodes in the load area of interest, from a cal-
culation performed on the AEP 30-node system [18]. During
this computation the active and reactive load at four nodes
(nodes: 17 - 20) waa increased in steps and the power fac-
tor was kept constant at 0.89. The needed increase in active
power generation was, during this and all other studies done
on the Ward-Hale and the AEP-systems, supplied from the
slack bus. Note that two different vertical scales have been
used in figure 2, the left scale was used for the minimum
singular values and the right scale was used for the voltage
magnitudes (given in the per unit system).
The jumps in the numerical value of U,,(.) are due to that
PV-nodes are changed into PQ-nodes when limits for reac-
tive power generation are hit, and hence increases the size of
o,8, , "' 30-node test aystem , ,
0.51 1I-y i
I I
002 O b 4 0 0 8 008 0 1 0 19
Rl i
Figure 2: un(J) and un(Gs) plotted as functions of the load
increase in the nodes (17 - 20), see text
the jacobian matrix. This increase in the dimension of the
jacobian matrix will, as described in section about the sin-
gular value decomposition, reduce the numerical value of the
studied index. The jumps in the proposed index will thus
indicate where generators in the system hit their limitations
for reactive power generation - reflecting the importance of
reactive power resources, which if inadequate can contribute
to voltage stability problems in power systems [6].
AEP 30-node teat syatem
0 35
0.25 O'I
0 5 10 15 20 25 30
bur #
Figure 3: The shifted negative voltage profile (see text) and
vn(Gs) plotted versus bus numbers
Figure 3 shows the right singular vector, v,, of the matrix
GS plotted together with the shifted negative voltage profile
for a specific operating point from the same study as the ex-
ample given above. At the studied operating point an(Gs)
was equal to 0.028 (see figure 2), i.e. the studied operating
point was close to the voltage stability limit. The plotted
'voltage profile' for node i was calculated as (1.0 -Vi), where
Vi is the voltage magnitude (in per units) at node i. The
above presented and other obtained results indicate that the
right singular vector, v,,, is proportional to the above defined
'voltage profile', i.e. the right singular vector is a linear trans-
formation of the voltage profile. I t should be noted that the
above result was found for a constant power load. A more
complex load model, e.g. a voltage dependent load model,
could give another result.
For a real symmetric matrix the minimum singular value
and the minimum of the absolute values of the eigenvalues will
be equal. Further the corresponding left and right singular
vectors will both be equal to the eigenvector corresponding to
the minimum eigenvalue. The reason for the above relations
between a singular value decomposition and an eigenvalue de-
composition of the same matrix can be seen from equations
(5) - (8) in the section about singular value decomposition.
The power flow jacobian matrix is a diagonal dominant real
matrix. The left and right singular vectors will consequently
be close to parallel, and hence also approximately equal to
the eigenvectors of the same power flow jacobian matrix. Fig-
ure 4 shows a plot of the two singular vectors un and vn for
the AEP 57-node test system [18], which confirms that the
left and right singular vectors are almost parallel, and con-
sequently also close to the corresponding eigenvector of the
matrix. I t could also be noted that the eigenvalue problem
for a real symmetric matrix is always well-conditioned [13].
0.15
0.1
0.05
AEP 57-node test svstem
-
0.4 -
-
0.2 ~
-
0.35 I
I
0.25 -
0.2
-
0. 15-
0.1 -
0.05 -
0 20 40 BO BO LOO
Figure 4: The components of un(J) and v,(J)
Numerical results indicate that the sign of uzv, could be
used to determine if the power flow solution is stable or not.
Studies rugged that uzv,, >0 for the high voltage (stable)
solution and U~ V, . <0 for the low voltage (unstable) solution.
The performed numerical studies also confirm the interpreta-
tions made for the minimum singular value and corresponding
singular vectors, i.e. interpretation 1 to 3 at the end of the
section about singular value decomposition of matrices. The
two calculated singular vectors can thus give valuable infor-
mation about the operating point of the system; the last col-
umn of U (the left singular vector) indicates the most critical
disturbance and the last column of V (the right singular vec-
tor) indicate how the state variables will change due to this
disturbance. No numerical difficulties have been encountered
during the performed studies.
Figure 5 shows the Nordic power system, where Sweden has
synchronous connections with Norway, Finland, and parts of
Denmark. A large part of the generation in the Swedish sys-
tem is in the northern part, while the main part of the load
is in the southern part of the country. During spring flood
there is thus a large transfer of power from north to south
through the Swedish main transmission grid. This transfer of
power is limited due to the risk of voltage instability and thus
it is of importance to reliably determine the maximum trans-
fer capacity of the network. During the performed study the
north-south transfer of active power was increased in steps for
the 1033-node model of the Swedish power system, which ba-
sically is the same as the network used for operational studies
on the Swedish main transmission grid. This increase of trans-
fer in active power through the network was accomplished by
an increased loading (both active and reactive parts) in the
southern part of Sweden while a proportional increase in ac-
tive power generation was made at hydro power plants in the
northern parta of the system, which will result in a, realistic
dispatch for the power system.
59
I I
Figure 5: The Nordic power system
1033-node network
o::l
0 3
OL O b 560 I000 I bO 2&lO 2500 3000 3500 4000
D Pt r anr f w
Figure 6: un(J) and voltage magnitudes plotted versus
increase in transfer of active power
Figure 6 shows the minimum singular value for the ma-
trix J, together with four representative (depressed) voltages
magnitudes from a load area in the southern part of the net-
work. The variables are plotted as functions of the increase
in active power transfer from northern to southern Sweden.
Two vertical scales have been used in this figure, the left scale
for the minimum singular value and the right scale for voltage
magnitudes. The large jump in the numerical value of ult(J)
for the 1033-node system (just before the power flow program
no longer converged) is due to that several large generators in
the network hit their limits for reactive power generation at
about the same point. I t can from this study and the other
studies presented in this paper be concluded that the mini-
mum singular value will be a better indicator of the proxim-
ity to the steady state voltage stability limit than the voltage
magnitudes at nodes.
The numerical value of the minimum singular value is not,
by itself only, a measure of the voltage stability margin. The
numerical value depends on the size of the power flow jacobian
matrix, i.e. it is system dependent (see above), as well as the
proximity to the steady state stability limit. Consequently
the scale of the minimum singular value used as a voltage
stability proximity indicator is system dependent. To find
suitable threshold values for a given system isthus a question
which needs further research.
CONCLUSIONS
This paper has presented a fast method to compute the
minimum singular value, together with the corresponding left
and right singular vectors, of a matrix. The proposed method
was used to calculate the minimum singular values for the
power flow jacobian matrix J as well as the matrix Gs.
Results from the performed studies show that the minimum
singular value of the power flow jacobian matrix J or the re-
lated submatrix Gs are good indicators of the proximity to
the static voltage voltage stability limit. The two singular
vectors obtained together with the minimum singular value
give valuable information about the studied operating point
of the power system. The right singular vector, vn, corre-
sponds to sensitive voltages (and angles) and the left singular
vector, U,,, indicates the most sensitive directions for changes
of active and reactive power injections.
Studies were made both on small test systems as well as
on a real sire power system model, results from these studies
show that the algorithm is fast and efficient. I t can, from
the performed studies, be concluded that the presented fast
method for calculation of the minimum singular value of the
power flow jacobian matrix and related submatrices could be-
come a valuable tool for assessment of steady state voltage
stability in power systems.
Acknowledgements
The authors gratefully acknowledges financial support from
Vattenfall (Swedish State Power Board) as well as allowance
to use and technical assistance with the power flow calculation
program REN developed by Vattenfall.
References
A. Cheimanoff and C. Corroyer The power failure of
December 19, 1978, Revue Gincrale de IElectriciti, vol
89, no 4, pp 280 - 296, April 1980.
C. Barbier and J.-P. Barret An analysis of phenom-
ena of voltage collapse on transmission system, Revue
Gdndrale de ldlectricit6, Special CIGRE issue, pp 3 -
21, J uly 1980.
Y. Harmand, et.al. Analysis of a voltage collapse inci-
dent and proposal for a time-based hierarchical contain-
ment scheme, CIGRE Report 38/ 99 - 09,1990.
A.J . Calvaer and E. Van Geert Quasi steady state
synchronous machine linearization around an operating
point and applications, IEEE Transactions on Power
Apparatus and Systems, vol PAS-103, no 6, pp 1466 -
1472, J une 1984.
K. Walve Modelling of power system components at
severe disturbances, CIGRE Report 98 - 18, 1986.
D.J . Hill, P.-A. Liif, and G. Andersson Analysis of long-
term voltage stability, Proceedings of the loth Power
Syrt em Computation Conference, pp 1252 - 1259, GraB,
Austria, August 19 - 24, 1990.
H.K. Clark New challenge: voltage stability, IEEE
Power Engineering Review, pp 33 - 37, April 1990.
R.J. Thomas and A. Tiranuchit Voltage instabilities in
electric power networks, Proceedings of the 18h South-
eastern Symposium on Syst em Theory, Knoxville, TN,
pp 359 - 363, 1986.
A. Tiranuchit and R.J . Thomas A posturing strategy
against voltage instabilities in electric power systems,
IEEE %nsactions on Power Systems, vol 3, no 1, pp
A. Tiranuchit, et .al. Towards a computationally feasi-
ble on-line voltage instability index, IEEE Transactions
on Power Systems, vol 3, no 2, pp 669 - 675, May 1988.
C.L. Lawson and R.J . Hanson Solving Least Squares
Problems, Prentice Hall, Englewood Cliffs, NJ , 1974.
G.H. Golub and C.F. Van Loan M a t h computations,
North Oxford Academic, Oxford, 1983.
G. Dahlquist, A. Bjorck, and N. Andersson Numerical
Methods, Prentice Hall, Englewood Cliffs, NJ , 1974.
J . Guckenheimer and P. Holmes Nonlinear oscillations,
dynamical systemr, and bifircation of vector fields, Ap-
plied mathematical sciences v. 42, Springer-Verlag, 1986.
T. Smed Interaction between high voltage AC and DC
systems Ph.D. Thesis, Royal Institute of Technology,
Stockholm, 1990.
F.R. Gantmacher, The Theory of Matrices, vol 1(2),
Chelsea Publishing Company, New York, NY, 1977.
J .B. Ward and H.W. Hale Digital computer solution of
power-flow problems, AIEE Ransactiona, vol PAS-75,
pp 398 - 404, J une 1956.
L.L. Freris and A.M. Sasson Investigation of the load-
flow problem, Proceedings IEE, vol 115, no 10, pp 1459
87 - 93, Feb 1988.
- 1470, Oct 1968.
BIOGRAPHIES
Per-Anders LBf (S89) was born in Stockholm, Sweden in
1962 and received his Civ.Ing. degree from Royal Institute
of Technology, Stockholm, Sweden in 1987. He is a gradu-
ate student at the Royal Institute of Technology, and he is
currently doing research towards the Tekn.Lic. degree in the
area of long term voltage stability in power systems.
Thomas Smed (s88) was born in Overtorneir, Sweden in
1961 and received his Civ.Ing. and Tekn.Lic. degrees from
Uppsala University, Sweden in 1985 and 1987, respectively
and the M.Sc.E.E. degree from Auburn University, USA in
1989. He is a graduate student at the.Roya1 Institute of Tech-
nology, Stockholm, and he is presently doing research towards
the Ph.D. degree in the area of interaction between AC and
DC power systems.
GBran Andersson (M86) was born in Malmo, Sweden in
1951 and received his Civ.Ing. and Ph.D. degrees from Lund
Institute of Technology, Sweden in 1975 and 1980, respec-
tively. He is Professor and Head of Department of Electric
Power Systems, Royal I nstitute of Technology, Stockholm.
David J Hill (M76) received his B.E. (Electrical Engineer-
ing) and B.Sc. (Mathematics) degrees from the University
of Queensland, Australia in 1972 and 1974, respectively. In
1976 he received his Ph.D. degree in Electrical Engineering
from the University of Newcastle, Australia. He is currently
Associate Professor in Electrical Engineering and Computer
Science at the University of Newcastle.
61
tion, where progressively more units were restricted. This caused an
accelerating rate of fall of EHV voltages, which culminated in power
systemcollapse. It was only possible to simulate the second fast phase, by
carefully monitoring the MV Ar output of all rotating units and being
aware of the time versus overcurrent characteristic of rotor protections.
The usefulness of the voltage stability index that is described in your
paper depends very much on the assumptions that are made in the load
flow studies. As this index would provide a guide to stability limit, it
should be related to the second fast phase. At this time, transformer tap
changing should have restored load magnitudes, so that constant load
could provide a reasonable assumption, if transformers to at least sub-
transmission voltage levels were to be included. (This assumptions would
eliminate the complexities associated with the different load-volage charac-
teristics of the different load components.) Two otehr variables should also
be included, the effect of increased MV Ar demand at substations with
reduced EHV levels, and the effect of field forcing on generators at their
rated excitation.
By this approach, the effects of these variables for a critical operating
situation, would allow the range of index values to provide a better guide
to the dynamic effects encountered with systemvoltage stability.
References
[l ]
W. R. Lachs & D. Sutanto. Voltage Instability in Interconnected
Power Systems: A Simulation Approach 90 IC 590-0 IEEE Power
Meeting New Delhi, Nov. 1990.
W. R. Lachs. System Reactive Power Limitations Paper A79
015-9 IEEE Winter Power Meeting, New York, 1979.
[2]
Manuscript received January 31, 1991.
Discussion
W. R. Lachs, (University of New South Wales, Kensignton, NSW,
Australia): The authors are to be commended on presenting a computa-
tionally fast indicator of voltage stability which can be made as a direct
adjunct to load flow studies. Load flows are the bread and butter of power
system planning and operational studies, so that the voltage stability
indicator should beof value to the power industry.
In the 1983 Swedish incident, which was due to systemvoltage instabil-
ity, the authors describe the 50 second delay after the disturbance, before
cascade line tripping culminated in collapse. This was similar to other
incidents, when there has been a first slow phase followed by a second,
fast disruptive phase during which existing protections could not control
events, and if they did function, accelerated the demise of the power
system. The key aspect of systemvoltage instability is that it follows the
breakdown of reactive power control on significant areas of a power
system. In a critical situation, it is the increase of series losses, in all
transformers and heavily loaded EHV transmission lines which causes the
unsustainable reactive power demand. The authors must becommended in
that their studies examined a 1033 bus network, which would have
included many of the systems transformers so that the important influence
of their series reactive power losses is completely lost, raising doubts
about such studies.
To better understand the method used would the authors comment on
the following points:
1. What voltage levels, in addition to 400kV and 220kV, were included in
2. At what voltage levels were loads connected?
3. What load-voltage characteristics were assumed?
4. Was provision made for automatic transformer tap-changing?
5 . What were the situations examined in studies (e.g. heavy load period,
any 4OOkV line outages, unfavourable generating array) with variable
power transfers fromthe Northern generation?
Load flows are a very useful tool in studying systemvoltage stability. A
series of load flow studies have been able to replicate both the initial slow
phase and the second fast disruptive phase, following a disturbance which
has led to systemvoltage instability and collapse 111. In this simulation the
effects of transformer tap changing, transformers losses, and generator
field forcing were included and this provided a good replication of the first
slow phase and the second fast phase.
To limit the size of the network under study, preliminary studies
examined a load area, including all transformation steps down to distribu-
tion voltage levels at which all loads were connected. These studies
modelled cables, lines, shunt capacitors, automatic transformer tap chang-
ing (with altered timer settings at each voltage ratio) and different load-
voltage characteristics. The effect of a sudden 6% voltage reduction at the
EHV substation busbar was examined and the following was found [2]:
Before the tap changing commenced, the voltage drop reduced load
magnitudes, as well as line, cable and capacitor charging. As a conse-
quence, at the EHV substation, the MW demand had reduced but the
MV Ar demand had increased.
Tap changing gradually ovverrode load-voltage characteristecs and pre-
disturbance load magnitudes were regained by 1 minute.
Tap-changing on all transformers in series raised distribution voltages
above pre-disturbance levels before distribution transformer tapping
could be reversed.
. This caused loads to exceed their pre-disturbance magnitudes, with a
maximumovershoot at 2 minutes.
. When distribution voltage levels are first restored, sub-transmission
voltage levels are still low. Consequently their series reactive power
losses are higher and line, cable and capacitor charging are lower. At
this time the EHV substation MW demand is restored but the MV Ar
demand exceeds its pre-disturbance magnitude.
The results of the preliminary studies could then be incorporated into
the voltage stability simulation studies which included power stations, the
transmission grid and the transformers at all EHV substations. After any
critical disturbance, only a number of EHV substations had suffered large
voltage reductions, and at these substations the loads were varied with
time to correspond to the functioning of transformer tap changers. In fact
the sole cause of dynamic change during the slow first period is due to the
simultaneous functioning of transformer tap changers with identical timer
settings.
The second, fast phase of system voltage instability was found to
commence when the first generator had its excitation restricted by its rotor
current protection. Once one unit was restricted, it started a chain reac-
the 1033 bus network?
M. K. Pal ( Publ i c Ser vi ce El ect r i c and Gas Company,
Newar k, N . J . ) : Al t hough t he use of t he mi ni mumsi ngul ar
val ue of t he f ul l or r educed power f l ow J acobi an as a
measur e of vol t age st abi l i t y i ndex i s not a new
concept , t he gener al di scussi on on t he si ngul ar val ue
decomposi t i on and t he pr ocedur e f or cal cul at i on of t he
mi ni mum si ngul ar val ue as gi ven i n t he paper i s
i nt er est i ng and enl i ght eni ng. Not e however , t hat t he
mi ni mum si ngul ar val ue of t he J acobi an may not al ways
ser ve as a good i ndi cat or of t he pr oxi mi t y t o vol t age
col l apse. The i nt er pr et at i on gi ven by t he aut hor s i s
not f ul l y convi nci ng.
Si ngul ar i t y of t he power f l ow J acobi an si gni f i es
st eady- st at e l i mi t , beyond whi ch a sol ut i on woul d not
exi st . Over t he year s, var i ous at t empt s have been made
t o equat e t he si ngul ar i t y of t he st eady- st at e power
f l owJ acobi an, or some ot her syst emJ acobi an, wi t h t he
vol t age st abi l i t y l i mi t . The f act t hat t he si ngul ar i t y
of t he J acobi an can be a val i d i ndi cat or of vol t age
st abi l i t y l i mi t has not been r i gor ousl y demonst r at ed,
al t hough many cl ai ms have been made t o t hat ef f ect . A
t hor o r evi ew r eveal s t hat t hese l ack mat hemat i cal
ri gory v. One i mpor t ant f act or was mi ssi ng i n most of
t hese wor ks. The l oad dynami cs, t he dr i vi ng f or ce
behi nd vol t age i nst abi l i t y, was not i ncl uded i n t he
anal yses.
The aut hor s t al k about t he i mpor t ance of l oad
model i ng i n vol t age st abi l i t y st udi es i n t he
i nt r oduct i on; but not hi ng i s done i n t he mai n body of
t he paper . The ent i r e paper i s based on const ant power
l oad. I t i s not di f f i cul t t o show t hat f or such l oads,
i f const ant gener at or t er mi nal vol t age i s assumed,
cor r espondi ng t o i nf i ni t e speed of exci t at i on cont r ol ,
t her e can be no vol t age i nst abi l i t y. On t he ot her
hand, i f gener at or r ot or f l ux dynami cs and f i ni t e
exci t at i on cont r ol speed ar e t aken i nt o account ,
vol t age st abi l i t y l i mi t i s wel l bel ow t he power l i mi t
obt ai ned f r om a power f l ow model . Al so, wi t h t hi s
model , oper at i ons at mul t i pl e power f l ow sol ut i on
poi nt s ( t he l ow vol t age sol ut i ons) wi l l pr ove t o be
st abl e. Of cour se, t her e i s a f al l acy her e. Ther e i s
no such t hi ng as const ant power st at i c l oad.
Ther e i s al so t he quest i on of l ar ge di st ur bance
62
vol t age st abi l i t y. A s acknowl edged i n t he i nt r oduct i on
of t he paper , t he i ni t i at i ng event i n most known
vol t age col l apse i nci dent s was a l ar ge di st ur bance,
such as l oss of a heavi l y l oaded l i ne or l oss of a
l ar ge gener at or . Do t he aut hor s have any opi ni on about
how t o handl e t hi s i ssue. The aut hor s' met hod wi l l
onl y addr ess st eady- st at e l i mi t .
The t r eat ment of a gene' rat or bus as a P, Q bus under
r eact i ve l i mi t i s not cor r ect and can pr oduce gr ossl y
er r oneous r esul t s. At r eact i ve l i mi t , t he gener at or
t er mi nal vol t age may st i l l be under exci t at i on cont r ol ,
except t hat t he vol t age r ef er ence i s r eset . I n t hat
case t he gener at or bus shoul d be model ed as a PV bus.
I f , on t he ot her hand, t he gener at or t er mi nal vol t age
i s not mai nt ai ned by exci t at i on cont r ol , i . e. i f t he
gener at or i s oper at i ng at const ant exci t at i on, a
modi f i cat i on t o t he power f l ow model woul d be
necessar y. I t can be shown t hat under t hi s condi t i on,
t he vol t age st abi l i t y l i mi t i s det er mi ned by const ant
vol t age ( gener at or f i el d vol t age) behi nd gener at or
synchr onous r eact ance.
Have t he aut hor s compar ed t hei r mi ni mum si ngul ar
val ue appr oach wi t h a sensi t i vi t y appr oach? I t i s
known t hat f or oper at i ng poi nt s wel l wi t hi n t he st eady-
st at e l i mi t , t he sensi t i vi t y el ement s ar e ver y cl ose t o
t he el ement s of t he bus i mpedance mat r i x. As t he
st eady- st at e l i mi t i s appr oached, t he sensi t i vi t y
val ues i ncr ease r api dl y. A di r ect compar i son of t he
sensi t i vi t y el ement s wi t h t he el ement s of t he bus
i mpedance mat r i x shoul d, t her ef or e, of f er a ver y r api d
t est f or pr oxi mi t y t o t he st eady- st at e power l i mi t .
Gener al l y, val ues of t he sensi t i vi t y el ement s l ess t han
about t wi ce t he val ues of t he el ement s of t he bus
i mpedance mat r i x shoul d assur e suf f i ci ent mar gi n.
Si nce t hi s appr oach can al so t ake f ul l advant age of
spar si t y, i t coul d pr ove t o be mor e ef f i ci ent ,
especi al l y si nce t he scal i ng pr obl em wi l l not ar i se,
and t he pr obl embuses wi l l be aut omat i cal l y i dent i f i ed.
The aut hor s' comment woul d be appr eci at ed.
[ A] M. K. Pal , Di scussi on of "An I nvest i gat i on of
Vol t age I nst abi l i t y Pr obl ems. " by N. Yor i no, H.
Sasaki , Y. Masuda, Y. Tamur a, M. Ki t agawa and A.
Oshi mo, 91 WM 202- 2 PWRS, I EEE/ PES wi nt er meet i ng,
Feb. 1991.
Manuscr i pt r ecei ved Febr uar y 19, 1991.
J-C CHOW AND R. FISCHL (Drexel University, Philadelphia,
PA): The authors are to be congratulated for developing a method
for computing the minimum singular value of the Jacobian
efficiently. We have a few comments and questions and would
appreciate the authors' response:
1. In the computation of o(Gs), algorithm 3 is almost the same as
the two previous algorithms except for step 1, the Vn,i =0 for
all i corresponding to the angle in vn(l). Specifically, It seems
that algorithm 3 is the same as Algorithm 2 if the matrix J is
replaced by Gs.
2. Could the authors give more information on the definition of the
Ao in algorithm 2 and 3. In our opinion, the Ao is either
A&) =I &k-1)-&) I or
A&) =I &k-l)-afa-3) I or
A&) =I 0( 2) - 0( 2- 2) 1.
Moreover, it would be useful to plot o(2k-1) and d2k) as a
function of iterative step k.
3. Can the authors comment on the rate of convergence of the
algorithm as a function of the system size and structure?
4. Have the authors examined the performance of their algorithm
for the system loading which give non-convergent power flow
equations?
T. Van Cutsem, University of Likge, Liege, Belgium :
the authors have proposed an interesting algorithm for efficiently eval-
uating how close to singularity the load flow J acobian matrix is. This
discussion aims at making a brief comparison with some other meth-
oJs for evaluating system robustness with resped to voltage collapse.
The most essential quality expected from a Voltage Collapse Prox-
imity Indicator (VCPI) is obviously to be a "good measure" of the
distance to voltage collapse, which requires i n particular to take into
account the "precipitating effect" of generator reactive power limi-
tations. Within this respect, the method described in this paper is
of the "sensitivity" type in the sense that the system behaviour is
analyzed around the current operating poinf. In particular, reactive
power limitations are reflected. but not in anticipation. Figure A be-
low attemps to better capture this : around the operating point 0.
both the real system (with reactive limitations) and the hypothetical
one (without limitations) have the same behaviour. Clearly, the first
one is weaker but a local analysis around Point 0 does not reveal any
difference.
" t
without limits
hi ts
I
Figure A
In Ref. [A] below, the adopted viewpoint is that a direct determi-
nation of the collapse point (C' in Fig. A) provides a better measure
of the system robustness (at point 0) by anticipating the effect of
reactive power limitations.
Now, as far as a sensitivity method is deemed adequate, i t is in-
teresting to compare the authors' method with another sensitivity
method, namely the aQ,/aQ, index. initially proposed in Ref. [ B]
below. This index is defined as the reactive generation increment re-
quired to obtain one'more Mvar at each bus of the system. Denoting
by Q,(O,V) the sum of all reactive productions in the system, the
above sensitivity is simply obtained by solving
which yields :
of which only the components of the second sub-vector relative to
the load buses are used. This index can be interpreted as the inverse
of a reactive power transportation efficiency, tending towards infinity
at some buses as the critical point is approached for these buses.
As such a sensitivity is obtained at every load bus (directly). the
method does not merely provide a single number to characterize the
whole system but rather a bus-by-bus indication of the system ro-
bustness.
On the other hand, the numerical reliability of the above compu-
tation is at least as good as the one of the authors' method since
both use the L and U factonof J (close to the maximum load point,
J becomes ill-conditionned and unrealistic aQ,/aQ, indices can be
found.. . ).
63
As regards t he computational efficiency however, t he effort r e-
quired by t he aQ,/aQt criterion amounts t o a single substitution i n
the available factorized system. i.e. it amounts t o about one hal f
i terati on of t he authors algorithm.. .
Hence. as far as a sensitivity method i s deemed adequate. t he
older aQ,/aQ, criterion seems t o remain competitive. Could the
authors further develop on t he advantages offered by t he minimum
singular value computation wi th respect t o t he above criterion ?
[A] T. Van Cutsem, A method t o compute reactive power mar-
gins wi th respect t o voltage collapse , Paper No. 90 WM 097-6
PWRS. presented at t he IEEE PES 1990 Winter Meeting, Atl anta,
USA. Feb. 4-8. 1990. t o appear i n IEEE Trans. on PS.
[B] J. Carpentier. R. Girard. E. Scano. Vol tage collapse proximity
indicators computed from an opti mal load fl ow , Proc. of the 8th
PSCC Conf.. Aug. 1984, pp. 671-678.
Per-Anders Lof, Thomas Smed, Goran Andersson, and David J. Hill:
The authors wish to thank the discussers for their interest in the paper and
their valuable comments.
M. K. Pal:
We agree that the minimum singular value applied to the power flow
J acobian matrix, J, is an indicator of steady-state stability in the power
system; in our paper the submatrix G, is presented as indicative of voltage
stability problems. The minimum singular value of the submatrix G,
could be used to construct a special static voltage stability indicator, since
under the assumption that detFe f 0, the matrix J is singular if and only
if the submatrix G, is singular. We thus propose the use of U, ( G, ) for
voltage stability analysis and UJJ) for studies of steady-state stability
problems due to either angle or voltage problems.
The intention of the statement that there can be no voltage instability if
the loads are constant power loads and the generator terminal voltage is
assumed to be constant is not clear. Of course, this is not true in
steady-state as the case with one synchronous generator connected by a
reactance to a constant load illustrates. But the dynamic behavior is a
separate question not addressed in our paper except insofar that a dy-
namic load flow model can be used. Our calculations show that the claim
is ttue for dynamic stability with constant loads. We certainly agree that it
is important ultimately to consider both questions.
The proposed method to scan the numerical value of the minimum
singular value while the loading in the systemis increased allows for any
(static disturbance to be added to the power systembefore the power flow
calculation is performed and the minimumsingular value is computed. We
can use un to compute loadability limits or a reactive margin and so allow
for large power disturbances. But again our paper does not address the
transients associated with the disturbance and of course this could be
important.
We agree that the modelling of limits for the reactive power generation
at generators should be done more accurately. We are investigating the
limits being modelled as voltage dependent. This voltage dependence
could be easily included in the used power system model, which is the
basis for the computation of the minimumsingular value. This index could
naturally be used side-by-side with sensitivity information as proposed by
Schlueter et al. [ A] . We have not compared sensitivity elements with the
elements of the bus impedance matrix, but this approach clearly needs to
be verified under stressed network conditions.
3. The bads were all modelled as constant power loads, i.e., fixed p and
Q loads.
4. At approximately 40 of the almost 150 transformers provision was
made for automatic tap-changing, a routine performed automatically by
the used power flow program. In the used network model there were
normally one or two levels of transformers between the main grid and
the voltage levels at which the loads were connected.
5. The used network data was not chosen to represent a true picture of
the conditions in the Swedish system, but to show that the proposed
method is a fast and efficient way to compute the minimum singular
value for a large power systemmodel. The increased transfer of active
Power Over a Vital network section, during the study, caused a number
of 400 and 220 kV lines to become heavily loaded.
The power system model used for the calculation of the minimum
singular value and the two corresponding singular vectors is the same as
the one used for the power flow calculation. The proposed algorithm is
based on a standard power flow program and will thus have the same
possibilities in modelling of power systemcomponents as the preceeding
power flow calculation. We agree that in the scenario of the discussion,
the minimum singular value (proposed as a static voltage stability index)
refers to the second phase. It will hence be of importance to include the
effects of limited reactive power resources as stated in the discussion.
T. Van Cutsem:
It is true that in our index the reactive power limitations are reflected,
but not in anticipation. However, the main use (today) of a static
voltage voltage stability index based on the minimumsingular value would
probably be long-term planning and operational planning studies carried
out off line. With a little extra effort beyond that needed to compute
reactive power margins directly, the proposed index will also show the
trend to collapse, see figure 6 where the four plotted voltages show little
response in their magnitudes while the minimum singular value of the
power flow J acobian matrix actually approaches zero and hence the
steady-state stablility limit of the power system.
During the computation of the minimumsingular value the two corre-
sponding (left and right) singular vectors will also be found without any
extra computational efforts; see the proposed algorithms. As shown in the
paper the two singular vectors (U, and vn) can give us valuable sensitivity
information about the studied operating point of the power system. The
jumps which will occur during the scanning of the stability index when
limits for the reactive power generation are hit will indicate important
changes in the system characteristics. This information could become a
hasis for how to plan a reactive power resources or take preventive actions
to save the systemfromvoltage collapse.
The discussed sensitivity aQ, /aQ, is seen as a valuable complement to
the static stability index based on minimumsingular values proposed in the
paper. While we believe that the minimumsingular value applied to either
of the two matrices J or G, will better show the trend when the power
systemis loaded towards the collapse point, the sensitivity index aQ, /aQ,
will (probably) be better for studies of where to add reactive power
J-C Chow and R. Fischl:
The matrix G, which was defined as G, :=G, - G,FiF, could be
considered in a generalized way to be a submatrix related to the power
flow J acobian matrix that gives the voltage sensitivities with respect to
changes in the reactive power injections at constant active power loading.
It should benoted that while the power flow J acobian matrix J is a sparse
matrix, the matrix G, will be a non-sparse matrix due to the influence of
the second term in the above definition. Both the sparsity of the power
flow J acobian matrix and the special structure of the related submatrix Gs
were fully utilized in the presented algorithms.
support.
1. It would, of course, be possible to use algorithmtwo in the Paper and
apply it directly to the matrix G, . ~~wev er , due to the non-sparsity of
this matrix, i t would not be recommendable. Therefore, algorithm
W. R. Lachs:
In response to the specific questions about our network model, we firstly
note that the used power systemmodel consisted of 1033 nodes, over 1100
transmission lines, and almost 150 systemtransformers.
1. The network included several voltage levels and the main ones were
400, 220, 130, and 70 kV. There were also some special voltage
levels included in the model (mainly lower voltage levels). e.g., at
nuclear power stations and at some transformations to the Swedish
railway systemas well as a few 300 kV nodes connecting the Swedish
network to an equivalent of the Norwegian power system.
2. Loads were, of course, connected at different voltage levels; the main
parts of the load in the studied systemwere connected either to the 130
or the 70 kV level.
64
three was specifically designed to compute the minimumsingular value
and the two corresponding singular vectors of the matrix G, without
destroying the sparsity of power flow J acobian matrix.
2. The numerical value of Ao is only used in algorithms two and three to
stop the iterative procedure when the preset tolerance has been achieved.
It is computed as the difference between two consecutive minimum
singular values, i.e., A o is calculated according to the following
formula:
Ao( k ) =I o ( 2 k - 1) - ,,(2k) 1 ,
3. The rate of convergence will (at least) theoretically have the same
properties as inverse iteration, which in its turn has the same conver-
gence properties as Newton iterations, i.e., quadratic convergence.
Theoretically, the inverse power method guarantees convergence to the
smallest eigenvalue if the quotient (the convergence ratio) between the
smallest and the second smallest of the absolute values of the eigenval-
ues are less than one and the initial guess at least contains some
component of the corresponding eigenvector [B] . Practically the con-
vergence has been good; but the dependence on system size and
structure needs closer examination.
4. No, the proposed method to quantify how close a particular operating
point is to the steady-state stability limit needs the power flow J acobian
matrix of the converged operating point. The convergence rate for the
proposed algorithmis, in general, at least not worse for stressed power
systems than for lightly loaded networks. Thus, it is possible to study
the profile of collapse until very close to non-convergence.
References
R. A. Schlueter et al. Methods for determining proximity to
voltage collapse, IEEE Transactions on Power Systems, pp.
285-292, vol. 6, no. 1, Feb. 1991.
G. Peters and J . H. Wilkinson. Inverse iteration, ill-conditioned
equations and Newtons method, SIAM Review, pp. 339-360,
vol. 21, no. 3, J uly 1979.
[I ]
[2]
Manuscript received September 17, 1991.