2864

IEEE Transaction on Power Apparatus and Systems, Vol. PAS-103, No. 10, October 1984
OPTIMAL POWER FLOW BY NEWTON APPROACH

Art Hughes David I. Sun Bruce Ashley Brian Brewer Sr. Member Member Member Member Bellevue WA 98005 13010 Northup Way ESCA Corporation

William F. Tinney Fellow

Consultant

flow Abstract - The classical optimal power problem with a nonseparable objective function can be Efficient, solved by an explicit Newton approach. robust solutions can be obtained for problems of any is effort or kind. Solution size practical approximately proportional to network size, and is relatively independent of the number of controls or The key idea is a direct inequalities. binding simultaneous solution for all of the unknowns in the Lagrangian function on each iteration. Each iteration minimizes a quadratic approximation of the Lagrangian. For any given set of binding constraints the process converges to the Kuhn-Tucker conditions in a few The challenge in algorithm development is iterations. to efficiently identify the binding inequalities.

performance requirements that include: a solultion time that varies approximately in proportion to network size and is relatively independent of the number of controls rapid and consistent constraints; or inequality optimality (K-T) Kuhn-Tucker convergence to the absence of user supplied tuning and conditions; no scaling factors for the optimization process; compromises in OPF problem definitions; and problems of any practical size or kind should be as easily solvable by an OPF as by a PF. Newton-based OPF programs can satisfy these criteria for practicality.
the is a idea of approach The key sparsity-oriented, simultaneous solution for all of the unknowns of a quadratic approximation of the Lagrangian General large-scale nonlinear each iteration. on by solved are usually problems optimization quasi-Newton rather than explicit Newton methods since the latter would be too burdensome or even completely The OPF is a notable exception; the intractable. sparse Hessian matrix of the Lagrangian function can be evaluated and operated on efficiently. explicitly decoupled by Efficiency can be further enhanced convergence remains the which for formulations superlinear. For a given set of equalities, a Newton the K-T conditions in a few converges to OPF The iterations. major challenge in algorithm development is to identify the binding inequalities efficiently.

INTRODUCTION
the approach to power flow (OPF) problem based on an explicit Newton formulation. Tests performed on large problems with prototype algorithms show that the approach fulfills all of the requirements The Newton formulation for practical OPF programs. appears to be as fundamental and effective for OPF as it, is for power flow (PF).

solution of the classical optimal

This

paper

describes

new

The classical OPF is a PF problem in which certain controllable variables are to be adjusted to minimize an objective function such as the cost of active power generation or losses, while satisfying physical and dependent controls, operating limits on various variables, and functions of variables. Because the objective must include losses, and the controls include reactive devices, the problem is characterized by a non-separable objective function. This characteristic, which sets the classical OPF apart from similar optimization problems, also makes it more difficult to solve.
accommodate

a oei controls, power power controls, reactive combination of both may be optimized. Because of the non-separability, optimizing reactive power control is more difficult than active-power control. In general, any method that can solve the reactive problem can also solve the combined active-reactive problem. Therefore, the emphasis of this paper is on the reactive problem.

The types of controls that an OPF must be able to power and reactive include active injections, generator voltages, transformer tap ratios, and phase shift angles. In a given OPF study, active

different OPF algorithms suited to the requirements of Although the Newton approach different applications. exists as a concept entirely apart from any specific method of implementation, it would not be possible to employing without develop practical OPF programs The concept and the special sparsity techniques. techniques together comprise the given approach. Other Newton-based approaches are possible.

flexible

As

described here, the Newton approach is a formulation that can be used to develop

The paper describes the interim results of an ongoing research project, RP1724-1, being performed by ESCA Corporation under contract with the Electric Power Research Institute (EPRI). The purpose of the project is to explore OPF solution methods. More work has been done than is reported here and more work remains to be Other papers done before the project is completed. describing project results are planned. NOTATION
used are The following definitions symbol throughout the text. Most symbols are also defined in the text where they first appear. Symbols used only in the APPENDIX are defined there.

To be practical, OPF programs need to have

A paper recommended and approved 84 WM 044-4 by the IEEE Power System Engineering Committee of the IEEE Power Engineering Society for presenta-

k, m

A
wrt
t

IEEE/PES 1984 Winter Meeting, Dallas, Texas, January 29 - February 3, 1984. Manuscript submitted September 1, 1983; made available for printing November 18, 1983.
tion at the

Prefix on scalars and vectors denoting incremental correction of quantity. "with respect to". Superscript meaning transpose.

Subscripts denoting buses (nodes) or bus pairs.

0018-9510/84/1000-2864$01.00 1984 IEEE

2865
Matrix entry denoting a large number.

Scalars

Qk vk

Pk

Yk
F L

t; km 0km

E)k

Xpk

Pi

Xqk

N S ol i Vectors
y X
z

Active power injection at bus k. Reactive power injection at bus k. Magnitude of complex voltage at bus k. Angle of complex voltage at bus k. Transformer tap ratio between buses k and m. Phase shifter angle between buses k and m. Symbol for any state or control variable. Objective function Lagrangian function. Lagrange multiplier for Pk. Lagrange multiplier for Q Lagrange multiplier for active inequality constraint i. Number of buses in network. Quadratic penalty weighting factor. Quadratic penalty function for inequality constraint i.

published. Although some success has been claimed for these methods, recent findings show that gradient methods cannot solve the OPF. This was convincingly shown in a recent paper by Burchett, Happ and Wirgau [6] in which they compared a reduced gradient method with a quasi-Newton method which is strong enough to solve the problem.
Sasson, Viloria and Aboytes [7] were the first to show that a sparse factorization of an explicit Hessian matrix for the OPF could be performed. However, they attempted to solve the PF equations as well as to minimize the objective function by using only the Hessian matrix in an augmented Lagrangian formulation. With Evidently, this does not work well. their formulation, the Hessian matrix had second-neighbor fill-in, making it considerably less sparse than the Hessian matrix of the approach described in this paper.
The aforementioned quasi-Newton method of Burchett, et al [6) utilizes second order information contained in an iteratively constructed reduced 'Hessian matrix. It is strong enough to solve the OPF. In each iteration the quasi-Newton method obtains a descent direction by operating on the reduced gradient with an approximation of the factors of the symmetrical but dense reduced Hessian. (The reduced Hessian has the dimension of the superbasic variables). Updating the dense approximate factors in each iteration, and the number of iterations required for the approximate Hessian to generate a good descent direction, contribute to a large and computation storage requirement for problems of practical size.

"p Iq

,u g z

z"l

g"f
Matrices

All variables v, E, 0, t. All Lagrange multipliers X-pand Xq Subvector of X. Subvector of X. Composite of subvectors y and X Lagrange multipliers for binding inequality constraints. Gradient of L wrt z. Subvector of z for real power variables. Subvector of z for reactive power variables. Subvector of g for real power variables. Subvector of g for reactive power variables.

PRELIMINARIES
This section briefly describes how the basic aspects of nonlinear optimization are applied in the given approach.

H J W

W'

W"l H' J H" d"l

Hessian of the Lagrangian. Jacobian for OPF. Bordered Hessian. Composite of H, Jp jt. Bordered Hessian for PO subsystem. Bordered Hessian for Qv subsystem. Hessian submatrix of W'. Jacobian submatrix of W'. Hessian submatrix of W". Jacobian submatrix of W".

Example Problem
The five bus network of Fig. as a specific example.
1

1 is used throughout

2
j

/0-l
3

BACKGROUND
been reviewed Progress on OPF analysis has periodically [1-3], and it was reviewed for this project. Attempts to solve the OPF problem date back over twenty years. Practical solutions for OPF problems with separable objective functions have been obtained with special linear programming methods [4], but the classical OPF has defied practical solutions. Of the many proposed methods, only a few have been tested on problems large enough to evaluate their performance. It seems that all of these methods fall short of being practical because of limitations in speed, problem size or robustness, or because of compromises with This the problem definition. conclusion is supported by the observation that the classical OPF has not become a standard application. T24

T35

45

Fig. 1

Example Network

Generators at buses 1 and 3 are dispatchable sources of active and reactive power. Load buses 2, 4, and 5 have scheduled active and reactive power. The phase shifter has controllable angle 04.5 and the transformers have controllable tap ratios t24 and t35. The reference angle of the complex bus voltages is ej.
The problem is to minimize the total cost of P1 and P3 while satisfying the scheduled loads and limits of different variables. P1y Ql P3Y Q3, 045, t24 and t3s are controllable.

Rather than attempt to evaluate many different methods, only three are discussed. They are relevant to the Newton approach and are representative of
others.

The reduced gradient method of Dommel and Tinney [5] has been frequently cited as a benchmark.. Several other reduced gradient methods have also beern

Variables
can

In an optimization procedure the variables change roles. Variables are classified as

2866
contGrol (superbasic), dependent state (basic) or, if they have reached their limits, constant (nonbasic). are these distinctions approach In the Newton Basic and superbasic variables are unnecessary. processed identically. Variables that become nonbasic constrained at their limiting values without are The actually changing their roles as variables. original variables establish the dimension of the vector of variables, y, and it is never changed.
their limits. In (2), L is a linear combination of PF equations. the Newton approach inequality constraints on In variables and functions are enforced by quadratic penalty functions ai which augment L as shown in (3) .
L = CP

ii1

P 33 3C

XPpk

2 2Q iakqkk

(3)

045, t24, t35g,

In the example, y consists of thirteen variables; 1, vl, e2, v2, 03, v3, e4, v4, 85, and v5. Although O1, the reference angle, remains constant, it is included in y for completeness.

The summation for i is over all binding inequalities.

L for the example, with no active penalties and omitting Q1 and Q3for dispatchable VAR sources, is,

Inequality Constraints
The following quantities have upper limits; 1. 2.

and/or

lower

lI l

33 3

p11

p2 2
-

Ap3 3
-

p4P 4
phase

p5P5

Xq2Q2 Xq4Q4 Xq5Q5

(4)

Dispatchable sources of P and Q.

Variables; angles.
voltages,
tap

ratios,

shifter

In some optimization methods L is augmented by other equations to enhance its positive definiteness. Such augmentation, which adversely affects sparsity, is unnecessary in the Newton OPF.

3.

Functions such as line flows, measures of security,

etcc.

Solution Conditions A minimum of the objective function occurs the K-T optimality conditions are satisfied. the evaluating for quantitative indicators conditions require no additional computation. following conditions must exist for a minimum:
1. 2.
The mismatches of all PF equations within tolerance.

Active Equalities
The set of equalities, A, always includes the PF equations for scheduled load and generation. It also includes the following set of binding inequalities: 1. 2.
The PF equations of any dispatchable sources or Q that are constrained at their limits.
The equations of any other constrained at their limits.

when The K-T The

A are

in

set

of

The inequality constraints are all satisfied.

inequality

functions
at

3.
4.

The projected roundoff).

The

gradient

is

zero

(except

for

3.

The trivial equations of variables their limits.

constrained

Objective Function
The objective function, F can assume several have no the but differences forms, different significant impact on the approach. F for the example is given in (1).
F=CP

sensitivity, Pit between each binding inequality constraint and the objective function is such that further cost reduction can be achieved only if the constraint is violated.
The projection of the Hessian in the feasible region is positive definite. Conceptually, this means that the multi-dimensional objective function is bowl shaped; therefore, the stationary point is a true minimum, not a saddle point.

5.

I11

+C

3 3

(1)
a stationary point is a minimum, that it is a global minimum.

C I and C 3 are the slopes of piecewise linear segments the active power cost curves of dispatchable of In an actual program suitable generators 1 and 3. logic would be needed to change C1 and C3 as P1 and P3 change in the solution process. This logic is omitted here because it has no impact on the approach. If desired, quadratic approximations of the cost curves could also be used in F.

Although local, positive definiteness

it

does

ensures that not ensure

SUBMATRICES This section describes the Jacobian and Hessian matrices, J and H, and displays them for the example J and H are submatrices in the Newton problem. formulation. Their arrangements in the displays in this section are conceptual only. In an actual implementation they are combined and rearranged into a single matrix or matrix pair. However, the arrangement H shown would be correct if its separate for factorization were needed.
Jacobian Matrix
J is a matrix of first partial derivatives of equations wrt y. In the Newton approach the number rows -in J is always 2N and the number of columns equal to the dimension of y. J for the example shown is (5).

Lagrangian Function
The Lagrangian, L, for the Newton OPF is shown
(2) .

in

CP .1 1

+CP 3 3

ljk Xp'Pk'

'qkQ~k

(2)

The summations are over all buses from 1 to N. However, the Lagrange multipliers, Xql and Xq3, become nonzero only when their respective equations for QL and in A. The values of the Lagrange multipliers are In the are determined by the solution process. five bus example, X l, and X 3 are zero because they operating within correspond to dispatVhable

s%drces

PF of is is

2867

Hessian Matrix
+45 124 T3s 01
P1
01

un

02 V2 03 5/3 04

5/4 05 "55

i J

derivatives

P2 (2
P3

i'

second partial symmetric matrix of of L wrt y. Its dimension is the same as that of y. Each element of H is a linear combination of second partial derivatives of PF equations. Details on forming H are given in the APPENDIX. H for the example is shown in (6). H is
a

03
P4

Each symbol h indicates a nonzero element and its a specific derivative. coordinates indicate For example, h with coordinates t24,v4 is 82L/8t E)v Only the upper triangle is shown. The sparsity of the 2x2 block structure of the main submatrix (heavy border) is the same as that of the network incidence matrix. Therefore, it can be factorized in the usual sparsity-directed way. Processing of elements outside of the heavy border can be effectively separated from the processing of the main submatrix. This will be clarified in the next section. H by itself, however, is not factorized in the Newton approach.

Q4
P5
05

The rows for currently inactive PF equations can be omitted from J. Each symbol j indicates a nonzero element and its coordinates indicate a specific derivative. For example, j with coordinates Q4, v5 is aQ4/av5. A nonsingular Jacobian matrix for the Newton PF could be formed by deleting certain columns of J. +45 T24 T35 01
+4!

VI

02

V2

03

V3

04

54

05

V5

|h
4

h
h h
h

h
h

h
h

h
h

Elements of H representing the couplings between variable pairs e and v, G and t, and and v are very small compared to the average magnitudes of other elements. This favorable property is exploited in the decoupled formulations. NEWTON OPF FORMULATIONS In this section the coupled and decoupled Newton OPF formulations are explained by temporarily ignoring the complications of identifying and enforcing the binding inequalities. The only active constraints assumed at this point are the PF equations for the given loads. (Without some additional constraints the problem would be unsolvable since the optimal solution would normally be far outside any normal operating range.)

T24
T31
01

h
h

h h
h

V1

Ih
h

h
h
h

h
h

h
h

h
h

V2
03

h h

h
h

h
h

h
h

v3
04

h
h
I

(6)

h h
h

h h

v4

05
v,5
I
h

4 h h j

AZ

-9
-auat45

h
h

h
h

j
j

j
j

04!
AT2d

h
h

hhjj
h hj h

h hjj

-au,T24
-aUaT8
-a

AT,

oo
h h h

asp
j
i j i h h h h h h h h

SL

AV2
AV2

jj

-luav2

-auaV2
(9)

AA,2

A9q2 h.
hjh-h hj
h
j

h h

jhj

Ao3
bv3

-auag,
-a Lav3
a kp

hj

a,p
h
4

jj

AG4
by4

-a uao,

hj

h h

,AA
AG4

-a-aua1y Lla A94

-OLOO,
5

Av5 AV!,
bAAs
AAq5

-aLlA,S -a

UaAq,

2868

Coupl ed Formulation
Let, z be a vector composed of subvectors y and X . Differentiating L twice wrt z leads to the symmetric matrix W symbolized in (7).

7.

Return to step 2 with updated z.

Decoupled Formulation

W =l

Jt]
(7)

symbolized in (10),
WI
W"t
x
x

A decoupled

formulation

of

the

Newton

OPF

is

= Az, -'t

(lOa)
( 10b)

l-J

Azl

= _g.

This matrix is well known (pg. 239 of ref. (12) but apparently unnamed. Here it is referred to simply as W. W provides the basis for the Newton OPF formulation shown in (8),
W
x

where the symbols are as defined in the NOTATION section. The primes on symbols are used to suggest an analogy with the decoupled PF [9]. The sparsity arrangement of (10) for the example is shown in (11).
1

Az

-g

(8)
1

-a

Lla+,,

where g is the gradient vector of first partial derivatives of L wrt z, and Az is a vector of Newton corrections in z.

2 3
4
h

A)2

-aLlao2
-a L/ a X5

For sparse factorization it is necessary to rearrange W. Equation (9) on the previous page shows the sparsity arrangement for the example.

AkP2

-a uae3

(lla)

Equation (9) also shows in detail the elements that comprise the systelb% First derivatives of L wrt X in vector g are the familiar residuals of the PF equations, i.e., the differences between the actual and scheduled injections of Pk and Qk.

AO,3
h
j

AS3

a l

AS4

AS5

-aual,4 I-aL/ak

-aL/aos

external submatrices can be effectively separated from processing of the main submatrix because elimination of the external variables only modifies certain nonzero blocks within the main submatrix. Furthermore, the ordering of the external variables has no effect on sparsity. (These same remarks also apply to the arrangement of H in (6).)

borders) whose 4x4 block structure is the same as that of the network incidence matrix. Processing of the

This arrangement creates a main

submatrix

(heavy

oo
h_
h

AT,4
AT35
h
Avs

-aLUaT24
-aL/aT35

I-aLlav
-a L/av2 -aLalq2
-a

hjjhh hj h h j
h

AV2
Ax3
As4 AAq4

hh

3
4

LIav3

((lib)

times as much computational effort as the same operations with the power flow Jacobian, and the matrix storage requirement is double. Efficiency and storage requirements can be further improved by developing decoupled formulations.

four

Factorization and repeat solution of W requires

-a L/as1
-

L/aAq4

-5

As5
AAq5

I-OuaxV

Coupled Solution
The coupled OPF, subject only to equality constraints, can be solved by the following simplified algorithm;

1.

Make starting estimates for z

(y,X).
PF.

y can be the same as the starting values for a X can be zero or any reasonable guess.

The decoupling shown divides the problem into Pe and Qv sysbystems. The elements of H that are omitted in the decoupling are negligibly small; the elements of J that are omitted, which are the same as those omitted in the decoupled PF [9], may not be negligible in some problems. Any defects that might arise from this particular decoupling would be due to the approximation for J, not H. There are remedies for defects due to decoupling of J. Other versions of decoupling are possible. For example, versions based on decoupling H but not J. The sparsity of the 2x2 block structure of the main submatrix of W' and W" (heavy borders) is the same as that of the network incidence matrix. The remarks about processing of the external submatrices for the coupled version also apply to this decoupled version. Factorization of W and W" together requires approximately the same computational effort as factorization of the Newton PF. The matrix storage is also similar.

2.

Evaluate g as a function of z.

3.

Check for solution.

If K-T conditions are satisfied, OPF is solved. Else

4.

Evaluate W as a function of z. Factorize W and solve for Az.

Update z by Az.

5.

6.

2869

Decoupled Solution
The decoupled OPF, subject only to equality constraints, can be solved by the following algorithm:

of Xqk indicates whether Qk would return to feasible range if its equation were again removed.

its

1. 2. 3. 4. 5. 6.

Make starting estimates for z. Evaluate gt and W' as functions of z. Factorize W' and solve for Az'. Update z by Az'. Evaluate g" as function of updated z. Check for solution. If K-T conditions are satisfied, OPF is solved.

The PF equations themselves provide only hard enforcement. Methods for soft enforcement of limits on power sources have been developed but their explanation is beyond the scope of this paper.

Enforcing Limits on Variables

Any variable can be made into a constant by simply eliminating the equation for its correction and setting the variable to the desired value. For example, assume voltage at load bus k is below its minimum. The limit could be enforced by setting vk to its limit and making roW/column for vk in W" a dummy. The equation for Qk remains in W" to enforce its scheduled value. The change of variables to constants, or vice versa, can be made in this manner whenever W' or WI' is refactorized, but more efficient methods are needed for Also, enforcing one or a few inequalities at a time. as pointed out, uncoordinated changes in variables should not be made. A variable should be moved to its relevant limit in coordination with the simultaneous correction of all other variables.

7. 8. 9. 10.

Else Evaluate W" as function of updated z. Factorize WI' and solve for Az". Update z by Az". Return to step 2 with updated z.
ENFORCING INEQUALITY CONSTRAINTS

This section discusses methods for enforcing inequality constraints. Methods for determining which inequalities to enforce are discussed in the next The following attributes are advantageous in section. inequality constraint enforcement methods:

1.

enforcing The method should be efficient for inequalities singly or a few at a time without refactori zing.
The method should provide for automatic control of the hardness of constraint enforcement. There are situations where a certain amount of freedom around Certain an is beneficial. inequality limit strategies for identifying the binding inequalities require control of the hardness of enforcement. Soft limits may also be needed to obtain useful nonfeasible solutions when feasible ones do not exist.
The method should not make uncoordinated step changes in the variables or functions in order to set them to desired values. All such changes should be made in the simultaneous solution of the changes Uncoordinated correction vector Az. disrupt convergence.

2.

Quadratic penalty functions ideally fulfill the requirements for inequality constraint enforcement in second-order methods. This is in marked contrast to the erratic and generally unsuccessful behavior of linearized quadratic penalties used in gradient-based OPF 'methods. In a second-order method the effect of a It quadratic penalty is accurately controllable. provides a two-sided constraint that can clamp a or allow a variable at an exact target value controllable amount of freedom for variation around the target value. It also provides the means for moving a to the target value in variable or function coordination with the simultaneous correction of all other variables. Equation (12) shows the quadratic penalty function used to constrain variable yi.
i.

(Y=

(12)

where Yi is the target value, yi is the current value and Si is a weighting factor that is automatically controlled to give the appropriate amount of hardness of enforcement. The first and second derivatives of di are:
1 S.

3.

dy.
2

(y.

Yi)

(13)

dy i 1

(14)

Inequality constraints are considered in three (1) Dispatchable active and reactive categories: power, (2) variables, (3) functions of variables.
Enforcing Limits
are
on

Dispatchable Power Sources

VAR sources The PF equations for dispatchable are feasible, omitted from A as long as they the and restored to A when necessary to prevent Qk and removing from infeasible. Adding becoming these equations changes in causes no structural As an example, suppose that at some W' and W". iteration dispatchable reactive source at bus k exceeds Then the equation for AXqk would be its maximum. activated by entering the current evaluation of its elements in its previously dummied row/column in W". Since Qk had been feasible up to this point, Xqk would be currently zero, and this is the value that would be next After the used in the initial evaluation. iteration, Xqk would assume a nonzero value. The sign

Application of the penalty will force yi to coincide with or remain as close to yi as desired. The first Its the L. quadratic penalty augments of g and its second derivative is added to aL/ayi derivative is added to a2L/ayi of W.
is The proper value for automatically Si controlled. If Si is large, Yi acts as a hard limit and If Si is yi moves arbitrarily close to the limit. small, Yi acts as a soft limit. The corrections in yi are coordinated with all other corrections on each in W" or W" is used iteration. The value of a2/ayi in automatically adjusting Si to the proper value for a soft limit. Any Si larger than a certain value will effectively produce a hard limit.

After a penalty has been imposed to constrain a variable, the sign of its derivative can be used to decide whether the penalty is still needed. -The

2870
derivative of a penalty on a variable is the negative of its Lagrange multiplier, pi, for the enforced limiting The derivatives of imposed penalties are also value. used to determine whether the K-T conditions are satisfied.
for the Lagrange multipliert enforced 1l inequality. It indicates whether it is necessary to continue enforcing the inequality, and it is also used in determining whether the K-T conditions are satisfied.

Quadratic penalties eliminate uncoordinated changes in variables, permit control of the hardness of constraint enforcement, indicate whether imposed constraints are still binding, and provide the incremental costs of the imposed constraints.
Enfor Limits on Special Functions

Individual Enforcement of Inequalities

The enforcement or removal of a single inequality constraint involves only a few small changes in W' or W" and these changes affect only a few rows of the factorization. There are two sparsity methods for obtaining repeat solutions for small changes in a factorized matrix without having to perform a complete refactorization: partial refactorization and compensation. Either or both can be used to impose or remove inequality constraints singly or a few at a time without refactorizing W' or W"I. Partial refactorization exploits the property that small changes in a sparse matrix affect only a few rows of its sparse factorization. For simple changes, which occur most frequently, a special factor updating scheme [10] is most efficient. For more complicated changes the normal factorization is used but applied only to the affected rows. To impose or remove a PF equation or a penalty for an inequality, proceed as follows:

It is necessary to be able to enforce limits on special functions. One example is the function defining the flow of power through a line. Limits on such functions could be enforced, when needed, by adding their equations explicitly to the Lagrangian. But this would be undesirable because it would radically change the favorable and otherwise constant sparsity structures of W' and W". It is much simpler and equally effective to enforce such functional inequalities with penalties. The following example of An inequality constraint on the flow of power through a line illustrates the general case. Other functions defining inequalities can be handled in a similar way.
The power flow through line (km) can be held approximately within a specified limit by requiring the power angle ( -0k Om) to be close to some precomputed value 9km. If this is not accurate enough, Okm can be iteratively adjusted to achieve any desired degree of accuracy. The quadratic penalty function i for enforcing an angle limit 8km is:
.~~~~~~~. S;.
2

1.

Modify the affected rows of the factors of W by partial refactorization.

or W"

2.

Modify g' or g" by adding or removing the derivatives for the penalty or equation change.
Perform a fast forward solution [11] using only the rows of W' or Wt affected by the updating.
Back solve to obtain the updated solution.

3.
4.

(8k

)2

(15)

where Si is

weighting factor.
are

The first and second derivatives of i in (16).

given

dk
k

dui

,
-S
i

(8
i

- e

(16a)

Small matrix changes followed by updated solutions can be performed by this scheme repeatedly. The updates in the matrix and solutions are cumulative. The efficiency of each new update is unaffected by the

preceding updates.

i
m

(e

_e )
m

(16b)

1.
ao

(16c)

2g dk

S.
1

(16d)

d0 m 2

With compensation [111], which is similar to the Schur complement technique [13) used in optimization literature, the factors are not altered. The effect of matrix changes on the solution is computed directly from sparse compensation vectors. 'Successive matrix changes are not cumulative; each new solution by compensation is a change from the original base solution. successive can be However, changes accumulated in the compensation vectors to make solutions for cumulative changes more efficient. The efficiency of the method declines as the number of successive changes increases. Compensation is more efficient than partial refactorization to a up crossover point that depends on the total number and type of changes.

d2a.

d2u
d_

I
m
=

-s
i

d8 dB

m k

d8

(16e)

DETERMINING THE BINDING INEQUALITY SET

In decoupled schemes the PO and Qv subsystems are solved in alternating cycles. Some of the most effective algorithms tested thus far perform fast trial iterations within each cycle to determine the currently binding inequalities. Depending on the application, the scheme of trial iteration used in one cycle may differ from the one used in the alternate cycle. Schemes may also differ depending on the stage of the In solution process. the following simplified descriptions the particular cycle is not specified. Each algorithm could be used for either the PO or Qv cycle. Thus, instead of specifying W" or W", the The symbol W is used to mean either or both.

derivatives of i and add them to corresDondina derivatives in g. Add +S, to a2L/3)k

first

To activate a penalty for

and add

(Ok

0m)

evaluate

penalty only augments existing non-zero terms in W'. Any number of penalties such as this can be imposed with no burden on the matrix. Si can be adjusted automatically to suit requirements. The first derivative (the two are equal except for sign) is the

a2LI/a9,

-Si

to

a2L/aHk36m

of W'. Note that the

the the and

2871. the intent is only to descriptions are incomplete; for the binding suggest strategies finding inequalities. algorithms described employ two kinds of solutions of the decoupled quadratic approximations of L: main iterations and trial iterations.
A main iteration for either cycle is as follows:

Algorithm II
1. - 3.
Same as in Algorithm I.

The

4.
5.

Find scalar K (Ki1.O) such that at z + K(Az) only one inequality has moved to its limit.

Update z by K(Az).
Proceed to alternate half cycle.

1. 2.

Evaluate W and g as functions of the latest of z and A. Factorize W and solve for Az.

update

6.

A trial iteration is performed either by partial refactorization or by compensation. The aim of trial iterations is to identify the currently binding In each trial the effects of different inequalities. combinations of constraint enforcement and release at a tentative solution point are examined, Various strategies are employed to minimize the number of trials needed to find the binding set. When the currently binding inequalities have been satisfactorily identified, the solution is advanced to the next step A trial with the binding inequalities enforced. iteration is always fast compared to a main iteration, but the ratio of their speeds depends on the number and kinds of matrix changes involved in the trial and the amount of monitoring required to examine the tentative solution.

Except for the effect of the alternate decoupled cycle, Algorithm II would maintain strict feasibility. It would, however, be prohibitively slow. It could be greatly improved by adjusting K to allow for a A controlled amount of violation on each iteration. merit function [12] could be developed to determine a good value for K by taking into account the kind, the violations. number and of Many magnitude variatiQns of the merit function and the strategy for its application are possible.

Algorithm III
1. - 3.
Same as in Algorithm I.

4.

Tentatively update z by Az.

Check inequalities at z + Az.
a.

Example Algorithms

5.

algorithms

how aim of this section is to suggest based on the approach can be developed. Three rather fundamental algorithms, each of which Other could have many variations, are outlined. different algorithms are also possible. The

proceed to alternate half cycle.

Else
b.

acceptable iteration, advance the solution to the tentative. step; return to step (1) and

If the inequalities satisfy the criteria for an

Algorithm I
1.
2.

Check for solution.

Check inequalities.
a.

If not solved, continue.

According to the trial strategy select a set of violated inequalities for enfQrcement and a set of enforced for inequalities previously release. Modify A according to the selection.

Change A for inequalities.

enforcement

of

all

violated

Continue.

b.

inequalities needed.

Change A to

release all previously enforced where enforcement is no longer

6.
7.

Enforce and release the selected inequalities by trial iteration to obtain Az(trial). Return to step(5)with
z =

Az(trial).

3.
4.
5.

Perform main iteration to obtain Az.

Update z by Az.
Proceed to alternate cycle.

Algorithm I is analogous to the schemes used -in PF programs to identify and enforce conventional inequality constraints. It does not maintain strict feasibility but allows constraints to be, violated before enforcing them on the following iteration. Constraints are enforced and released simultaneously in As observed in PF programs, simultaneous groups. schemes succeed in identifying the binding inequalities but they take many iterations. Algorithm I could be improved by starting with soft penalties and systematically hardening them as the binding set becomes clarified. different Many simultaneous enforcement strategies are possible.

Algorithm III maintains a specified amount of At the end of each series of trial iterations following a main iteration the enforcement of inequalities is exactly or approximately resolved depending on the criteria for an acceptable iteration. In the effect, trial iterations test different combinations of enforcement in order to find the best one. Many variations of Algorithm III are possible.

feasibility.

Criteria For Enforcing Inequalities has The enforcement of different types of inequalities different effects. The effects are approximately predictable from the known behavior of power systems and the solution process. This knowledge can be used in establishing criteria for enforcing inequalities and in developing merit functions. When a violated voltage limit at
a

load

bus

is

2872 enforced, it tends to produce changes in the aame direction in other voltages in its area of the network. Simultaneous enforcement of hard limits on several violated voltage inequalities in one area can create disruptive reactive power- flows. Enforcing voltage inequalities one at a time, or several at a time with soft constraints, alleviates this difficulty. Criteria can be developed for selecting the inequality most likely to be binding out of a group of voltage inequality violations. Enforcing an inequality on a dispatchable source of active or reactive power forces the remaining dispatchable sources to make up for the deficiency or surplus. This action can force other dispatchable sources of the same kind to exceed their limits. These cascading effects be resolved by trial shopld iterations before proceeding. Violations of inequalities in this class shoul4 be enforced simultaneously because they generally tend to create more violations of the same kind.
Exact enforcement of all inequalities on each iteration is generally wasteful. This waste can be reduced by developing criteria for acceptance of an itferation that takes into account the number, kind and magnitude of violations as well as bus mismatches. These few examples of criteria are illustrative, but not exhaustive.

Network:

912 buses.

1637 branches.
1200 p.u.

Total MW generation:

(100 MVA base)

Starting Condition:
Objective Function:

unconverged network solution.

Minimum active power losses

Controls: 255 dispatchable reactive sources and 41 controllable transformers. Inequality Constraints: Upper and lower limits on dispatchable reactive sources, controllable transformer tap ratios and all bus voltages.

A version of Algorithm III for a decoupled formulation was used for the test. The strategy of inequality constraint enforcement was as follows: On the first iteration only the limits on generator voltages and transformer tap ratios were enforced. On the second iteration limits on reactive sources were also enforced. On the third and subsequent iterations all inequalities were enforced. All enforeements were with hard limits. Each Qv cycle was made completely feasible by trial iterations before the algorithm was advanced to the next cycle.

Results of Test Problem The history of iterations for the test problem is shown in TABLE I. Each iteration actually represents both a PO and a Qy cycle. However, the trial iterations and inequality constraint enforcements apply only to the Qv cycle.
TABLE I. Optimization Summary For A 912-Bus System.

Selective Monitoring

The forward solutions for trial iterations in Algorithm III can be performed efficiently because only a few columns of the factors are directly involved. The back solutions can be speeded up in a similar way by solving only for those quantities that need to be monitored. After each trial iteration it is only necessary to check two groups of inequalities; (1) those that were nonfeasible on the last trial; and (2) those that were marginally feasible but threatened with violation by the enforcement of other inequalities. The first group would normally consist of bus vQltages for which it would be necessary to check the effect of enforcing one on reducing the violations of the others. This group would rapidly diminish on each successive trial iteratiop. Since most of the effort of trial iterations is in the back solution, large gains can be made by selective monitoring.

(1)

Main Trial Iter. Itor.

No.of No. of

(2)

Active RMS of Max Bind'g Power Max P Max Q Projec'd Av or Limits Loss Mamtch Msmtch Gradient Atap

(3)

(4)

(5)*

(6)*

(7)*

(8)

1
2

11
8

142

11.448 0.286
10.325
10.271

6.136
0.820 0.344

.005554
.001167

.1824
.0689 .0462

143
131
133

0.418
0.303

3
4
5

23
2

.001417
.001367

10.302 0.097

0.126 0.028
0.036

.0186
.0193
.0225

TEST RESULTS
At the time of this writing the development and testing of algorithms is continuing. The intent of the investigation is to determine how different algorithms on different perform problems, not to produce a definitive solution method. All t;esting is being done on the VAX 11/780 computer. To give some evidence of how the approach works, the results of one experimental algQrithm on one test problem are shown. This example is representative of results being obtained with different problems and different experiment-al algorithms. It is expected that much better algorithms will be developed.
Test Problem

1
2

132

10.306 0.024
10.316 0.048 10.327 0.066

.000253 .000366
.000318

6
7

131
131

0.041
0.014

.0128
.0027

131
131

10.332 0.062 10.334 0.048

10.331
0.030 10.333 o.o14

.000094
.000047
.000046 .000043 .000034

9
10
11

0
0 0
0

0.004
0.004 0.004 0.002

.0006
.0002
.0001
.0001

131 131 131

12

10.334 0.002

* 100 MVA base.

Legend For Table I

Normally, multiple sources of active power can be dispatched. In this particular example, however, a single slack is used. The main attributes of the test problem are:

The network of the test problem is a portion of power -system of the Northeastern United States. The problem is reactive cpntrol optimization -only. the

Column 1 is the count of iterations. An iteration consists of one main iteration for the PO cycle and one main iteration plus a variable nqmber of trial

for the Qv cycle. The system is updated at the end of the main iteration of the PO cOycle. For the Qv cycle the system is not updated until the end of the trial iterations.

iterations

2873
Column 2 is the count of trial iterations needed to identity the currently binding inequalities for an acceptable iteration. If the main iteration itself is feasible, no trials are needed and the count is zero.
Column 3 is the number of binding inequalities the end of the iteration. at
prove that all OPF problems have unique solutions, strongly suggests that at least some of them do.
it

SUMMARY AND CONCLUSIONS The Newton approach is suitable for development;of practical OPF programs. This conclusion is supported by tests of Newton-based algorithms With prototype codes. has not been fully Although efficiency exploited in the prototype codes used in testing, it is possible to estimate the performances of efficient codes for the same operations. There is always some risk in drawing conclusions about a new approach that has not been extensively tested in production Use. It seems unlikely, however, that any difficulties can arise that would invalidate the main outline of the given approach. If an explicit Newton method cannot be made to solve a problem, any The existing quasi-Newton method would be no better. only known alternatives are weaker, more approximate methods that involve more computation. The key ideas of the given Newton approach are,
1.
An explicit Newton formulation'. Decoupling.

Column 4 is the value of the objective function at the start of the iteration.
Column 5 and 6 are the largest residuals of the PF equations at the start of the iteration.

Column 7 is the RMS of the gradient of L for the basic and superbasic variables at the beginning of the iteration. The RMS value is the square root of the sum of the squared gradients, divided by the number of the This nurdber is a basic and superbasic variables. measure of the steepness of the objective function. At the solution it should be zero except for roundoff.
iterative Column 8 is the maximum per unit correction in a voltage or transformer tap ratio. At the solution it should be zero except for roundoff.

DISCUSSION OF RESULTS
can

2.

The total computational effort for the solution be estimated from the total of main and trial each code main For a well-written iterations. iteration would require about 1.2 times as much effort The effort of trial as a Newtoh PF iteration. iterations is variable. With good sparsity techniques the cost of a trial should average about one tenth of a full iteration. This makes the total effort for the 18 Newton PF test problem equivalent; to about iterations.
Most of the trial iteratidns were performed at an Eleven trials were needed in the first early stage. iteration to find the currently binding taps and generator voltages. Eight trials were needed in the second iteration when reactive source limits were also enforced.

3.

Quadratic penalties constraints.

for

enforcing

inequality the

4.

inequalities.

Special

strategies

for

finding

binding

5. Special sparse matrix/vector techniques.

Some of the attributes of algorithms based on approach are, 1.
2. the

Superlinear convergence to the K-T conditions. Solution time PF s.

comparable

to

few

conventional

Twenty three trials were needed in the third iteration when load voltages were included in the enforced inequalities. After the third iteration only five more trials were needed to identify the binding The 131 binding inequality set1 for the solution. constraints consisted of 14 VAR limits and 117 bus Voltage limits.

3. 4.
5.

Solution time and matrix storage not adversely affected by number of controls or inequalities.
No need for user-supplied tuning interaction to obtain a solution.

factors

or

active/reactive only.

Adaptability to

all three subdivisions dispatch, active only,

of OPF; reactive

Iterations seven through twelvve were needed only solve the power flow equations while enforcing a fixed set of' binding inequalities. Convergence was slow near the end because of the weakhess of the decoupled PF formulation when the PF residuals become small. A change in solution scheme at this point would be advantageous.
to

A solution algorithm based on approach can be roughly divided 1.

the given into,

Newton

Problem definition: constraints, etc.

etc.

objective function,

controls,

It is also interesting to note that the objective reached its lowest; value on the third function iteration arid then gradually increased as the power flow mismatches were dimninished. More det'ailed monitoring and analysis of what is happening in the course of the solution is expected to provide insight for improving algorithms.

2.

Solution formulation:

decoupling,

approximations, constraints,

3.

Methods for enforcing inequality penalties, trial iterationsi etc.

4.
5.

Strategies for finding the binding inequalities. Sparsity techniques.

In developing an algorithm, choices must be made the several possibilities in each part. The

The possibility that OPF sOlutions may not be of considerable concern. Several OPF unique is problems have been solved with Widel-y different and with entirely different starting conditions algorithms. In all such tests identical solutions were obtained for the same problems. While this does not

among

2874
choices in any one part will be affected by the choices made in others. The number of combinations of choices is large and each application area may require As more knowledge and experience different choices. are gained, the best choices for each application should become evident. Practical, production-grade OPF programs can be developed from what is presently known.

G?kk jBkk
tknm
0km

Driving point admittance at bus k exclusive of contributions from any controllable transformers with their tappers at bus k. If branch (k,m) is a controllable transformer, tkm is its tap ratio. Otherwise, tkm=1.O.
If branch (k ,m) is a phase shifter, 0km is its shift angle. Otherwise, ,km = 0.

ACKNOWLEDGEMENTS
This research is sponsored by EPRI under contract I. RP 1 724-Phase ESCA Corporation is the prime contractor. Hermann Dommel, Philip Gill, Walter Murray, Michael Saunders, Larry Stanfel, William Tinney, and Margaret Wright are the consultants to the project. Dr. John Lamont is the EPRI project manager.

In the following derivatives,

km
=

Specific appreciation is expressed to Dr. John for Lamont initiating the project and for his to Walter continuing guidance; Murray, Margaret Wright, Philip Gill, and Michael Saunders for their the overall on expert advice application of Hermann Dommel for to Dr. optimization methods; sharing his broad perspective and experience on power and to Dr. Larry Stanfel for his system analysis; investigation of the relationship between the Newton method and Linear Programing in a real time operations environment.
The authors also wish to thank the project's industry advisers for the utility perspective they provided. APPENDIX

( k

am +40 - 7 ) km km

First Partial Derivatives of P

for J

and

'.

aPk aPk
a Ok k
ap?r

= -v Iv

m v km Y km sin4'km
t

a Orel
-

km
t

km
Y

sin;

km

apl, ao

k1a

-'VI'v

km

km

sin4'

km

Expressions For Matrix Elements

The matrix W is composed of submatrices J and H. Expressions for J are well known from the Newton PF; expressions for H are less familiar. More than fifty expressions are required for H, but many fewer are required for H and H".
Each element of H' or H" is of the form 3L/!yk,y1Y. L is just a combination of PF equations, each term of H' or H" is a sum of second partial derivatives of the PF equations wrt, Yk and ym. For any pair of variables Yk and Ym,- nonzero second partial derivatives exist only for those PF equations that contain Yk and ym. The number of nonzero terms in the sum for any element of H' and H" is small. The expressions for representative elements needed for W' and W' are developed in the following. Expressions for the first partial derivatives of the PF equations are needed for Ji and J". Expressions for second partial derivatives of PF evaluations are needed for H' and H".

First Partial Derivatives of

for J" and g

avk

Vm km km sin 2vk (Bkr tkm Bkm) k~~~~~~~~~~~~k

+
+

Since

ao
av

y k _v t ~ k km km sin;km

at

km!

= 2v

km

km

+ vk v Y sin; m km km

Representative Second Partial Derivatives of P

2 a p, k

for H

Power Flow Equations The polar form of the PF equations is shown below.
P

2ao2 ak

-v

zvm

km

km

cos4

km

= k
v

2, (C

~~~2G )+(v kk -Et km km k Y,vm

kk

2
a km m l;

km

km

cos4')
km

k Vm

tkm

cok Ykm

aQ k 'Q-

kkm

a m

k (
+
=

Z.2

Etkm Bkm)

(vkX

tkm

Ykm
/G

sin km

Representative Second Partial Derivatives of

for H"

(Ck km
Y
km

j B km )
G
km

Transfer admittance of branch km.

B

2+

km

7km

= arc tan B

km

km

Qk _2 2 =2( Bkk av2 k

a
2

tk

Bk)

Summations are over all branches (k ,m) connected to bus k.

avkavm

Qk

= t

Y sin pk km km

2875
Example
of

Hessian Elements

11.

Following are some of the expressions for elements of H' and H" for this example problem. Refer to the expression for L of the example in (2)
.

Stott and W. Alsac, B. "Sparsity-Oriented Compensation Modified Network Solutions", Ibid, No. 5, pp. 1050-1060, May 1983.
0.

F. Tinney, Methods for vol. PAS-102,

12.

a2L
aoj

P. E. Gill, W. Murray and M. H. Wright, Practical Optimization, Book, Academic Press, 1981.

I__

-I

2.
2

-P2
X

XQ2

2
2

13.

P. E. Gill, W. Murray, M. A. Saunders, M. H. Wright, Sparse Matrix Techniques in Optimization, Report SOL 82-71, Department of Operations Research, Stanford University, California, December

xP4

ao2

__4 4
ao2

1982.

Biography
Bruce Ashley received his BSEE from the University of Washington in 1980 and attended Purdue University under a Purdue Electric Power Committee Scholarship. He joined ESCA in December, 1981, and is a member of the IEEE PES, IAS, and the Computer Society. Brian Brewer received his BSEE degree in 1980 and MSEE degree in 1982 from the University of Washington, and is a member of IEEE PES and Tau Beta Pi. He joined ESCA in July 1982. Art received Hughes BSEE degree from the University of Manitoba in 1972 and MS and PhD degrees from the University of Texas at Arlington in 1976 and 1978. He is presently employed by the ESCA Corp. David I. Sun received BSEE and MSEE degrees from Resselaer Polytechnic Institute in 1974 and 1976, and PhD from the University of Texas at Arlington in 1980. He is presently employed by the ESCA Corp.
William F. Tinney received BS and MS degrees from Stanford University in 1948 and 1949. He worked for the Bonneville Power Administration from 1950 to 1979 and was head of the System Analysis Section at the time of his retirement. Most of his experience has been in power system computation. He is presently a consultant in this field.

.av4,av5

a2L

.2

a_

_4_

__a _ X)
_

Q4

________-

a v4av5

xQ5

av4av5

Expressions for other elements of H' and H" follow directly from these examples.
REFERENCES
1.
Alsac and J.L. Stott, 0. B. Marinho, "The Optimal Power Flow Problem", in Electric Power Problems; The Mathematical Challenge, Book, SIAM, pp 327-351, Philadelphia, 1980.

2.

A. M. M. Sasson and H. "Some Merril, Applications of Optimization Techniques to Power System Problems ", IEEE Proceedings, vol. 62, pp. 959-972, July 1974.
H. A Happ, "Optimal Power Dispatch Comprehensive Survey", IEEE Transactions on Power Apparatus and Systems, vol. PAS-96, pp. 841-854, May/June 1977.

3.

H.

4.

B. of

Stott, J. L. Marinho and 0. Alsac, "Review Linear Programming Applied to Power System Rescheduling", IEEE PICA Conference Proceedings, Cleveland, pp. 142-154, May 1979.
H. W. Dommel and W. F. Tinney, "Optimal Flow Solutions". IEEE Transactions on and Apparatus Systems, vol. PAS-87 , 1 866-1876, October 1968.
R.
Power

5.

Power pp.

6.

C. Burchett, H. H. Happ and K. A. Wirgau, Scale Optimal Power Flow", Ibid, vol. "Large PAS-10l, No. 10, pp. 3722-3732, October 1982.

7.

A. M. Viloria and F. Sasson, F. Aboytes, "Optimal Load Flow Solution Using the Hessian Ibid, vol. Matrix", 31-41, PAS-92, pp. January/February, 1973. F. L. Alvarado and M. K. Enns, "Blocked Sparse Matrices in Electric Power Systems", Paper A 76 362-4, IEEE PES Summer Meeting, Portland, July 1976.

8.

9.

Flows", IEEE Systems, vol.

B.

Stott

and

Alsac, "Fast Decoupled Load 0. Transactions on Power Apparatus and PAS-93, pp. 859-869, May/June 1974.

10.

E. C. Housos and G. D. Irisarri, "A Sparse Variable Metric Optimization Method Applied to the Solution of Power Syst;em Problems", Ibid, vol. PAS-101, no. 1, pp. 195-202, January 1982.

2876
Discussion S. K. Chang and J. Lequarre (Systems Control, Inc., Palo Alto, CA): This paper has presented a refreshing new approach to the solution of optimal power flow problem. At Systems Control, we have developed a Loss Minimization Program (Reactive Control Optimization) employing a similar Newton-based approached and the test results indicate that this method is very promising for practical applications. We would therefore like to share our experience by offering the following comments: 1. Although not stressed in the paper, we feel that a good initial estimate of z is essential for improving the robustness of the algorithm. This is due to the fact that Newton's method is ideal only in terms of its local convergence properties, but not in terms of global convergence. An initial start with flat voltage and zero X will generally cause difficult or erratic convergence. A typical scheme would be as follows: Perform one load flow iteration to obtain initial state variables corresponding to a reassigned set of control variables. Then compute the initial X from the relation [JvJTh = v, where [Jv] is the nonsingular power flow Jacobian matrix, and gv is the gradient vector of the objective function with respect to the state variables. For the Loss Minimization problem, the initial assignment of maximum voltages to the generator control variables was found experimentally to give best convergence characteristics. 2. The partition in the paper of the W matrix into an external submatrix and a main submatrix suggests the possibility of illconditioning if a controllable transformer has negligible conductance. For example, the diagonal term (At, At) corresponding to a controllable transformer in the Loss Minimization problem can be written as (At, At) - term = 2Vk2 [g(l - XpK) + bXQkl = 2Vk2 b.XQk g = o where Vk is the voltage magnitude at the tap-side bus k g,b the transformer conductance and susceptance, respectively the Lagrange multiplier for real and reactive power XPk, XQk flow equations at bus k, respectively. If bus k is of PV-type and XQk is zero, the term (At, At) is zero and can cause numerical difficulty if the partition strategy is applied. A modified ordering or processing scheme, e.g., placing the tap variable in the main submatrix, is therefore required to circumvent this problem. 3. In spite of its heuristic nature, the strategy of enforcing/releasing multiple constraints at a time seems to be essential for efficient identification of the binding inequality set in large scale OPF problems. To apply this strategy, however, the possible phenomenon of "zigzagging" should be properly guarded against so that a constraint is not repeatedly enforced and released during the trial iterations. Although finding a method for identifying the binding inequalities still is a major challenge, it seems that a strategy that utilizes zigzagging-guarded multiple constraint enforcement/release coupled with deterministic single constraint enfordement or release would be a practical choice. The authors' comments concerning this point will be appreciated. Finally, we would like to commend the authors on a well-written and important paper.
Manuscript received February 13, 1984 B. K. Johnson (Power Technologies, Inc., Schenectady, NY): The authors have presented a very interesting paper on methods of optimal power flow solutions. Having developed a technique similar to their coupled solution several years ago,[1J I read the paper with particular interest. A point to note is that if the objective function, F is total cost of generation then the Lagrange multipliers Xpk amd Xpa equal the incremental cost associated with their particular constraint equation, i.e., the incremental cost of power or vars at system buses where power or vars are constrained. This suggests that this information might be useful for determining the charges which should be assessed for interchanges between utilities or the most economic location for capacitor banks. Have the authors pursued such an extension of their method?

[1]

REFERENCE V. D. Albertson, B. K. Johnson, and W. Scott Meyer, "Exact Economic Dispatch With Quadratic Convergence", IEEE Conference Paper No. 68 CP 671-PWR presented at the IEEE Summer Meeting, June 23-28, 1968. Manuscript received February 21, 1984

Walter L. Snyder, Jr. (Leeds and Northrup Co., North Wales, PA): The authors are to be complimented on a well developed, practical, and workable approach to the optimal power flow problem. This discussion presents a more compact and generalized mathematical statement of the problem objective, and explores the treatment of active generation for both binding and nonbinding constraints. The authors' approach minimizes the following Lagrangian type function: L(z)= L(y,X) = [C * P] - Ri * dP] + (2) * [dx * [S] * dx] (equation di) where: [and] denote row and column vectors, respectively The d "operator" denotes a finite change, not differentiation p] = P(y)], and includes both real and reactive power (or either one in a decoupled formulation) dp] = P(y)] - P(sch)] dx] = [A] * Y] - x(sch)] Eq. (di) should be identical to the authors' formulation and lead to the same solution equation, but expresses the problem in matrix form, explicitly including the quadratic penalty functions. The advantage of the form given in (di) is that it allows one to more clearly review and evaluate the complete objective. It should be pointed out that the lambda terms of (di) contain dP rather than P as in the authors' paper. Could the authors please verify that P in their lambda term is actually a mismatch, dP, while P in their cost term represents the actual injection and not a mismatch. If such is not correct, would the authors please clarify these expressions. The discusser views C and X as playing the same role, where both are never non-zero at the same time for a given element of P. If C is zero, X is a variable which is solved such that P is constrained to a scheduled value by enforcement of the Kuhn-Tucker conditions. A non-zero C, hoever, is viewed as a prespecified, fixed incremental cost thereby taking the place of lambda. For non-zero C, P is no longer constrained to a scheduled value since C is not a variable, rather the C*P term is minimized in conjunction with the rest of the cost terms. For consistency, a scheduled value of 0 could be assumed in the C*P terms, although this scheduled value is never enforced. The approach would tend to reduce the dispatched elements of P to zero were it not for maintenance of a power balance via the J and H sensitivities which relate dispatched P to scheduled P. The authors' comments on the preceding paragraph would be appreciated. In particular, the treatment of real power C and lambda coefficients with respect to binding and non-binding constraints is not clearly explained in the paper. Are the real power lambdas held at zero while P is dispatched within limits, and are the associated costs set to zero when P is at a limit? If C is left non-zero when P is binding, does the non-zero lambda effectively override the C*P term? In Eq. (4) of the paper, the inclusion of real power but not reactive power lambda terms for buses and 3 lends to the confusion, since both real and reactive power are defined as dispatchable on the two buses. Similar confusion is evident in Eqs. (6) and (9) of the paper where reactive lambda terms are 'dummied out' (set to zero) for buses 1 and 3 while real power lambda terms are not. Again, this suggests that real power on buses 1 and 3 is scheduled while the paper states that it is dispatchable. It should be pointed out that the zero diagonal terms in (6) and (9) will present no problem since fill in will occur when the matrix is actually factored due to the jacobian terms of the associated diagonal matrix packet. It is assumed that the only purposes of the arbitrarily large diagonal terms is to ensure a zero correction to lambda and v on execution of the 'divide by diagonal' step of the forward-backward substitution process. Finally, the last term of Eq. (di) at the beginning of this discussion is seen as a more general expression of the quadratic penalty terms used to enforce limits on the variables themselves as well as on linear functions of the variables. [A] is simply an incidence type matrix (not to be confused with the set, A, of binding constraints), and is a trivial identity for limits on the variables themselves. Was this method considered by the authors for "beyond the scope of this paper" soft enforcement of limits on power sources? In Eqs. (16a) and (16b) of the paper, should a scheduled power angle term be included as was done in Eq. (15)? Overall, the presentation of this paper was still excellent and was extremely motivating to the discusser, injecting new life into the optimal power flow problem.
Manuscript received February 23, 1984

2877
B. Stott, 0. Alsac and A. Bose (Electrical & Computer Engineering, Arizona State Univ., Tempe, AZ): For over twenty years, researchers have had limited success in developing efficient, reliable solution methods for large "classical" OPF problems with nonseparable objective functions. These problems are encountered mainly when scheduling reactive power for the minimization of objectives such as production cost or losses. The present paper introduces an approach with real potential for a breakthrough. In doing so it reverses a recent trend towards the use of sophisticated modern general-purpose constrained-optimization methods and/or codes, which appear to solve OPF problems correctly but relatively very slowly. High speed is aimed for by reverting to optimization first principles and focussing the development effort on structure-suited implementation, resulting in an approach with good convergence power and excellent network sparsity preservation. The attraction of the approach centers on the facility with which the equality-constrained OPF problem can be solved. The Newton solution of its Lagrangian necessary-condition equations is conceptually simple, performs second order minimization with no tuning, and is amenable to efficient exploitation of problem structure. Unfortunately, this approach has no accompanying powerful well-established means for handling inequalities. Therefore heuristic schemes, tailored to particular OPF formulations, need to be developed for switching limit constraints into and out of the binding set via equation addition/removal or quadratic penalties. The same general approach has been tried in small-scale research a number of times before (see especially [A]). Among the major improvements introduced by the present work is a superior way of organizing the necessary-condition equations for efficient decoupled sparse solution. It is shown that quadratic penalties can be used effectively for statevariable limits, as well as for functional inequalities. Improvements in efficiency through the use of advanced sparse matrix techniques are emphasized. The most critical factor for success in the approach is to find schemes that determine the binding limit set reliably and rapidly, on a wide range of power systems, constraint types and OPF formulations. This is precisely the major problem encountered in many previous OPF methods and, as far as is known, never yet satisfactorily resolved. The present paper emphasizes this as the key problem that must be overcome by further development. It gives very rough outlines of several candidate trialiteration schemes. As the paper itself states, it is introducing a methodology as much as a developed method. Many questions of detail remain to be answered. A few of these questions are as follows. It is stated that enhancement of positive-definiteness of the Lagrangian is unnecessary, implying the absence of convergence problems. Has sufficient experience with different systems and situations been accumulated to verify this? Is decoupling sensitive to branch R/X ratios? What is the effect of decoupling when voltage magnitude limits need to be enforced by active power scheduling. What is the infeasibility behavior of the approach, and what infeasibility strategies are most appropriate with it? Penalties are advocated for functional inequalities, but other than for simple line limits, they can alter the structure of matrices W' and W"'. How are important limits such as interchange, reserve and contingency constraints to be handled? How is solution oscillation avoided using piecewise cost functions as in Eq. (1)? Can controlled hvdc links, with variable converter-transformer taps, conveniently be included? Can the approach conveniently and efficiently accommodate all modeling and local-control features of a conventional power-flow solution, when needed? In conclusion, now that the approach's potential has been clearly exposed, its general effectiveness and applicability for industrial use need to be established. There are many practical requirements for OPF solutions not covered by the basic classical problem formulation. Extensive further development and testing with different power systems, objectives, constraints, etc., will be necessary. We look forward to further results from the project. W. W. Lemmon and W. R. Barcelo (Middle South Services, Inc., New Orleans, LA): The authors' contribution to the field of Newtonian OPF methods is appreciated. As the authors point out, many Newtonian approaches are possible, but a simplification of the true bordered Hessian matrix is essential since it has 12 non-zero entries for each line between (P, Q) type buses. In 1977 [1] we presented a Newtonian-based OPF method for the real-time enforcement of security constraints using the bordered Hessian matrix. Our method uses the highly sparse "diflex" approximation which assumes that the (P,Q) busses have passive loads. In this method, a compensation term is added to the Hessian to account for voltage deviations during solution iterations. This approach required only 4 matrix entries per transmission line instead of the 12 entries required by a pure Newtonian method. We have used the method in a study mode to optimally dispatch active power while maintaining security constraints. The logic is available for both active and reactive dispatching [2] but the reactive part has not yet been implemented. Our constraint enforcing technique is basically "Algorithm II" for active source limits and security constraints and "Algorithm III" (modified) for reactive sources and voltage constraints. Algorithm II is significant in that a proof of convergence exists. [11. The Objective Function is production cost and refactorization is both selective and compensational. P-Q decoupling is performed only on an appropriate subset of busses; also, the PO and QV parts are subjected to simultaneous "trial iterations" (which we called "subiterations"). The authors' use of relaxation on the PO and QV subsystems is analogous to the decoupled PF technique and in our opinion has high potential. However, when security constraints (special function constraints) are being held, the decoupling assumption can fail for the Hessian submatrix H. This "recoupling" is independent of whether the constraints are held by X's or by penalty functions. In the presence of recoupling, limiting trial iterations to individual PO or QV cycles, as the authors propose, may lead to inappropriate selection of the binding inequality constraints due to each subsystem's lack of knowledge of the other's constraint enforcement capabilitids. This may result in an unduly large number of trial iterations or even divergence. We shall describe the reason for this recoupling phenomenon and suggest three possible solutions. Recoupling occurs when a constraint is affected strongly by both voltage and angles-we call this intrinsic recoupling. For example, when trying to impose a thermal constraint such as MVA or current magnitude on a line that is transmitting appreciable reactive power, a strong (0, V) coupling term is added to the Hessian. The authors' example of constraining real power flow does not show this effect. Recoupling also can occur whenever constraints are being held tightly enough to cause the resultant Xp profile to become distorted. We call this X - differential recoupling. In order for 0, V) terms in the Hessian to be negligible, it is necessary that the (Xp difference between connected busses be reasonably small-which is assured in the absence of security constraints by low line losses. While holding certain security constraints, we have observed Xp's between connected busses that are orders of magnitude apart! This type of recoupling occurs because frequently voltages are helpful in controlling constraints that depend explicitly only on angles: for example, raising voltages (within allowable limits) has the ultimate effect of decreasing angular differences. Because of the great potential of the decoupling approach, we have attempted to develop solutions to the recoupling problem. Three of the more promising solutions which may be applicable to the authors' technique are as follows: Solution 1: Give the reactive subsystem "busy work". If all busses have sufficiently large non-relaxable quadratic penalties on their voltages, the coupling terms are made negligible because voltage can no longer be used to help maintain security constraints. Unfortunately, this solution uses reactive power only for voltage profile control, which violates the intent of the OPF solution. Solution 2: Partial decoupling without relaxation. This is the solution which we have provisionally chosen for our optimizer. The bordered Hessian W is decoupled except for the recoupling terms and iterations are performed simultaneously on the PO and QV subsystems. Of course, when no recoupling is present, it would be desirable to relax on the two subsystems in the normal manner since simultaneous iteration takes about twice as long for given accuracy criteria. Solution 3: Multiple block relaxation. Normally, the recoupling terms (if any) are not very numerous. Thus it should be possible to decompose the bordered Hessian, into, for example, three approximately decoupled submatrices W', W", and W"', where W"' spans a relatively small set ofPO and QV variables that are strongly coupled. Each itera-

REFERENCE

[A] H. Duran,

flow problem," Proc 8th PICA Conf., 1973.

1984

"A Simplex-like method for solving the optimum power pp. 162-167, Minneapolis

Manuscript received February 24,

2878

tion would then comprise a relaxation on three subsystems: a large PO, QV pair and a small set of connected mixed variables. We would appreciate any comments the authors may have concerning the recoupling problem and our proposed solutions. Also, would the authors please comment on the accuracy of the matrix factorization in their approach when large weighting factors are required on the penalty functions to hold the inequality constraints and how these weighting factors are selected. We commend the authors for their organization of the material on the constrained OPF problem and their presentation of the variety of algorithms possible for determining the binding inequality set. We look forward to further papers on this subject by the authors and are especially interested in the convergence properties of the decoupled approach.
REFERENCES
1. W. R. Barcelo, W. W. Lemmon, and H. R. Koen, "Optimization of the Real-Time Dispatch with Constraints for Secure Operation of Bulk Power Systems." IEEE Trans. PowerApp. Syst., vol. PAS-96, No. 3, p. 741-757, May/June 1977. 2. W. R. Barcelo, W. W. Lemmon, and J. D. Stenquist, The Middle South Tulane Security Modelfor a Bulk Power Management System, Middle South Services Proprietary Documentation, 1975.
Manuscript received February 25, 1984

other investigators in having an EPRI research contract that enabled us to carry out the costly programming and testing. The importance of techniques cannot be overemphasized. The method and techniques are inseparable, but the equations and matrices given in the paper do not explain the techniques. For highest efficiency, sparsity techniques must be adapted to each specific application; general purpose software cannot do this. S. K. Chang and J. Lequarre: We are pleased that the discussers have developed a similar Newton approach, and that their experience with it confirms our claims about its promise. The suggested idea of performing one conventional power flow iteration with nominal control settings before optimizing seems to work well as a starting procedure. A converged power flow solution or a feasible state is unnecessary. As pointed out, a controllable transformer with negligible resistance will introduce a near zero diagonal term that cannot be processed in the given matrix partitioning without numerical breakdown. Putting such transformers in the main submatrix corrects the numerical difficulty in most cases, but may degrade efficiency. If node k in the discussers' equation is radial as well as PV-type, a local degeneracy will occur with a no-resistance transformer even if it is in the main submatrix. All transformers have sufficient resistance to prevent numerical breakdowns. If the resistance is missing in the data, a minimum value can be automatically defaulted. Resistance should always be modeled accurately for loss minimization. The recommended partitioning is highly advantageous, and it can be retained for all transformers. We concur with the cautions about zigzagging in strategies that enforce/release multiple constraints simultaneously, although if a good active set management strategy is used, zigzagging should occur only rarely. When it does occur, a simple ad hoc scheme would be to keep count of the individual zigzags and to reduce the number of constraints being simultaneously changed should a specified count limit be exceeded. In a particularly difficult situation, this scheme could eventually reduce to changing one constraint at a time for the given iteration, although this, also, should rarely occur.
B. K. Johnson: The cited paper co-authored by the discusser was probably the earliest application of the Newton approach to the OPF problems. The Lagrange multipliers of the power flow equations could, as suggested, be used in various ways as a basis for transaction-costing which takes into account incremental transmission as well as production costs. The OPF is also a key module in important applications that have been impractical until now because of the lack of an efficient OPF solution; capacitor allocation is one such application. The Newton approach should be highly efficient in this OPF subproblem.

A. Monticelli and F. F. Wu (Dept. of EECS, Univ. of California): The authors have developed a fast, accurate, and robust approach to optimal power flow. The method has potential for many other applications. For example, it can be applied to the solution of state estimation problem with equality constraints [A], where at each iteration a matrix equation of the form of Eq. (7) is solved. The Newton method takes a large step at each iteration. In the OPF by Newton's approach developed in this paper, the direction of movement at each iteration depends on the selection of binding constraints. If one takes the Newton's direction based on the binding constraints at the current solution point, due to constraint violation, the stepsize may be very small (Algorithm II). The trial iteration is a scheme to determine which Newton's direction is better. It seems to us the scheme used in the trial iteration is crucial to the success of the method. We would appreciate if the authors would comment on: (1) the comparison between Newton-OPF/Algorithm II with the conventional gradient approach, (ii) to elaborate on the method they used for trial iteration, and (iii) to comment on whether the gradient approach can benefit from the use of trial

iteration.

REFERENCE

[A] F. A. Aschmoneit, N. M. Peterson, and E. C. Adraian, "State

Manuscript received March 2, 1984

Estimation with Equality Constraints," Proc. 1977, PICA Conference, Cleveland, OH, pp. 427-430.

W. L Snyder, Jr.: A more detailed exposition of the basic formulation than appears in the paper is given here in an attempt to answer the questions raised in the discussion. A new variable for dispatchable active power is introduced and a quadratic cost function is used instead of piecewise linear costs. The following definitions of symbols are for a bus with dispatchable active power:

W. F. Tinney, D. I. Sun, B. Brewer, and A. Hughes: We are grateful for the excellent discussions. This closure first addresses issues common to several of the discussions, followed by replies to other questions listed under the discussers' names. We apologize for omissions in the bibliography, as some papers describing similar approaches have gone unreferenced. We've found that it is difficult to limit such a list once it is started. Some of the unreferenced papers are listed in the discussions, but more exist. In retrospect, it is apparent that several forms of the Newton approach have been developed previously by others; our contribution lies in developing the techniques for making it practical. We had an advantage over

PGk Pk fk Ck F

Active power generation (variable) Active power mismatch (function) = PGk - EPkm Cost of generation (function) = V2ak (PGk)2 + bk(PGk) Incremental cost (function) = ak(PGk) + bk Objective function = Efk

Using these definitions the Lagrangian for the five bus example becomes, L = fl + f3 - XPIP, - XP2P2 - XP3P3 - XP4P4 - XP5P5 - X Q2
-

xq4Q4

Xq5Qs

(.1)

In (C. 1) the variable PG1 occurs in functions f, and P1, and the variable PG3 occurs in functions f3 and P3. Assume that axes for the variables

2879

PG, and PG3 are added to matrix W' for the decoupled version. The conditions for bus 3 within W' are shown in (C.2).

cient than processing individual elements. It is probably an unnecessary precaution, but we are not certain. The schedules angle terms were omitted in Eq. (16). The equation should be,

2/,PG2
h
-1
j

-1
j

APG3

PG3
a.s,e3

(C.2)

a =k si (ek O en

ekm)

(16a)

Amp hA3

8p3
aai

(e= i(k

-e

Em

km)

(16b)

where

Ap3
2

C3

"p3

These discussions were very helpful in pointing out ambiguities in the problem formulation. We hope this response has answered the questions satisfactorily.
vantages of the approach and raises important questions about it. The quest for improvement in schemes for determining the binding constraints

and

B. Stott,

0.

Alsac and A. Bose: This discussion summarizes the ad-

The axes for dispatchable active power, PG, and PG3, do not appear explicitly. They are reduced into the rest of the matrix. Eliminating the axis for PG3 gives,

G *P I8
(9

]C
(

ax

CPG3)(

6PC )

To initialize the solution, the incremental costs, C. and C3, are evaluated for the initial values of PG, and PG3, and XP, and Xp3 are set to the values of Cl and C3. Note that for linear generation cost model (as in the ex-

ample of the paper), the tern:

(a

F/<pc2) is

replaced by

large

number, co, on the diagonal and the RHS of XPk axis. The result of these steps is symbolized in Eq. (1 la) of the paper. Their effect is to force a predetermined change in AXp, and AXp3 in the solution, in the same manner as explained for penalties on variables. Solution of the matrix equation ensures coordinated changes in all of the variables. On each iteration, PG,, PG3, XP, and Xp3 are updated, and C, and C, are re-evaluated using the updated values of PG, and PG3. At the K-T conditions Xp, = C, and Xp3 = C3. If a bus with dispatchable active power reaches a limit, it is converted to a P-scheduled bus. Since there is no cost for dispatchable reactive power in the example, the XPk axes of these buses are simply dummied out. If a reactive source reaches a limit, it is converted to a Q-scheduled bus (the Xqk axis is re-inserted). The interpretation of the formulation given in the discussion is also essentially correct for a Newton OPF approach and it may be viewed as PF extensions. The seemingly more complicated formulation given here is used mainly in order to express the functions of dispatchable power costs in terms of the variables PGk instead of the functions Pk. IF Pk were used instead of PGk in fk, it would create many additional terms in the matrix. The formulation also has important advantages for some special OPF applications. As pointed out, the zero diagonals in each diagonal block will fill in before they become pivots in the factorization. However, as an additional safeguard against numerical problems in the test program each block submatrix was processed as a unit, and the diagonal blocks were explicitly inverted. Processing block submatrices is somewhat less effi-

could continue indefinitely, but the existing ad hoc schemes are adequate for practical purposes. The lack of a definite scheme for automatic adjustment of controls in conventional power flow programs has not prevented their practical use. Finding the binding constraints is a challenge in all optimization methods. It is not unique to the OPF or the Newton approach. Although the Newton formulation seems to be strongly positive, definite safeguards to ensure positive definiteness should be provided. Augmenting the Lagrangian with additional quadratic penalties is a possible scheme. Decoupling in the Newton OPF, similar to that in the fast decoupled power flow, is degraded by high R/X ratios. The example problem in the paper had many lines with high R/X ratios. Reducing the ratio in one critical line reduced the number of main iterations for the example solution from 12 to 8. Presumably, the remedies for high ratios used in the decoupled power flow will also work in the OPF. After the first few main iterations in which the binding constraints are approximately determined, the bordered Hessian submatrices W' and W" become nearly constant and no longer have to be recomputed and refactorized. If changes are necessary they can be made by highly efficient partial refactorizations. This considerably reduces the cost of additional iterations needed to reduce the power flow mismatches resulting from high R/X ratios. It also makes solutions of all problems faster. Some complicated functional constraints, such as area interchange and MW reserve, are not necessarily best enforced by explicitly including quadratic penalties in the bordered hessian. The complications arise from significantly disturbing the sparsity structures of the matrices, and/or from numerical considerations. The alternatives include applying the Schur complement technique, or explicitly introducing additional equations to the matrices. Schur complement is attractive when the number of equations is small; embedding the equations directly is efficient, but the programming is further complicated by the nonuniform block
structures.

Various OPF applications will require all of the models now used in conventional power flow programs except certain local controls that would not make sense in a global optimization. It seems reasonable that all models for the Newton power flow could also be implemented in the Newton OPE but they have not been individually evaluated. Multiterminal hvdc links could be difficult. Another obstacle is how to optimize sensibly in a network with equivalents. Piecewise cost functions appeared in the paper only for illustration; quadratic costs were used in testing. If costs are supplied in piecewise form, it is best to automatically fit them with analytic functions from which first and second derivatives can be computed. Handling of explicit piecewise linear costs the solution process requires needs logic to prevent oscillations at their breakpoints. The very similar approach to OPF by Duran was overlooked. With good sparsity techniques it could have worked much the same as our approach. He apparently tested it only on small problems.

2880

W. W. Lemmon and W. R. Barcelo: The discussers' successful method for active power dispatch by the Newton-based diflex method was helpful to us and it should have been listed as a reference. We still do not fully understand -the diflex approach, and in our investigation we could not see how it could be adapted to enforce voltage constraints efficiently. A comparison between the diflex OPF and our decoupled version for the active/reactive problem should be interesting. The diflex OPF for active power seems to be the first Newton approach used on large practical problems. Although the exact Newton formulation without decoupling results in a large bordered Hessian matrix, it is not intractable even for the largest power networks. Whether to use the exact formulation or some decoupled approximation would be decided on the basis of computational tradeoffs. Decoupled formulations usually are more efficient, but the exact formulation need not be ruled out. The need for "recoupling" could be a serious impediment to the success of decoupled approximations for some applications. We greatly appreciate the discussers' sharing of their experience with it and their suggestions for solving it. Either of their suggested solutions 2 or 3 should accomplish recoupling with a minimum sacrifice of the advantages of decoupling. Efficient programming is the main challenge. The fact that the constraints requiring recoupling are generally few in number is a big help. We are inclined towards Solution 3 because we can visualize an efficient way to program it. When recoupling is not necessary, the program would be unaffected. To enforce simple-bound limits on a non-degenerate set of variables, quadratic penalties are effective. Numerical accuracy is unaffected by the number of binding limits or by larger than necessary penalty weighting factors. It is best, however, to have the weighting factor no larger than necessary since this makes efficient removal of a penalty by compensation (or factor update) a more accurate process. A value of 1,000 times the Hessian diagonal may be a good compromise. The exact value is not critical.

In enforcing functional limits with quadratic penalties, which creates a large off-diagonal as well as diagonal terms, it is possible to experience numerical difficulties. If this occurs, penalty weightings could be reduced in magnitude and shifted to produce the desired degree of enforcement. For this reason, penalties do not have to be extremely large to be effective. Further experience will contribute to the knowledge about how penalties can and cannot be used.

A. Monticelli and F. F. Wu: Some non-OPF problems have matrix formulations identical to those of the bordered Hessian but they are not solved in the manner described in the paper. State estimation is one example. The suggestion that there might be advantages in solving these problems in the same factorized block form as the OPF seems well worth investigating. This form of solution would remove most restrictions on the form of the upper left hand submatrix. A definitive strategy for trial iterations has not been worked out. Performance of a strategy depends on the number of kinds of constraints enforced/released and the number and kind of quantities monitored on each trial. Programming tecniques are as important as the selection strategy and each affects the other. The strategy resulting in the fewest trials is not necessarily the best since the speed of a trial must also be considered. The weakness of reduced gradient methods for OPF is such that they tend to hang up before reaching the exact K-T conditions. The hang-up point is not unique and it depends on starting conditions. The difficulty seems to be that the reduced gradient is not in the direction of the minimum; i.e., it does not provide a good move direction. Trial iterations would not help because a gradient-based move constantly encounters constraints that do not lie between the current point and the minimum
Manuscript received April 30, 1984