40 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS VOL. PAS-87, NO.

1 JANUARY 1968

The current on the cable then is An alternative form of (2), which relates leakage current
density to a specific cable, is the following:
Iop d(x2 + y2) -1/2
I(z)= -2R dx Iop
2ir2dRy3 (4)
Iopx
I'J 2?rR(x2 + y2)312 (3) where i is the maximum leakage current density in A/M2, and
d is the cable diameter in meters.
where I(Z) equals the current on cable at distance x.
Also, the unit leakage current (I') is ACKNOWLEDGMENT
F The author gratefully acknowledges the assistance and co-
= _ dI Iopd2(x2 + y2) -1/2 operation of the members of the Pacific Northwest Bell Tele-
dx 27rRdx2 phone Company and the BPA in making the earth potential
measurements.
IoP [ 2X -y2] REFERENCES
27R L(X2 + y2)512J
[l] F. M. Stumpf, "Induction effects of hvdc transmission on
The maximum leakage current when x = 0 is telephone communications," 1965 IEEE Internat'l Conv. Rec., pt. 9,
vol. 13, pp. 77-88.
[2] R. S. Gens and R. F. Stevens, "The high voltage dc test program
I (max) = Ry3. of the Bonneville Power Administration," IEEE Trans. Power
Apparatus and Systems, vol. PAS-82, pp. 1054-1060, December
1963.
The previous equation is identical to (2); the latter was derived [3] E. Sunde, Earth Conduction Effects in Transmission Systems.
New York: Van Nostrand, 1949, p. 44.
from basic transmission equations and makes no assumption [4] E. Sunde,[31 p. 161.
about the equivalence of the cable and earth potentials. [5] E. Sunde, [3] p. 200.

Optimum Control of Reactive Power Flow

JOHN PESCHON, MEMBER, IEEE, D. S. PIERCY, WILLIAM F. TINNEY, SENIOR MEMBER, IEEE,
ODD J. TVEIT, MEMBER, IEEE, AND MICHAEL CUPNOD

Abstract-The general problem of minimizing the operating cost INTRODUCTION

of a power system by proper selection of the active and reactive THE GENERAL problem of minimizing the instantaneous
productions is formulated as a nonlinear programming problem in operating cost of a power system subject to the equality
accordance with previous work by Carpentier of Electricite de constraints imposed by the physical laws governing the trans-
France. This general problem is particularized to the minimization mission system as well as the inequality constraints imposed by
of transmission line losses by suitable selection of the reactive pro- the equipment ratings was formulated in 1961 by Carpentier[l'] of
ductions and transformer tap settings. An efficient computational Electricite de France. The resulting optimization problem was
procedure based on the Newton-Raphson method for solving the shown to be one of nonlinear programming. The sufficiency con-
power-flow equations and on the dual (Lagrangian) variables of the ditions of the theorem of Kuhn and Tucker[21 were applied to
Kuhn and Tucker theorem is discussed. This minimization procedure derive the optimization equations which must be satisfied at the
has been applied successfully to a 500-node system studied by the optimum.
Bonneville Power Administration for which an effective power-flow In the general case, the control variables are:
program had been developed previously. The dual variables asso- 1) power settings of the generation units that control the active
ciated with the primary (electrical) variables are obtained in the injections and determine the phase angles in the transmission
course of the computation, and their engineering significance for system
power system design and tariffication is emphasized. The procedure 2) generator excitations and other means for reactive injec-
has been extended to the general case of combined active and tions, such as capacitors, which determine the voltage levels in
reactive optimization. the transmission system.
Carpentier treated as a particular case the optimization of real
power by assuming a constant and known voltage throughout
Paper 31 TP 67-47, recommended and approved by the Power the transmission system. A successful computer program capable
System Engineering Committee of the IEEE Power Group for of solving the Kuhn and Tucker optimization equations one by
presentation at the IEEE Winter Power Meeting, New York, N. Y., one (as in the Gauss-Seidel method) was apparently established
January 29-February 3, 1967. Manuscript submitted October 27,
1966; made available for printing June 14, 1967. at Electricite de France for a 50-node system in this special case.
J. Peschon, D. S. Piercy, and 0. J. Tveit are with the Stanford Independently of this work, Siroux,[3] also of Electricite de
Research Institute, Menlo Park, Calif. France, formulated the optimization problem in an almost
W. F. Tinney is with the Bonneville Power Administration, identical manner and arrived at the Kuhn and Tucker conditions
Portland, Ore.
M. Cuenod is with the Soci&t6 G6nerale pour l'Industrie, Geneva, of optimality by a remarkable process of engineering intuition
Switzerland. based on the marginal costs of Kirchmayer. [4]

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:24 from IEEE Xplore. Restrictions apply.
 PESCHON et at.: OPTIMUM CONTROL OF REACTIVE POWER FLOW 41

In 1965 the special case of optimizing a power system with SUMMARY OF CARPENTIER-SIROUX FORMULATION
respect to the reactive injections and transformer tap settings An electric power system of N nodes is considered. There may
was undertaken by the authors for the Bonneville Power Ad- be a production Pi, Qi and a consumption Ci, Di of active and
ministration (BPA).[51 Since the BPA system is supplied ex- reactive power, respectively, at the ith node, which is further
clusively by hydraulic generation, the production of which is characterized by its voltage V; and phase angle Ci. For pure
determined largely by water availability considerations, the only production nodes, Ci = Di = 0, and for pure consumption
remaining control variables are the reactive injections, unless one nodes, Pi = Qi = 0. A frequent intermediate case is Pi = 0 and
wishes to attack the much more ambitious variational problem of Qs $ 0.
optimally utilizing the hydraulic production units interconnected
by means of a lossy transmission system. Carpentier's mathe- The total productions, EPi and EQi, must equal the total
matical model was adapted to this special case, and programs i i
were written to implement the Gauss-Seidel procedure he had consumptions ECi and ZED plus the respective transmission
suggested in his paper. [1] For several reasons-notably, the i i
erratic convergence properties, the difficulty of computing per- losses. The cost J (in dollars per unit time) of operating the
missible corrections in the presence of constraints, and the in- system depends only on the active powers Pi, the reactive powers
compatibility with the existing BPA power-flow program,[6] 71 Qi being free of cost once the equipment required for their
which employs Newton's method-a different computational production (e.g., capacitors) has been installed. Thus,
approach was developed. This approach, which appears to be of J = F(P1,... ,PN)- (1)
considerable generality, consists of making vector adjustments
of the primary control (electrical) variables in accordance with If, in a given power system, the Pi, Ci, Qi, and Di are specified,
the information derived from the dual variables associated with then the voltages Vi and the phase angles Ci between these volt-
a feasible but not necessarily optimum solution. Since all of the ages and V1 (whose phase angle is arbitrarily set equal to zero)
control variables (reactive generations) are adjusted simul- are fixed. The related variables Si, Pi, Qi, Vi, and Ci are subject
taneously, the program converges to an optimum solution in to upper (M) and lower (m) constraints of the form
much the same manner as it does to a feasible solution. With
present program implementation a typical 500-node system Pi2 + Q2 < Si2,M (2a)
takes about ten minutes to reach an optimum solution on an Pi > Ptn (2b)
IBM 7040 computer. Contemplated improvements are expected
to cut this time by more than 50 percent. Qi < QiM (2c)
Since the Carpentier-Siroux papers were never published in the Qi Q<i
> (2d)
English-speaking technical literature, their formulation of the
general economic optimization problem will be summarized. Vi < vim (2e)
Thereafter, this general formulation will be particularized to
reactive optimization. The computational procedure developed Vi > Vim (2f)
in the course of the study will be discussed and justified, and 0,,, I<Tia,.
|0i - (2g)
sample results will be given. The present limitations of the pro-
cedure will be pointed out and fairly straightforward remedies The engineering justification for these constraints is given by
will be given. The extension of the procedure to the general case Carpentier. [1]
of simultaneous active and reactive optimization and an applica- The problem of power system optimization then consists of
tion to power system operation under emergency conditions will minimizing J, subject to two sets of constraints:
be discussed in a forthcoming paper. 1) satisfaction of the inequality constraints imposed upon Pi,
Qi, Vi, and Ci - 0a in accordance with (2)
NOMENCLATURE 2) satisfaction of the network relations that must exist between
the variables Pi, Qi and Vi, Oi.
N number of nodes For convenience of analysis (but not of programming) the
Pi active power generated at node i network relations (power-flow equations) may be written as
Qi reactive power generated at node i
Ci active power consumed at node i
Di reactive power consumed at node i Pi -Ci = It= E V sin ( - C -bi)
a Zia
Ii active injection at node i, = Pi-Ci
Ki
V*
reactive injection at node i, = Qi- Di
voltage at node i
+ E Vi2 sin 3ia
Oi phase angle of voltage at node i measured with
reference to node 1 (01 = 0)
a
Zia (
Zia, impedance magnitude of tie line ia Qi-Di= -E cos (ei-Xa-8a)
ctZicr
siat loss angle of tie line ia;38ia = sin-' Ria,/Zia
Yii admittance magnitude with respect to ground at Vi.2
Tia
node i
maximum admissible phase angle difference between
+ E
a
-Z7
Zia
2
CSa"a - Y,,Vi2. (4)

connected nodes i and a The stated problem is one of static optimization in the presence
Si apparent power production at node i, = V\/P2 + Qi2 of constraints. The optimization conditions provided by the
P.M minimum and maximum permissible real production
pim,$,t theorem of Kuhn and Tucker[21 constitute a practical solution,
at node i which can be implemented for real-time control.
Q mM QM minimum and maximum permissible reactive pro-
duction at node i THEOREM OF KUHN AND TUCKER
SiM maximum permissible apparent production at node i Given
Vim, Vim minimum and maximum permissible voltage at node i 1) cost F(x) of the vector variable x
A Jacobian matrix of the power-flow equations 2) inequality constraints of the form
sensitivity matrix
pig autotransformer ratio, = V,/FVi. G(x) < 0 (5)

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:24 from IEEE Xplore. Restrictions apply.
 42 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS JANUARY 1968

3) equality constraints of the form INTERPRETATION OF DUAL VARIABLES

H(x) = 0 (6) It is proved by CarpentierWl] that at the optimum the dual
variables Xi and u, measure the marginal or incremental cost of
and assuming convexity for F, G, and H, the conditions under active and reactive consumption at node i, i.e.,
which F(x) is minimum is dZ = 0, with
OF
= _. -
AF
S = F(x) + aTG(x) + f3TH(x) (7) (16)
ac 2 0, p arbitrary (8)
where aT is the transpose of the vector a, and where r =D- ADi (17)
a1Gs(x) = 0. (9) Similarly, the dual variables Mi, mi, ei, ei', ui, ui', and ti,a measure
The a; and *3i are the dual variables associated with the in- the marginal cost associated with the constraints (2a)-(2g),
equality constraints of (5) and the equality constraints of (6), respectively. For example, the dual variable
respectively. Condition (9) implies that ai = 0 if constraint -
laF AF
G1(x) is not reached, and ac > 0 in the alternative.' Ui =
av*M AV
(18)

KUHN AND TUCKER CONDITIONS OF OPTIMALITY APPLIED relates the saving AF to a relaxation AViM in the upper voltage
TO A POWER SYSTEM bound ViM.
The marginal cost, (16) and (17), thus provides a logical basis
Let Xi and jlj denote the dual variables associated with the for tariffication, whereas the marginal costs of the type repre-
equality constraints (power-flow equations) (3) and (4) and let sented by (18) point out the equipment constraints and thus aid
the dual variables Mi, mi, ei, ei', ui, ui', and ti, be associated in system expansion studies.
with the inequality constraints (2a) through (2g), respectively.
The function 2 [see (7)] is then written as NUMERICAL SOLUTION OF OPTIMIZATION EQUATIONS
£ = F(P1, ... ,PN) + Xi(Ii - Pi + C) + Z/i(Ki -Qi It is instructive to verify that the number of optimization
i i equations (11)-(15) together with the power-flow equations (3)
and (4) indeed equals the number of unknowns. At the same
+ DO) + EMi(Pi2 + Q12 - S,2,M) + Emi(Pin
i
- P) time, the programming steps required in a numerical solution
i
become apparent. By way of example, the 3-node system shown
+ Eei(Qi - QiM) + Eeis(Qim - Q1) + Eui(Vi - ViM) in Fig. 1 will be discussed.
Assuming first that the reactive generations Q, and Q2 are
+ Euit(Vim - Vi) +
Etia(1i - Oa - TiaM). (10) sufficient (Al = U2 = 0), one expects intuitively that at least one
i ia node, say number 1, operates at maximum voltage ViM, in which
case ul X 0. Under those circumstances, the unknown primary
Setting d£ = 0, the following group of optimization equation variables are
results for a node with both production and consumption. For a
pure consumption node, (11) and (12) do not exist. P1, P2, Q1, Q2, V2, V3, 02, 03
Differentiation with Respect to Pi and Qi since 0, is arbitrarily set equal to zero.
The unknown dual variables are
OF
Xi
bpi
+ 2MiPi - mi (11) X1, X2, X3, M3, Ul.
The total number of unknowns is thus 13; the equationsavailable
lA= 2MiQi + ei - ei'. (12) for their resolution are
Differentiation with Respect to Oi and Vi (3,), (32), (33), (41), (42), (43), (11,), (112), (13,), (132),
Xia +Z E Xa + Ai +E yAa +E (tta -
tai ) =0 (141), (142), and (143)
(O0i a b0 o a )0 at
that is, 13 nonredundant equations. Equation (133), which is the
(13) sum of (131) and (132), is omitted in the count.
If the guess V, = V1M, V2- < V2 < V2M, V3i < V3 < V3m iS
Xi - + E Xa + i 1 + E
ZMa + u - ui1 0. = correct, then the optimization equations can be resolved with
6 a ov ov a loiV reasonable ease, for example, by means of the Newton-Raphson
approach.'7] If, on the other hand, this guess was incorrect, a
(14)
In addition, the power-flow equations (3) and (4) and the f
exclusion equations
M1(Pi2 + Q12 + S2,M) = 0 mi(P1i - Pi) = 0
ei(Qi - Q1M) = 0 ei (Qim - Qi) = 0 (15)
Ui(Vi -
ViM) = 0 Ui (Vim -
Vi) = 0
(0i - Oa - Tia)
ti = 0
C3 D3 V3 83
must be satisfied at the optimum.
Fig 1. Three-node example system. Nodes 1
1 A simple geometric interpretation suggested by Luenberger can and 2 are production nodes (0), node 3 is
b e found in Peschon et al.[r] a pure consumption node (0).

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:24 from IEEE Xplore. Restrictions apply.
 PESCHON et al.: OPTIMUM CONTROL OF REACTIVE POWER FLOW 43

different set of optimization equations would result. For example, OPTIMIZATION EQUATIONS FOR THE REACTIVE CASE
if the optimizing voltages were Vim < V1 < VlM; V2 = V2M; Since now the active injections are fixed except at the swing
V3m < V3 < Vsm, then the unknown primary and dual variables bus (node 1), the real power constraint in (15) is omitted; also,
would be in view of the dominant effect of P upon 0, the angle constraint in
(15) is omitted. The apparent power constraint, which involves
Pl; P2, Ql) Q2; Vl, V32 021083 quadratic terms, is likewise dropped for reasons of convenience.
The resulting optimization equations are then written as
Xl, 2,X3, /.L3) U2 follows:
respectively. Equation (141) would no longer contain the dual OF
variable u,, but (142) would contain U2, which was zero in the = P = p (19)
previous example. The power-flow equations (3) and (4) would
also be written differently, since V2 is now imposed instead of Vi. where ,B is the power cost in $/MWh at node 1,
For a third example, one might assume that V2 should be at (20)
V2m, but that the required reactive generation Q2 would exceed yi = ei - ei
Q2m. Under these circumstances, one would expect the optimum
to be characterized by (21)
Xi a +E - + mi a +
?j0 E HaJ0i =-o
Vlm < V1 < Vl-; V2' < V2 < V2M; VP, < V3 < V3M aKa
+EXa, bIcr + ;ia +ZE
ais aKi-
Xi a
Qlm < Ql < QlM? Q2 = Q2M.
+ ui - uiI = 0 (22)
The unknown primary and dual variables are now
-
e (Q-QiM) = 0 (23)
P,, P2, Ql, V1, V2, V3, 02, 03
e I(Qfi - Q.) = 0 (24)
X1, X2, X3, el = Al, 13 as (V; V.m) 0
- = (25)
respectively. The optimization equations again contain a different U; (Vim - Vi) = 0. (26)
set of unknowns.
These few examples point to a major difficulty in implementing The power-flow equations (3) and (4) of course remain unchanged.
a computer program for resolving the optimization equations: To verify that (14) through (26) together with (3) and (4)
one does not know at the outset which of the variables should be suffice to determine an optimum solution, the example system of
treated as knowns and which as unknowns. Fig. 1 is used again; if, at the optimum,
To overcome this difficulty, Carpentier proposed the following
computational scheme: V1 = VAM, Vlm < V2 < V2M; V3m < V3 < V3M; Qlm < Q < Q1M;
1) A set of nonconstrained primary variables P, Q, and 0 are Q2m < Q2 < Q2M
assumed which need not satisfy the power flow equations (3) and
(4) at the outset. A set of dual variables X and A. is also assumed. the primary unknowns are
2) This set is substituted into one optimization equation and P1, Qly Q21 V2, V3, 02, 03
the change required in one of the variables to satisfy this equa-
tion is computed. and the unknown dual variables are
3) If the required change is allowed by the constraint equa-
tions, it is carried out; if not, the required change is forced into 1l X2, X3,132 Ul-
another variable and the dual variable associated with the con- The equations available to determine these twelve unknowns
strained primary variable is introduced into the appropriate are the twelve equations:
optimization equations.
4) Proceeding to the next equation the corrected value ob- (19), (211), (212), (221), (222), (223), (3,), (32), (33),
tained in step 3 is used. A permissible change for another variable (41), (42) and (43).
is computed.
Obviously, each equation has a natural variable associated Since at the optimum V, = Vim, (221) must contain the dual
with it and the required correction should be performed on that variable ui.
variable, if at all possible; for example, the natural variable for
(3;) would be Pi if i is a production node, and at if i is a con- COMPUTATIONAL APPROACH TO FIND THE OPTIMUM
sumption node. The implementation of a computer program
capable of carrying out these steps for a large system is quite In view of the obvious fact that efficient transmission of power
involved as a result of all the logic required; a program capable of implies high-voltage levels throughout the grid, it was guessed
handling the 3-node system of Fig. 1 was written and extensively that at the optimum all nodes capable of reactive production
tested. This program performed satisfactorily for a certain range would be at maximum voltage V = VM except for those nodes
of consumptions C and D, but did not converge for relatively where the reactive production constraint Q < QM would be
high and relatively low values of D. It is believed that these violated. For these nodes, the optimum would be characterized
convergence difficulties are characteristic of the step-by-step by Q = QM. To check if indeed this strategy satisfied the opti-
procedure. mization equations, the following procedure was carried out:
In view of the very considerable programming effort required 1) With V = VM at all reactive production nodes, all the
and the convergence problems encountered, it was decided to remaining primary unknowns (Pi, Q, 0; V for reactive consump-
develop an entirely different computational approach for the tion nodes) should be obtained from the power-flow equations
special case of reactive optimization. In the remainder of this (3) and (4).
paper, the discussion will be limited to the program developed for 2) For those nodes where Q > QM, Q = QM should be imposed
this special case; in a later paper, the modifications required for and all the primary unknowns from the power-flow equations
active optimization will be described. recomputed.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:24 from IEEE Xplore. Restrictions apply.
 44 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS JANUARY 1968

3) The primary variables thus found should be substituted into spectively. Following the previously quoted theorem, those com-
the set of equations (21), (22), with A1 = f; a dual variable u is in- ponents of V' and Q' having negative dual variables must be
cluded for all nodes where V = VM and a dual variable /u for all decreased during the next iterations until they become zero.
nodes where Q = QM In order to obtain fast convergence, it is convenient to assume a
4) The set (21), (22) is resolved for all the unknown dual linear relation between the primary variable to be decreased and
variables X (except A1), A, and u. It should be noted that the its dual variable, as shown in Fig. 2 for voltage Vi'.
system of simultaneous linear equations (21), (22) is conve- If the slope mi A AVV'/Aui were known, the optimizing value
niently resolved in two stages: first A2,. . .,,, and all nonzero Ai are Vi'* could be computed directly as
obtained from (21) and that part of (22) which does not contain
u in terms of the unknown A1. Thereafter, X and M are substituted Vi'* = Vi'(O)
miui(O). - (27)
into the part of (22) that contains u, and u is calculated; ul is This slope mi is easily obtained from two consecutive computa-
the only remaining unknown in (22X). tions, where the prescribed voltages Vi'(0), Vj'(1) lead to the
If this strategy is to be optimum, then all the dual variables u
and those of the dual variables ,u associated with reactive pro- dual variables ui(O) and ui(1), respectively. The voltage Vj'(2)
duction nodes must be positive, as the Kuhn and Tucker condi- to be imposed in the third iteration is then related to Vj'(0) and
tions would be violated otherwise. It was found that certain of Vj'(1) by
these dual variables were negative, which suggested that the
corresponding voltages or reactive injections should be decreased. Vi'(2) = Vi'(1) - V7'(1) -
Vi'(O) ui() (28)
In order to derive from this observation an iterative computa-
tional scheme for reaching the optimum, the following theorem In actuality, the relation between Vi and ui (or Qj and ,u) is
was established: not linear, and the optimizing values Vi'* and Q/* will not be
Theorem: Let the vector of voltages and reactive generations obtained during the third iteration. Therefore, further iterative
prescribed in the power-flow equations be denoted by V' and Q', adjustments in accordance with the following recurrence relation
respectively. Then, if as a result of imposing V' and Q'. . ., the are required:
set P, Q, 0, V satisfies the power-flow equations (3) and (4) and
the constraints (23) through (26) but not the optimization equa- Vi'(k) = Vi'(k - 1) - Vi'(k - 1) -
Vi'(k - 2) uit(k - 1).
tions, the imposed pair V', Q' optimizes a fictitious power system
S', whose electrical properties are identical with those of the (29)
actual system S except for the constraints QM, Qm, VM, Vy.
The proof of the theorem is straightforward if one associates a For the reactive generation Qj' to be decreased, the adjustment
dual variable Ai = ei and ui with each of the variables Qi( and equation is similarly
Vi', respectively, which have been prescribed in the power-flow
equations. Substitution of the set P, Q, 0, V satisfying (3) and (4)
into the singular linear system of (21) and (22), together with the Qj'(k) = Qj'(k- 1) - (k -1) -Q(k -2) j(k - 1).
result of (19), yields the dual variables
(30)
Xl 0 =
xi -(i = 2, . . . ,N) Since a dual variable is computed for every imposed primary
-i for the nodes where Qi is prescribed (this includes the variable, the adjustment (29) and (30) can be performed simul-
consumption nodes where Qi = 0) taneously for all those voltages and reactive generations that
ui - for the nodes where Vi is prescribed. need to be decreased. This very desirable feature is referred to as a
vector adjustment operated upon the correction vector of com-
The variables MAi and ui may be positive or negative. If positive, ponents
the fictitious optimum system S' is restricted by constraints of
the type QiM, ViM; if negative, it is restricted by constraints of Vi' (k) -Vi' (k -1)
the type Qim, Vim.
The cost of operating the actual system S can hence be reduced Q/(k) -
Qj'(k 1).-

by increasing the prescribed Qi', Vi' when the corresponding Ai

and ui are positive to the extent allowed by the actual con- Computation of Duat Variables
straints, and vice versa.
This theorem constitutes the basis for the optimization pro- The dual variables u and M required to implement the adjust-
cedure developed in the course of the study. ment algorithm of the previous section are obtained from the set
of equations (19), (21), and (22), which also contain the dual
Determination of the Correction Vector variables X and those of the , associated with reactive consump-
The result of the first iteration is a set of dual variables u and M tion nodes. Since the feasible solution fixes all the electrical
associated with the imposed primary variables V' and Q', re- variables, the partials
bIi 6Ia bKi 6Ka bIL bla 6Kz ?Ka
and
o' boi) Iboi bOi bVI
I
bVi' avi'V '
bV
Vj' (O)
are all known; consequently, the set (19), (21), and (22) con-
stitutes a system of simultaneous linear equations the unknowns
of which are the dual variables.
It was noted in the course of the study that the matrix corre-
sponding to the set (19), (21), and (22) with all the equations
containing u removed is the transpose of the Jacobian matrix A
corresponding to the power-flow equations (3) and (4). This is
very important from the standpoint of efficient implementation
Fig. 2. Assumed linear relation of Vi with ui' at the optimum, on the computer. In the power-flow program the linear system
Vi'= Vi'* and ui = 0. Vl'(0) is the voltage imposed in the based on A is solved by a very efficient scheme for sparse ma-
first iteration and ui(0) is its corresponding dual variable.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:24 from IEEE Xplore. Restrictions apply.
 PESCHON et al.: OPTIMUM CONTROL OF REACTIVE POWER FLOW 45

trices. This scheme, explained in several other sources,-[8]ill It is shown in Peschon et al., [5] Section III-B, that the matrix S
factors A into upper and lower triangular matrices. These can be calculated easily once the Jacobian matrix A of the power-
triangular matrices may be used to obtain repeat solutions of flow equations has been inverted. From the general relation (31),
linear systems based on A or AT in the same manner as if A-' the sensitivity of Vi with respect to Va', VB', etc., is readily
and (AF)-1 were available, but with much greater efficiency. obtained as
Therefore, the direct solution of (19), (21), and (22) is available (32)
as a byproduct of every iteration of the Newton method of spiPAV#,
AVi =
pia. Va + '. ..

power-flow solution and may be obtained with only a small On the other hand, it is known from (18) that the sensitivity of
modification of the program. the saving AF is related to the adjustable voltages VI,
-

In Peschon et al., 51 Section V-B-1, a 5-node example is treated VP',... by

in detail to show that indeed the matrix corresponding to the set
(19), (21), (22) is the transpose of the Jacobian matrix associated -AF = uitAVt' + u#AV#'.... (33)
with the power-flow equations (3) and (4). A rigorous proof of To obtain the optimum solution in the presence of indirect
this relation has been given by Dauphin et al. of Electricite de constraints, it is therefore desired to maximize the incremental
France. [12] saving -AF with respect to AVa', AV#',..., subject to the
equality constraint (32) in which AVi is chosen in a manner such
Special Cases that Vi = Vim. The resulting suboptimization constitutes a
As a result of the previously described optimization procedure, straightforward linear program.
it is usually found that: In this paper, only a brief summary of the approaches available
1) Many reactive generation nodes will operate at VM if there in the case of indirect constraints is given for the following reason:
is sufficient reactive generation, at QM if there is not. the condition Vi > ViM arises usually when the system is lightly
2) Some of the reactive generation nodes will operate below VM loaded-that is, when the line losses are comparatively small and
or QM, respectively. The optimizing values of V and Q are easily when consequently the gain of an exactly optimum solution over
obtained from the above described adjustment algorithm [see a nearby optimum solution is negligible. In a forthcoming paper,
a detailed treatment of indirect constraints will be given in con-
(29) and (30)].
3) The voltages at the reactive consumption nodes naturally nection with combined real and reactive power optimization for
fall within the permissible region V", VMI . which the linear program defined by (32) and (33) has been
Sometimes, especially when the system is lightly loaded, the actually used.
reactive power produced by the line itself causes the voltage to EXAMPLE OF REACTIVE OPTIMIZATION
rise above VM at certain of the reactive consumption nodes.
One possible modification of the optimization procedure in the To illustrate these procedures, the computer results of the 9-
presence of such indirect constraints is as follows: node system shown in Fig. 3 are given. It is reemphasized at this
Let i denote the consumption node at which the voltage Vi was point that the method is not limited to systems of that size, but
found to exceed ViM. Then, impose Vi = ViM in the power-flow has been successfully applied to a 500-node system. The results
equations and free a nearby reactive production node a. The re- of the 9-node example are given in Table I. The following should
quired voltage Va is obtained from the power-flow program. be noted:
Associate a dual variable ui with the imposed voltage Vs. As
before, compute all dual variables and check that they have the
0-3
right sign. If so, the optimal solution has been obtained. If not, Y33- 0.925* 10 Y, = 1.05 -10 3
z=2956
Yl=11
ll
4_s.--I
the voltages at other adjacent production nodes -y, etc., must
~~o
4( An
TD=:-C=94
g, D=80 \ a 8=0.1493 94 D- 140 I

P =O -20 P=8605
be lowered as determined by the sign of us, ui, etc.
The programming logic required to handle this special case is zz 121 08 z= 12.65 z =28.28
somewhat more involved because the reactive injections at 8=0.1325 8=0.3217 8 =0.1418

several production nodes may have to be adjusted subject to the P :900

C =200
constraint Vi = Vim, and it is not always obvious beforehand C =3206
D= 105
4\66 = 0.55 * 10-3 1fi =150 z.17.2
8=0.1882
Y44 1.775I-I.- 9= 20
D= -!00
which production nodes should be adjusted. P=O Y55=0.2.10 3

To determine the effect upon V1 of altering the imposed volt-

ages Va', Va', etc., at nearby production nodes, it is convenient
z= 12.16 z=80.4
8=0.1652 j 8=0.0992
to compute the sensitivity matrix relating the unprimed to the
primed variables, i.e., P=448 CD:-175
D

17324
P-0
-

10-3D===XjC
C - 200 z 60.49 20
=20=
Y99 09 - - Y =0.5063 04 10
APi-
=

0.925l= YN88 = 0.25 10-3 -I0`3 Y977 0.475

AV = [J QV'] (31)
Fig. 3. Nine-node example problem. Generation nodes are marked by
AQ i

a double circle. The constraints are V5M 240; VgM = 236 kV; all
-AO other voltages are constrained at 242 kY.

TABLE I
NUMERICAL RESULTS FOR EXAMPLE

Node 1 2 3 4 5 6 7 8 9
V 242.00 241.36 242.00 238.89 240.00 235.75 242.00 236.76 236.00
Q 68.5 0 143.5 0 -1.3 139.0 103.0 0 76.2
Ui 0.364 0 0.042 0 1.144 0 0.3307 0 0.69
, 0 -0.009 0 0.0248 0 0 0 -0.0071 0
X 2.000 2.063 2.240 2.108 1.990 2.4892 2.387 2.587 2.462
F = 1721 $/h = gEPl.
P1 = 860.5 MW.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:24 from IEEE Xplore. Restrictions apply.
 46 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS JANUARY 1968

1) A nonzero dual variable u is associated with production In the presence of these autotransformers, the power-flow
nodes 1, 3, 5, 7, and 9 where the upper voltage constraint VM is equations are rewritten as
reached; u5 has largest magnitude, which signifies that V5M
should be increased in preference to the other voltage constraints. lh = Z asin (Oi Oa -ia) + i6ia
2) A nonzero dual variable , is associated with consumption a Zia a Ziia
nodes 2, 4, and 8. At nodes 2 and 8, the system operating cost
would be reduced by increased reactive consumption. + E pa d sin(O - ,- B) + E p2i. --sinOi
3) The dual variables X measure the incremental cost of con- j3 Zip Zip
sumption at each of the nine nodes. The consumption cost
differs from the production cost (except at the swing bus) because
at least part of the demand increment is not generated locally, + E pZ 8 sin (7i- -bi) +
ly Ziy ly zi
b7 sin y (34) EV2
but needs to be imported. This is not so for node 5, where addi-
tional consumption would actually reduce transmission line
losses by a small amount since X5 < 2.00.
4) The total production MP: is 2208.5 MW, whereas the total Ki=- E V-V, cos(Oi-oa-bia) + EZ-Cos V.2
ia
consumption is 2C: = 2084 MW. The difference of 124.5 MW a Zia a Zia
equals the total transmission line losses.
Other simple numerical examples can be found in Peschon - V2 ZYia - cos (Oi -o bi- - )
et al.,[5] Section V-F. a 2 Zip
AUTOTRANSFORMERS
V.2
V2 E
~~~~~~1
+ E p2t- Cos e5, - Pp - Yi
The model of the system used up to this point for reactive d Zip p 2
power optimization does not include the presence of autotrans- V-V7 V.2
formers. pi7 Zi7 COs (7i- e- ) + E z COS 7iT
In what follows, the modifications required by the optimiza-
tion procedure to account for autotransformers are briefly dis-
cussed. These transformers introduce additional control vari- -V,2 2 Yi. (35)
ables-namely, the transformer ratios, assumed to be continu- 7y2
ously variable. The optimization therefore not only involves the When the system is optimized with respect to reactive power
reactive generations, but also the transformer ratios. flow, the autotransformer ratios pip, pi, constitute independent
The discussion is concerned only with voltage transformers; optimization variables. These ratios are assumed to be con-
the mathematical approaches required for phase transformers tinuously variable between the bounds pm and pM, i.e., the
are very similar and will therefore not be given explicitly. presence of the steps is disregarded.
The autotransformers of ratio pig in the line ig and located at The function £ is written as before [see (10)1, except that the
node i are represented by the equivalent circuit shown in Fig. 4 dual variables (r and r') associated with the constraints on p
where all variables and parameters are referred to the left side must be considered. The resulting optimization equations can be
of the transformer. found in Section V of Peschon et al. [5] Optimization with respect
It is convenient to distinguish between three different types of to autotransformer tap settings was successfully performed for a
lines from node i to the adjacent nodes a, (, and y, shown in 500-node system.
Fig. 5. In line ia there is no transformer; in line i3 there is a
transformer at i; in line iy there is a transformer at 'y.
SUMMARY AND CONCLUSION
In this paper, the Carpentier problem formulation for mini-
i Pip mizing the instantaneous operating cost of an electrical power
system by judicious choice of the active and reactive injections is
summarized and the necessary conditions of optimality are
stated. This general problem formulation is thereafter partic-
ularized to the special case of minimizing the transmission line
0-~~~~~~~_
2 losses by judicious choice of the reactive injections and trans-
former tap settings, the active injections being fixed.
t Pip~ T ~~P
To obtain a practically computable solution of the optimiza-
tion equations, an apparently new adjustment procedure of the
electrical control variables based on the sign of the associated
dual variables corresponding to a feasible solution is developed
Fig. 4. Transformation of a circuit with autotransformer and justified. This adjustment procedure has been applied to
into a circuit without autotransformer. actual systems of several hundred nodes, and convergence has
been obtained in about twice the time of an adjusted feasible
solution by Newton's method of power flow. The practicality of
real-time reactive optimization has thus been demonstrated.
In addition to this main result the very significant additional
information contained in the dual (Lagrangian) variables has
been emphasized. The computation of these variables is very
simple once a program capable of solving the power-flow equa-
tions by Newton's method exists. The dual variables not only
determine the gradient of the objective function (i.e., the line
losses) with respect to the electrical control variables and thereby
suggest an adjustment algorithm, but also measure the sensitivity
of this objective function with respect to various equality (i.e.,
Fig. 5. Autotransformer locations considered. demand) and inequality (e.g., equipment limitation) constraints.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:24 from IEEE Xplore. Restrictions apply.
 PESCHON et al.: OPTIMUM CONTROL OF REACTIVE POWER FLOW 47

In conjunction with the sensitivity relations between the elec- authors have presented a method attempting to optimize both kW
trical variables, which are readily available from the power-flow and kvar dispatches in order to minimize transmission losses and
solution, these dual variables lead to an efficient optimization in system operating cost. Their efforts and contributions should be
the presence of indirect constraints of great importance in the commended.
general optimization problem. However, in practical application to many power systems, the
objective for optimizing kvar dispatch may not always be to mini-
The reactive optimization described in this paper has been mize transmission line losses. Some of the related problems are:
applied to the mathematical models of actual systems but has 1) Contrary to the kW output, the kvar sources of a power system
not been used in the real-time operation. In addition to an are not all from generators. To optimize the kvar dispatch of all
efficient optimization algorithm, a real-time control system available sources, consideration must be given to line and station
requires careful consideration of the control actions to be taken equipment loading limitations both under normal operating condi-
in the presence of such imperfections as sensor inaccuracies, tions and under conditions when a failure of certain system facility
transmission system model inaccuracies, noninfinitesimal sam- occurs.
pling times, and generally speaking, perturbations resulting from 2) In many systems, the bulk-power system voltage and reactive
faults. These practical problems have all been studied and are control is employed to supplement distribution area voltage regula-
tion in order to optimize total system reactive capacity and voltage-
reported by Peschon et al. 161 regulating equipment requirements from generators to distribution
Perhaps the main significance of the effort described in this feeders.
paper derives from the extension potential of reactive optimiza- 3) During light-load periods, transmission system voltages must
tion to the economically more important problems of combined be lowered to reduce the effects of charging capacity and the dis-
active and reactive dispatch optimization of pure thermal as tribution area voltage-bucking requirements.
well as mixed hydrothermal and pure hydro systems. This These problems must be considered in determining the system
subject will be discussed in a forthcoming paper. kvar dispatch in addition to minimizing transmission system losses.
Perhaps the authors may wish to include these considerations in
REFERENCES their future work.
['1 J. Carpentier, "Contribution a l'6tude du dispatching 6co-
nomique," Bull. Soc. Franc. Elect., ser. 8, vol. 3, August 1962. (For a
partial translation into English of this work, see J. Peschon, Tech-
nical Memorandum 3 to the Bonneville Power Administration
(BPA), SRI Project 5503, Stanford Research Institute, Menlo Park,
Calif., October 11, 1965, 19 pp.)
[21 H. W. Kuhn and A. W. Tucker, "Nonlinear programming," J. L. Carpentier (Electricit6 de France, Paris, and Public Power
Proc. Second Berkeley Symposium on Math. Stat. and Prob., University
of California Press, Berkeley, Calif., 1951. Corporation, Athens, Greece): The authors are to be congratulated
[3] J. Carpentier and J. Siroux, "L'optimisation de la production for developing this very efficient technique for solving injection-
l'Electricit6 de France," Bull. Soc. Franc. Elect., March 1963. (For a method relations (11)-(15), and I want to add my personal thanks for
partial translation into English of this work, see J. Peschon, Tech- their having translated some of my publications into English.
nical Memorandum 4 to the Bonneville Power Administration I appreciate the philosophy at which the authors have arrived.
(BPA), SRI Project 5503, Stanford Research Institute, Menlo Indeed, the numerical methods proposed in 1961 for solving injection-
Park, Calif., October 14, 1965, 16 pp.) method relations were based upon relaxation techniques as in the
[4] L. K. Kirchmayer, Economic Operations of Power Systems. Ward and Hale method for load flows; they appeared at that time to
New York: Wiley, 1958.
[6] J. Peschon, D. S. Piercy, 0. J. Tveit, D. G. Luenberger, be the only way to get the optimum in reasonable time.
"Optimal control of reactive power flow," Stanford Research However, since that time, the Newton method and techniques for
Institute, Menlo Park, Calif., interim report to BPA, SRI project exploiting the sparsity of the network admittance matrix appeared
5503, contract 14-03-53699, February 1966. for load flow problems; and it is sure that, with the possibilities given
[61 J. E. Van Ness and J. H. Griffin, "Elimination methods for by the latter techniques, the best way to solve injection-method
load flow studies," Trans. AIEE (Power Apparatus and Systems), relations is to include Newton's method in some manner; I think the
vol. 80, pp. 299-304, June 1961. authors are truly on the right path.
[71 W. F. Tinney and C. E. Hart, "Power flow solution by New-
ton's method," IEEE Trans. Power Apparatus and Systems, vol. Three programs were developed for solving injection-method
PAS-86, pp. 1449-1460, November 1967. relations with relaxation techniques:
[81
N. Sato and W. F. Tinney, "Techniques for exploiting the one with voltage magnitudes given in advance
sparsity of the network admittance matrix," IEEE Trans. (Power one general
Apparatus and Systems), vol. 82, p. 944, 1963. one general, with a more approximate technique for determining
[91 J. Carpentier, "Ordered eliminations," Proc. Power System the initial units.
Computation Conf., 1963. The first and third programs were used for large networks.
[10] H. Edelmann, "Ordered triangular factorization of matrices," As the authors of the present paper describe it, it was difficult to
Proc. Power System Computation Conf., 1966. obtain convergence in some cases because of the use of relaxation
['] A. Ralston, A First Course in Numerical Analysis. New York:
McGraw-Hill, 1965, p. 408. techniques, the strongest source of divergence being the existence of
[12]
G. Dauphin, G. D. Feingold, and D. Spohn, "M6thodes limits of transits on the lines.
d'optimisation de la production des groupes d'un r6seau 6lectrique," Now different techniques for solving injection-method relations or
Electricit6 de France, Bull. de la Direction des Etudes et Recherches, equivalent relations are under development.
vol. 1, 1966. Among them, the so-called differential-injection method, which
has already given very interesting results, seems promising for prac-
tical application, especially to control problems.
I would like to ask a question of the authors. Calling (at the
optimum):
m' number of nodes whereQi is neither minimum nor maximum
s' number of nodes whereVi is minimum or maximum,
Discussion it may be demonstrated (using, for example, differential-injection
theory) that g = m' - s' is always> 0. Moreover, I find that in
practice g always remains very small. Did the authors find the same?
W. S. Ku (Public Service Electric and Gas Company, Newark, N. J.): This is ratherimportant, since g measures to some extent the
Several methods have been widely used in the power industry to degree of nonlinearity of the problem: if g is naught, linear program-
optimize the kilowatt dispatch of generators by takinginto account ming techniques will succeed fully. If g > 0, nonlinear programming
incrementalfuelcost and incremental transmission losses. The techniques are necessary, with moredevelopments when g is greater.

MIanuscript received February 20, 1967. Manuscript received April 3, 1967.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:24 from IEEE Xplore. Restrictions apply.
 48 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS JANUARY 1968

It should be noted that for the general case, where active power is where the following two conditions hold:
also optimized, g = m + m' -s - s', with 1) There are few violations of indirect constraints; by this, we
m number of nodes where there exists a unit with active power mean that the voltages at consumption nodes do not tend to rise
neither minimum nor maximum above the voltages of adjacent production nodes.
s number of lines on limits of transit. 2) The number m' - s' is small, i.e., few reactive productions are
such that the corresponding node voltages are not at their maxima.
In the practical examples we have treated to date, these two condi-
tions hold. However, we have continued to search for optimization
methods that remain effective when either or both of these conditions
do not hold. These conditions are violated for lightly loaded systems
E. Jamoulle (EBES, Brussels, Belgium): It is doubtless interesting and other situations of practical importance.
to try to relate to a general methodology the systematic and truly By developing a procedure where the cost function and the power-
operational solution process presented by the authors of this paper. flow equations would be linearized about a feasible solution, the
It is obviously a gradient method. In effect, the proposed algorithm optimum solution would be obtained by solving the linear program
works in the field of feasible solutions, and each iteration gets closer operating on this linear model. This approach works well in the
to the optimum by moving the initial feasible solution in the direc- presence of indirect constraints but will not work unless m' - s' = 0
tion of steepest descent of the cost function. Although convergence is or unless the linear program is further constrained by artificial re-
very rapid, the method does not guarantee the exact theoretical strictions on some of the variables. The reason for this is that the
solution in a finite number of steps. solution of a linear program must lie on a constraint surface, whereas
However, the original and basic idea-which appears to be quite the condition m' - s' > 0 implies one or several of the problem
general-on which this procedure is based, is that the direction of variables are not constrained at the optimum.
steepest descent can be derived from the knowledge of the dual vari- The approach which we are now pursuing for solving the problem
ables of a restricted primal problem. On the other hand, the authors with indirect constraints efficiently is the well-known penalty-fune-
use Newton's method to solve the power-flow equations; the efficiency tion method where an appropriate cost term is assigned to each in-
of this method has been amply proved in recent years. direct constraint violation. This modification, combined with second-
Newton's method linearizes the problem at each successive point order gradient procedures for accelerating the convergence of the
reached. It uses the fact that Euler's equations constitute the dual optimum seeking algorithm, appears quite promising. An additional
system, in the sense of linear programming, of the primal linearized advantage of this method is that the sensitivity of cost to inaccurate
system at the operating point. input data (here C and D) can be readily assessed, and a perturbation
The optimality conditions of Kuhn and Tucker are identical in this law of dispatching of the form: Au = KAp; K = constant matrix,
sense, with the so-called complementary variation equations, a classic can be readily established from the results of the second-order
of linear programming. gradient procedure.' In the preceding equation, Au is the change
This really interesting property, which can be credited to the required in the independent (control) variables for an observed small
solution of the power-flow equations by the Jacobian method, thus change Ap in the demand vector (C, D) to remain at the optimum.
makes it easy to find by a direct route, bearing in mind the par- Our practical experience with this penalty-function approach has so
ticular problem being studied, the marginal costs sought and, there- far been very good.
fore, the direction of steepest descent. In reply to Mr. Ku's discussion, we wish to make the following
Finally, from a theoretical and practical point of view, the pro- statements:
cedure used seems to be entirely suited for extension to the general 1) The line and station equipment loading limitations are ac-
case of short-term optimization of active and reactive powers in an counted for by the inequality constraints [see (2a)-(2g)] under
electrical network. normal operating conditions. Under abnormal conditions, i.e., when
an important system facility has failed, the first requirement is to
Manuscript received February 23, 1967. prevent the system from collapsing, and the most economic opera-
tion of the system becomes a secondary consideration. The optimum
dispatching program described in our paper is in no way concerned
with the operation of a faulted system, but the same mathematical
approach can be applied to optimize the system's performance after
the occurrence of a fault.2
J. Peschon, D. S. Piercy, W. F. Tinney, 0. J. Tveit, and M. Cuenod: 2) The method presented can be readily extended to the situation
We would like to thank Mr. Carpentier, Mr. Ku, and Mr. Jamoulle referred to by Mr. Ku where the bulk power system voltage is
for their helpful dliscussions. In reply, we wish to make the following employed to supplement distribution area voltage regulation. One
points: only needs to prescribe the voltage Vi or its permissible range
Mr. Carpentier points out the convergence difficulties he encoun- (Vim, ViM) required at node i and search for the optimum reactive
tered by applying to the optimization equations a relaxation method productions sutbject to these constraints. There remains a reactive
of the Gauss-Seidel type. At the beginning of our study we attempted optimization problem as long as the number of reactive production
to apply to the reactive optimization this relaxation method sug- sources exceeds the number of nodes with prescribed voltage.
gested by Mr. Carpentier in his well known paper, and we were 3) The required lowering of transmission-system voltage during
plagued for well behaved systems by diverging answers whenever the periods of light load is also taken into consideration by the method
initial guessed solution was not sufficiently close to the optimum presented in our paper, specifically by the inequality constraints.
solution. For systems with unusually high or low reactive consump- The main contribution of Mr. Carpentier's problem formulation for
tion, these relaxation methods refused to give the optimum solution, which we developed a computationally practical solution is precisely
even when the initialization was very close to the optimum final re- the inclusion of inequality constraints to account for equipment
sult. The main reason, however, for discarding the relaxation method limitations and operating restrictions.
was the enormous amount of programming logic required to force the Finally, we are happy to read Mr. Jamoulle's point about the dual
correction upon a variable that had not yet reached a constraint. variables (marginal costs) being obtained as a matter of course
This last difficulty is characteristic of reactive optimization where during the optimization computations. In our opinion, knowledge of
the optimum solution is located on numerous voltage or reactive these sensitivity variables may well turn out to be more important
constraints but not of active optimization where the optimum active to the system operator than the additional savings obtained from an
productions are seldom located on the P constraints. exactly optimum dispatching scheme.
We agree with Mr. Carpentier that there are several possible
ways to solve the optimum dispatching problem. The procedure
presented in our paper constitutes a fairly standard first-order
gradient method where the dual variables of a feasible solution are ' J. Peschon, D. S. Piercy, W. F. Tinney, and 0. J. Tveit, "Sensi-
used to find an improved set of independent primary variables. tivity in power systems," presented at the PICA Conference, Pitts-
Our experience with this method has been very good in those cases burgh, Pa., May 15-17, 1967.
2 L. P. Hajdu and J. Peschon, "Optimum load shedding policy for
power systems," Stanford Research Institute, Menlo Park, Calif.,
Manuscript received April 3, 1967. Final Report for BPA, under contract 14-03-64089, February 1967.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:24 from IEEE Xplore. Restrictions apply.