Paper accepted for presentation at 2003 IEEE Bologna Power Tech Conference, June 23th-26th, Bologna, Italy

Stochastic OPF by Constraint Relaxation

Lucia M. Kimball, Member IEEE, Kevin A. Clements, Fellow IEEE,
Slobodan Pajic, Student Member IEEE, Paul W. Davis Member IEEE

Abstract— This paper describes an algorithm for the so- linear and a nonlinear formulation of the stochastic OPF.
lution of the stochastic optimal power flow problem. The Linear examples illustrate the extraordinary economy of
stochastic OPF minimizes the total cost of a base case op-
erating state plus the expected cost of recovery from con- applying constraint relaxation to the optimization.
tingencies such as line or generator outages. The algo-
rithm presented is based on a relaxation method in which II. Linear problem formulation
constraints are modeled only as they become active. This
method efficiently models all transmission constraints using The stochastic OPF for a linear DC network can be for-
a rank-one update to the distribution matrix. Numerical mulated in the following way
examples illustrate the dramatic reduction in problem size
achieved in practice by constraint relaxation. They also con- K
X
firm that the expected cost of a dispatch that accommodates Minimize cT0 u0 + pk cTk uk (1)
failures according to their probability of occurrence is less
than the cost of a preventive operating mode. k=1

Keywords— Optimal Power Flow, Contingency Analysis, Subject to g0T u0 = pd (2)

Relaxation Methods D0 u0 ≤ d0 (3)
gk u0 + hTk uk
T
= pd (4)
I. Introduction
Dk u0 + Ek uk ≤ dk (5)
Optimal power flow (OPF) is a class of tools used for k = 1, · · · , K
both the operation and planning of power systems. The
simplest OPF formulation is as a one-stage optimization where u0 are the pre-contingency control settings, uk are
problem in which the performance of some base-case sit- the post-contingency control adjustments for contingency
uation is optimized with respect to a specified objective k, pk is the probability of contingency k.
function such as cost of generation or losses. The objective function (1) includes the total cost of op-
The basic OPF formulation has been expanded in or- eration in the pre-contingency or base case as well as the
der to recognize contingency conditions [5], [7]. In the ex- expected cost of recovery from all contingencies. Con-
panded formulation, it is required that the power network straint (2) represents the power balance constraints for the
be able to operate under a specified set of contingencies base case, and (4) enforces power balance for the post-
such as generation outages and line outages. Typically the contingency states. Constraints (3) and (5) include trans-
base-case costs are higher when contingency constraints are mission constraints and bounds for all control variables.
imposed because the system is operated in a more conser- In the post-contingency case, these bounds include realis-
vative manner; e.g. efficient generators may be scheduled tic limitations on the post-contingency adjustments repre-
to have lower outputs than they would have when the con- senting the system capability to recover from the contin-
tingency constraints are not modeled. gency, in particular ramp-rate limits. (The coupling effects
It is widely recognized that it is impossible to model of ramp-rate limits will be shown explicitly in section III
all possible contingencies. As a consequence, only the for the nonlinear OPF.)
most likely contingencies are usually accounted for. Most The efficient solution of the stochastic OPF represented
contingency-constrained OPF algorithms can be considered by (1–5) is accomplished via an interior point method. The
preventive since they impose additional constraints to allow Lagrangian with logarithmic barrier function (to enforce
for a feasible operating state in the event of a contingency nonnegativity of the slack variables si ) for the interior point
but do not take into account the ability to change con- method is
trol settings in the event of a contingency. An approach
that allows for controls to be reset after the occurrence of L = cT0 u0 + λT0 (g0T u0 − pd ) + π0T (D0 u0 + s0 − d0 )
a contingency is known as a corrective method. OPF with XK 
post-contingency corrective rescheduling was first investi- + pk cTk uk + λTk (gkT u0 + hTk uk − pd )
gated in [5] using a Bender’s decomposition. However, the k=1
probability of occurrence of the various contingencies is not

+ πkT (Dk u0 + Ek uk + sk − dk )
directly modeled in either [5] or [7].
The stochastic OPF presented here was first proposed in X Nk
K X

[2]; the optimization is undertaken as a two-stage stochastic −µ ln sk,i (6)

programming problem. In this paper, we present both a k=0 i=1

This work was supported by NSF grants ECS-9810288 and ECS- Here µ is a nonnegative barrier parameter and Nk is the
0086706. length of vector sk . The Karush-Kuhn-Tucker or first-order

0-7803-7967-5/03/$17.00 ©2003 IEEE

necessary conditions are defined by the zero of the gradient fk (xk , uk ) ≤0 (19)
system h(u0 , uk ) ≤0 (20)
K
X k = 1, · · · , K
∇u0 L = c0 + g0T λ0 + D0T π0 + gkT λk + DkT πk (7)
k=1 where x0 and xk are state variables for the base and con-
∇λ0 L = g0T u0− pd (8) tingency cases, respectively, and u0 and uk are the pre- and
post-contingency control settings.
∇ π0 L = D0 u0 + s0 − d0 (9)
The set of equality constraints consists of power balance
∇ s0 L = π0 − µS0−1 e (10) equations for the base case (16) and each contingency case
∇uk L = pk ck + hTk λk + EK T
πk (11) (18) (cf. (2) and (4)). The inequality constraint set con-
T T tains transmission constraints and bounds for all control
∇λk L = gk u0 + hk uk − pd (12)
∇ πk L = Dk u0 + Ek uk + sk − dk (13) variables for the base case (17) and each contingency case
(19) (cf. (3) and (5)). The last set of inequalities (20) are
∇ sk L = πk − µSk−1 e (14) coupling or ramp-rate constraints; these are incorporated
for k = 1, . . . , K into (5) in the linear formulation but are shown separately
here to illustrate how they are accommodated.
where Sk = diag(sk ), k = 0, . . . , K. The solution of this
Sequential quadratic programming coupled with an inte-
system can be approximated using Newton’s method.
rior point method, as proposed in [6], can be used to solve
The size of the system (7–14) is determined in part by
this optimization problem. The Lagrangian for (15–20),
the product of two large numbers, the number K of contin-
with nonnegativity constraints again imposed on the slack
gencies being modeled and the number of inequality con-
variables si and σ through a barrier parameter, is
straints in (3) and (5). Of course, contingency planning
requires modeling a substantial number of line failure and
L = c(x0 , u0 ) + λT g(x0 , u0 ) + π0T (f (x0 , u0 ) + s0 )
generator outage contingencies. Furthermore, there are at
k 
least as many inequality constraints as there are transmis- X
+ λTi g(xi , ui ) + πiT (fi (xi , ui ) + si )
sion line flow constraints. Hence, the system (7–14) can be
i=1
dauntingly large. 
Fortunately, the amount of computational effort required + pi dTi ui + γ T (hi (u0 , ui ) + σi )
to solve the Newton system can be greatly reduced by an nc nc k
K X nr
!
active set method in which transmission constraints are
X X X
−µ ln si + ln sik + ln σi
not introduced until they are violated. Since only a small i=0 k=1 i=1 i=1
percentage of the total transmission constraints usually be-
come active, the size of the system is greatly reduced. (Compare (6).)
In addition, rows of the distribution matrices included in A stationary point of the Lagrangian function is a zero of
the matrices D0 and Dk are only created as needed. A spe- the following system of KKT conditions from the interior
cially tailored method for creating rows using a rank-one point formulation:
update expedites those computations. This critical tech-
nique is described in appendix VI. ∇x0 L = ∇x0 c(x0 , u0 ) + GTx0 λ + FxT0 π0
The structure of the actual Newton system that is the
surrogate for the KKT equations (7–14), the role of time ∇u0 L = ∇u0 c(x0 , u0 ) + GTu0 λ + FuT0 π0 + HuT0 γ
coupling constraints, and the precise placement of active ∇xk L = GTkx λk + FkTx πk
k k
inequality constraints are all described in the following sec-
∇uk L = GTku λk + FkTu πk + HuTk γ + pk dk
tion for the general nonlinear stochastic OPF. k k

∇λ L = g(x0 , u0 )
III. Nonlinear problem formulation
∇ π0 L = f (x0 , u0 ) + s0
The nonlinear stochastic OPF for an AC network can be ∇λk L = gk (xk , uk )
formulated in a similar framework, a single optimization
problem which includes a base case and a set of contin- ∇ πk L = fk (xk , uk ) + sk
gency cases coupled with ramping constraints. With the ∇γ L = h(u0 , uk ) + σ
generator ramping constraints shown explicitly, this more ∇ s0 L = π0 − µS0−1 e
general formulation can be written ∇ sk L = πk − µSk−1 e
K
X ∇σ L = γ − µΣ−1 e
Minimize c(x0 , u0 ) + pk ck (uk ) (15) s0 ≥ 0, sk ≥ 0, σ ≥ 0
k=1
for k = 1, · · · , K
Subject to g(x0 , u0 ) =0 (16)
f (x0 , u0 ) ≤0 (17) where Σ = diag(σ) and e is a vector of ones of appropriate
gk (xk , uk ) =0 (18) dimension. (Compare (7–14).)
 TABLE I
To form the outer loop in the solution algorithm, ap-
Test cases
ply a Newton linearization by expanding the KKT equa-
tions about x0 , u0 , xk , uk . Several more steps then lead
to a small system for the inner loop. First, eliminate al- case IEEE system Number of contingencies Nc
gebraically the state variables ∆x0 and ∆xk and the mul- 1 30 bus 6
tipliers λ and λk . Then linearize the remaining reduced 2 118 bus 53
system about s0 , sk , and σ. Finally, eliminate the con-
trol variables ∆u0 and ∆uk . The resulting bordered-block
TABLE II
diagonal system to be solved in the inner loop is
Problem dimensions
   
∆π0 r0
V0T
 
C0  ∆π1   r1 
 C1 V1T      case Nc Total coupling constraints
   ∆π2   r2 
 . .. .
..   .  =  . 
    1 6 162
  ..   .. 
 

T 
2 53 7579
 CK VK  
 ∆πK   rK 

V0 V 1 · · · V K M
∆γ rγ
(21) approach is to set the base case to accommodate all con-
Here the block matrix C0 corresponds to the base case, tingencies without resetting any control. In these simple
the blocks Ck , k = 1, · · · , K, correspond to each of the examples, in which the failure probabilities are uniformly
contingency cases, and the bordering blocks Vk arise from distributed, the stochastic approach is at least 2% less ex-
the generator ramping constraints that couple the subprob- pensive than the conservative approach.
lems.
Potentially, each diagonal block in (21) is as large as V. Conclusions
the number of all the line flow and control variable con-
straints in a single case, and there is one more block than Stochastic OPF is challenging because each contingency
the number of contingencies; (21) could be enormous! But considered potentially introduces a new problem as large as
constraint relaxation limits the entries in each ∆πk to just the original but with a different distribution matrix. This
the constraints active for contingency k, a number which paper presents an algorithm that introduces two significant
is typically quite small compared with the size of the base computational efficiencies,
case problem. The actual solution of this (significantly) • restricting attention to a small set of violated constraints,

smaller bordered-block diagonal system can be obtained and

by an elimination process such as that described in [8]. • testing for newly violated line-flow constraints for each
different line-fault contingency without explicitly comput-
IV. Numerical examples ing the distribution matrix for that contingency.
The IEEE 30 bus and 118 bus networks provide examples The first is accomplished by constraint relaxation, the
illustrating this approach. The test cases are described in second by using rank-one updates to compute post-
table I. Generation costs as a function of MW output are contingency power flows.
similar to those used in [6].
The challenge of the stochastic OPF is the large dimen- VI. Appendix: Efficiently testing line flow
sionality of the associated nonlinear programming problem. constraints
When system constraints include time coupling effects such A key implementation challenge is recalculating network
as ramp constraints, the problem size increases dramati- line flows for each line failure contingency. A sparse rank-
cally given a small increase in the size of the system; see one update technique, which we describe here for the DC
the comments following (21). The number of time coupling network, is an effective solution. Note that the DC network
constraints for each test case in detailed in table II. matrix is included in that portion of the coefficient matrix
The efficiency of the relaxation method for solving the of the inequality constraints (5) which represent line flow
stochastic OPF relies upon the fact that a small number limits. The same approach can be applied to the p-θ and
of the total constraints become active during the solution q-V matrices in a fast decoupled AC model.
process. As table III shows, in each case less than 9% of the
total set of time coupling constraints become active. Con-
TABLE III
straint relaxation dramatically reduces the effective prob-
Percentage of active constraints
lem size.
The output of the stochastic OPF is a set of control vari-
able settings that can be changed within system constraints
to produce secure and economical settings in the event of case Total coupling Number active Percent of total
a contingency. These settings have been chosen to mini- 1 162 14 8.6%
mize the sum of the base case cost and the expected cost 2 7579 193 2.5%
of recovery from the contingencies. A more conservative
 In the DC network model the bus angles are related to where D is saved from the factorization (22) of the base
the real power injections via case network B matrix.
γl
Bθ = p pl = pl + eTij D−1 emn eTmn p = pl + βl δl (23)
xl
where θ is the vector of bus angles and p is the vector of
real power injections. Furthermore, the real power flow on where
branch l connecting buses i and j is given by eTij D−1 emn
βl =
θi − θj xl
pl =
xl δl = γl eTmn p
where xl is the reactance of line l. Since Equation (23) expresses the post-contingency power flow
θ = B −1 p, pl as the sum of the pre-contingency flow pl and an update
vector βl δl . The update vector is computed using factors
we have of the base case network B matrix and employing sparse
1 1 vector methods.
pl = (ei − ej )T B −1 p = eTij B −1 p
xl xl Consider a multiple line outage contingency in which k
where ei is a vector containing all zeros except a 1 in loca- lines are removed. Assuming that line k connects buses mk
tion i and eij = ei − ej . With B factored as and nk , the modified B matrix is written as

B = U T DU (22) B = B − EX −1 E T .

and eij = ei − ej we can write If emk nk is a vector containing a 1 in location mk and −1

in location nk then E = (em1 n1 em2 n2 · · · emk nk ) and
1 T
pl = e p
xl ij
 1
0 ··· 0

x1
1
where  0
x2 ··· 0 
−1
X = .
 
. .. .. 
U T eij = eij  .. .. . . 
1
U T Dp = p 0 0 · · · xk

Note that one forward substitution and one divide-by- where xl is the reactance of line l. Applying the Sherman-
diagonal suffices to compute p. Computation of eij requires Morrison-Woodbury matrix inversion theorem to B we get
one fast forward substitution. Calculation of pl requires −1
two sparse dot products since eij is sparse. B = B −1 + B −1 E(X − E T B −1 E)−1 E T B −1 .
Suppose we consider a contingency in which line l con- Now consider the real power flow on branch l connecting
necting buses m and n is removed. Then the modified B buses i and j. If we let X̃ = X − E T B −1 E then
matrix is given by
1 T −1
1 pl = e B p
B=B− emn eTmn xl ij
xl
1 T −1
where emn is a vector containing a 1 in location m and = e [B + B −1 E X̃ −1 E T B −1 ]p
xl ij
−1 in location n. Using the Sherman-Morrison-Woodbury 1
−1
matrix inversion theorem, B is given by = pl + eTij B −1 E X̃ −1 E T B −1 p
xl
−1 B −1 emn eTmn B −1 Here pl is the base case or pre-contingency real power flow.
B = B −1 +
xl − eTmn B −1 emn Letting B −1 E = U −1 D−1 E where D−1 comes from the
= B + γl B −1 emn eTmn B −1
−1 factorization of the base case network B matrix as defined
in (22) and E = (em1 n1 em2 n2 · · · emk nk ) where emk nk
where is found by forward substitution in (23) then we have
1
γl = .
xl − eTmn B −1 emn 1 T −1 T
pl = p l + e D E X̃ −1 E p = pl + glT dl (24)
In this case, for the real power flow on branch l connecting xl ij
buses i and j we have
where
1 T −1
pl = e B p 1 T −1
xl ij pl =e B p
xl ij
1 T −1
= e [B + γl B −1 emn eTmn B −1 ]p 1 T
xl ij gl = E D−1 eij
xl
1 T γl
= eij p + eTij D−1 emn eTmn p dl = X̃ −1 E p
T
xl xl
Equation (24) expresses the post-contingency real power
flow pl in terms of the pre-contingency flow pl and a k × k
update matrix that can be computed efficiently. This
approach permits computing the post-contingency power
flows without recalculating the entire post-contingency dis-
tribution matrix B. Hence, the post-contingency transmis-
sion constraints in (5) can be efficiently tested for viola-
tions. Those that are violated are added to the constraint
set for future iterations; i.e., the ∆πk block of (21) is en-
larged accordingly.

VII. References
[1] N. Alguacil, A. J. Conejo, “Multiperiod Optimal power Flow
Using Benders Decomposition”, IEEE Transactions on Power
Systems, Vol. 15, No. 1, February 2000.
[2] K. Clements, P. Davis, “Contingency Constrained Optimal
Power Flow for Deregulated Electricity Markets”, NSF proposal,
March 24, 1999. Grant ECS-9810288
[3] G. Infanger,Planning under uncertainty: Solving large-scale
stochastic linear programs, Boyd and Fraser, 1994.
[4] M. Madrigal, V. H. Quintana, “An Interior-Point/Cutting-Plane
Method to Solve Unit Commitment Problems”, IEEE Transac-
tions on Power Systems, Vol. 15, No. 3, August 2000.
[5] A. Monticelli, M. V. F. Pereira, S. Granville, “Security-
Constrained Optimal Power Flow With Post-Contingency Cor-
rective Rescheduling”, IEEE Transactions on Power Systems,
Vol. 2, No. 1, Feb. 1987.
[6] I. M. Nejdawi, K. A. Clements, P. W. Davis, “An Efficient Inte-
rior Point Method for Sequential Quadratic Programming Based
Optimal Power Flow”, IEEE Transactions on Power Systems,
Vol. 15, No. 4, Nov. 2000.
[7] B. Stott, O. Alsac, A. Monticelli, “Security Analysis and Opti-
mization”, Proceedings of the IEEE, Vol. 75, No. 12, pp. 1623–
1644, Dec. 1987.
[8] G. Irisarri, L. Kimball, K. Clements, A. Bagchi, P. Davis, ”Eco-
nomic Dispatch with Network and Ramping Constraints via In-
terior Point Methods,” IEEE Trans. Power Systems, Vol.13, No.
3, pp. 236-242, Feb, 1998.

VIII. Biographies
Lucia M. Kimball (Member, IEEE) received the B.S. in math-
ematics from the University of Massachusetts, Lowell in 1987 and
the M.S. and Ph.D. degrees in applied mathematics from Worcester
Polytechnic Institute in 1989 and 1997. She is currently an Associate
Professor in the Department of Mathematical Sciences at Bentley
College, Waltham, Massachusetts USA.
Kevin A. Clements (Fellow, IEEE) received the B.S. degree
in Electrical engineering from Manhattan College, Bronx, NY, in
1963, and M.S. and Ph.D. degrees from the Polytechnic Institute
of Brooklyn, Brooklyn, NY, in 1967 and 1970, respectively. He is
currently a Professor at Worcester polytechnic Institute, Worcester,
Massachusetts USA.
Slobodan Pajic (Student Member, IEEE) was born in Kikinda,
Serbia. He received the B.S. degree in electrical engineering from
the School of Electrical Engineering, University of Belgrade, Serbia
in 2001. He is currently pursuing his M.S. degree in electrical engi-
neering at Worcester Polytechnic Institute, Worcester, Massachusetts
USA. His research interests are in the areas of state estimation in
power systems, optimal power flow, power system computation, and
methods for the solution of large-scale systems
Paul W. Davis (Member, IEEE) received the B.S., M.S., and
Ph.D. degrees in applied mathematics from Rensselaer Polytechnic
Institute, Troy, New York. In 1970 he joined the Mathematical
Sciences Department at Worcester Polytechnic Institute, Worcester,
Massachusetts USA, where he is currently a Professor and Dean of
Interdisciplinary and Global Studies.