Optimal power flows

J Carpentier
Electricit& de France, 3 Rue de Messine, 75008 Paris, France

The state of the art o f optimal power flows at the end o f

19 78 is reviewed. The different aspects of the problem
statement are established in the most complete way possible;
then methods for the solution of the static optimization are
presented and filed, with a brief description of the most
typical ones, trying to show the respective merits o f each
family, and with a discussion o f points o f special interest.
Finally, dynamic aspects with applications to online
dispatch are investigated.

I. I n t r o d u c t i o n
Before tile introduction of optimal load flows, so-called
'economic dispatch' was used to determine the best way to
share the real load between several generating thermal units
having a total capacity greater than the generation required.
Everything was as simple as in paradise : losses were taken
into account through B coefficients, control orders were sent
directly to the units and nothing happened but savings.
The trouble started around 1961 when the use of networks
close to their limits led to a fear of line overloadings; security
constraints had to be introduced and optimal power flows
were rapidly born. Instead of a 'compact' model consisting
of one equation with real powers only, it appeared necessary
to consider all the variables defining the state of the system
and to solve the economic dispatch and the load flow
problems at the same time. An optimal power flow may
thus be dethaed as the determination of the complete state
of a power system corresponding to the best operation
within security constraints. Best operation usually means
least fuel cost: security may range from the generation
feasibility up to very sophisticated constraints, so that the
optimization problem may become huge.
Since 1961, a number of optimal power flow computation
methods have been proposed. In this paper, after presenting
various possible statements of the problem, our aim is to
point out the most typical methods, as they appear today
with a thought towards their possible usefulness, and bearing
in mind that omissions are always possible. Except for a few
papers, the author compelled himself to read (or re-read)
each paper mentioned in the references; so, please, forgive
him if the list is too limited. Complements to references
may be found in the excellent survey on optimal power
dispatch carried out by Happ 6s a few years ago.
The author found writing this review very exciting for two
reasons: first, looking at 15 years' work corroborated his
opinion that sometimes there exist barriers between
countries -- researchers from one country working on
subjects solved and published elsewhere five or ten years
before; secondly, the revolution in computer hardware
means that cheap, fast and powerful computers will soon

Vol 1 No 1 April 1979

appear on the market and these will certainly make a

contribution in reducing the gap existing between theory
and real time applications for optimal power flows. The
author hopes that this paper may help in putting researchers
in closer contact and in putting theory closer to real time
applications.

II. Statement of the problem

II. 1 Preliminary assumptions
Before stating the problem, we assume that
thermal generating units were committed (nuclear units
being included among thermal units)
hydroelectric plant real generations were fixed
the transmission network was defined, at least provisionally.
11.2 Physical control variables
In order to find the best operation within the constraints
imposed, the following variables are employed:
(1) real generated powers of the thermal generating
units
(2) voltages (or reactive generated powers) of the
thermal and hydroelectric units and of the synchronous
compensators
(3) variable transformer tap ratios for voltage magnitude and for phase shifts if any
(4) other variable reactive power sources such as
capacitors and reactances
(5) in special conditions, hydropower, emergency
start-up, load shedding and even network structure changes
As a rule, we shall call these variables the 'control variables'
of the problem. In general, only variables (1), (2) and (3)
are considered. Let us call:
'complete' or 'real-reactive' optimal power flow a
problem with the variables (1), (2) and (3) and possibly
(4) and (5)
'real' optimal power flow, the suboptimization problem
with the real variables (1) only, the variables (2), (3) and
(4) being fixed and (5) possible,
'reactive' or 'reactive only' optimal power flow, the
suboptimization problem with the reactive variables
(2) and (3) [(4) possible], the variables (1) being settled
(except at the slack bus to meet the real power balance).

0142-0615/79/010003-13 $2.00 1979 IPC Business Press

11.3 Objective function

The 'best operation' consists, in most cases, of minimizing
the real production cost F corresponding to the fuel cost
directly due to power generation, excluding no load,
start-up and shut-down costs. As a rule, F is a polynomial
in output power, P, usually of degree less than or equal to 3
and often represented by a piecewise linear function.
In special cases, an environmental cost function replaces
the fuel cost function or is mixed with it; load curtailment
may also be minimized in some emergency cases.
The considerations above are valuable for the complete and
real problems. In the reactive only problem, the losses are
usually minimized, but more sophisticated objective functions may sometimes be considered.
11.4 Security
Security requirements are such that the operation is possible
for generating units, transmission lines, transformers and
customers - first when the system remains intact, and
secondly under contingency conditions, i.e. when one or
several elements of the system suddenly trip. According to
Dy Liacco 2s,38 the words feasible and secure have been
used in the past for the two cases just mentioned. Since
the English vocabulary appears to have changed in recent
years, here we shall use the usual French designations:
when operation is possible for the intact system, we shall
say 'n security' is met
when operation is possible for the system under contingency, with k elements tripped (k ~> 1), we shall say
'n-k security' is met. Contingencies are usually limited
to the loss of 1 element, which corresponds to the case
of the ' n - I security'
These labels have the merit of being short and accurate.
n - 1 security (and sometimes n - k ) must be considered
because the sudden trip of an element may always occur
before any corrective action. Security involves meeting the
inequalities linking the control variables of the problems.
A list of the most common security inequalities encountere?, in optimal power flows is given below (n or n - k in
brackets indicate the corresponding degree of security):
limits on the real power output P of a generator (n)
limits on the reactive power output Q of a generator (n)

limits on a function of these two quantities (e.g.

p2 + O2)(n)

limits on transformer tap ratios (n)

limits on bus voltages (n)
limits on currents, apparent power, real power or the
phase difference through a line or a transformer (n and
n - 1, sometimes n - 2 )
limits on the generator voltages just after contingency,
especially when a generator has tripped ( n - l )
The major objective of security requirements is to avoid
a general failure in the system, the cost of which is considerable: e.g. in France, the general failure of 19 December

1978 lasted only about 3 h but caused a loss of production

for the country estimated to be at least the equivalent of
50 years' savings through economic dispatch.
This point leads us to two important questions concerning
the problem statement: discussion of the n - 1 security and
the overconstrained cases. A n-1 (or n-k) security constraint must be met only if dropping it would cause a
general breakdown or at least a sufficient amount of nondelivered power. The 'overconstrained case' happens when,
due to constraints, an optimal power flow has no possible
solution using conventional variables of types (1) to (4).
Then, it is necessary to use a special process, with type (5)
variables, such as load shedding or changes in the network
structure.
11.5 Extensions of the statement for static optimization
In the static optimization considered so far, extensions to
special cases or refinements may sometimes be present in
the problem statement, e.g.:

valve point loading

sparse generating capacity
environmental considerations
corona losses

HDVC lines

multiarea operation optimization

11.6 Dynamic despatch and control
Optimal power flows are of basic importance for long-range
planning and daily scheduling, but their final purpose is
real-time secure and optimal operation. In static optimization, real and reactive loads are the given quantities. In real
time, loads change, giving a dynamic aspect to the problem.
Three ways of solving this problem may be given, in order
of increasing refinement:
(1) Perform static optimal power flows periodically and
use the usual control devices. This process presents an
important drawback: without any addition, frequency
control (or classical B coefficient dispatch) are inconsistent
with the matching of line flow constraints. Similarly existing
or planned voltage secondary regulations do not seem to
meet the reactive constraints.
(2) Perform static optimal power flow periodically and
compute a by-product as 'participation factors' allowing
control devices to meet the constraints; this corresponds
to following a foreseen trajectory for the loads. This
process represents progress but does not allow for unforeseen
changes in the loads.
(3) Perform static optimal power flows periodically and
compute a by-product allowing control devices to meet
the constraints corresponding to the actual loads. The
latter process is not classical. We shall see later that a whole
family of optimal power flow solutions gives the elements
necessary to build the by-product. Processes (2) and (3)
correspond to two statements of the 'dynamic optimal
power flow'.

Electrical Power & Energy Systems

III. General features of the solutions

II I. 1 Some historical breakthroughs
The history of optimal power flows may be followed in the
references, which are arranged in chronological order. For
optimal power flows, fundamental prehistory corresponds
to economic dispatch, a basic synthesis of which was given
by Kirchmayer 2 in 1958. Security was not taken into
account, as also in the work of Squires 4 which certainly
was the first attempt to solve load flow and economic
dispatch at the same time. In 1962, n security appeared
in a fundamental work from the present author, the socalled 'injections method 's was introduced, where the
optimal power flow problem with security was stated,
optimality relations established and solved by a fairly crude
method.
Then, after some years of little activity, 1968 saw the
arrival of four new important methods, appearing simultaneously: the Dommel and Tinney reduced gradient
method12; the Sasson Hessian method13; the Wells linear
programming method H and the author's differential injections method 14 using the generalized reduced gradient. The
basic new concepts contained in these methods are still
valuable today and, to the author's knowledge, the decade
following was mainly dedicated to improvements both on
the theoretical and practical sides, the application of the
optimal power flow in real time being undertaken now.

sum of the generated real powers at bus i,

if any
sum of the generated reactive powers at bus i,
if any
real load at bus i
reactive load at bus i
real power injection at bus i = Pi - - C i
reactive power injection at bus i = Qi - Di
real power generated at bus is given by the
generator k (actual or fictitious for piecewise
linearization)
sum of the real loads of the system
'reactive generalized injection' in i, used in a
load flow process. R i may be equal to Vt or K i
= Q i - O i o r t o another quantity when mentioned
subscript for 'transposed'. Usually vectors
without T will be column vectors
column vector with N r o w s Vi: same definition
for all the variables above with index i

Pi
Qi
Ci
Di

Ii
Ki
Pik

C
Ri

T
V

E =

'

W=

,J =

~E

J'" J'

= Jacobian (2N x 2N)

classic economic dispatch

p = p(W) = p(E) Total real losses of the system

Z, Z m, Z M Variable, minimal and maximal corresponding
values
AZ
variation of Z
~21
constrained quantity in the line or transformer 1 for the intact system; basically the
current;when mentioned, apparent or real
power flow, or phase angle difference
g2~
value of f21 after the element d of the system
tripped
T1
upper limit for ~ 1
T'1
upper limit for g2d (usually f2d > ~~l)
pg
variable transformer tap ratio, for voltage
magnitude
ok
variable transformer tap ratio, for phase shift
u
vector of the (physical) control variables of
the system (here control ~ independent)
x
vector of the state variables of the system,
defined as the variables different from the
control variables (here state * dependent)

general static optimization noncompact methods

X:[~]

general static optimization compact methods

F = F(X) = F(x, u) objective function, usually operation

total fuel cost

Peculiar aspects of static optimization methods are then

investigated:

g(JO
= 0= 0 tj (physical) load flow equations
g(x, u)

111.2 Method families

Among the various solutions reviewed, the main difference
appears to be the 'compactness' of the method. We shall
say a method is 'compact' if, directly or as a by-nroduct, it
provides a model of the system expressed only versus the
(physical) control variables defined in the problem statement, with security inequalities, if any, all being included
in the model. Thus, the classic economic despatch is compact. Hessian and Dommel Tinney reduced gradient are
noncompact, since their system models use all the variables.
Compactness generally involves greater complexity but
allows wider applications.
Starting from this viewpoint, we shall review in turn:

special purpose extensions

discussion of the n - 1 security
overconstrained cases
Finally dynamic optimal load flows and control will be
considered.
111.3 Main notations and symbols
We shall use the following notations and symbols:

i ~ [0, NI

Vi
Oi

bus number; 0 refers to the slack bus

voltage magnitude at bus i
voltage phase angle at bus i

Vol 1 No 1 April 1979

vector of the variables of the system (whole set)

h(X) ~< 0 }
h(x, u) <~0 set of inequality constraints
L

Lagrangian function

IV. General static optimization methods I: classic

economic dispatch
The classic economic dispatch 2 is limited to real optimization, taking losses into account, without security and with
a dynamic character. Its formulation is compact.
It is based upon the property that the control variables of
a system are linked by a single equation, the real po~er
balance, so that the problem is written:
Min F(P) = ~ Fi(Pi)

(1)

with

(2)

~ Pi = C + p(P)
i

suited to accept constraints as tile compact methods and

are not really practical for extension to real tinre operation.
V.2 Direct Kuhn and Tucker methods

Basically, by a separate computation, tile losses p are

expressed versus the generated real powers Pi using a quadratic formula p = p(P) such as:

In these methods, the optimality conditions are established

through the Kuhn and Tucker theorem and solved directly
(whereas in all the other methods the gradient criterion is
used). The earliest was the "injections' method.

~}p

(3)

- - = 2 E BliP/+ B

aPi

Lagrange's method gives the optimality conditions:

d L = 0 with L = F ( P ) -

X(~/Pi -P

Minimize F(P) under the R)llowing constraints:

C)

li(O, V)

Qi + Di : 0 i 'injections' relations, V i

(6)

(7)
(s)

aF/a&

- ?, = constant

(4)

(~p/~iPi)

Q~" ~ Qi <~Q~4

Equations (4) are the so-called 'coordination equations'

which are straightforward to solve taking (3) into account.
The B coefficients are computed using various approximations and may be employed over a very wide region, a few
sets of coefficients being sufficient for all kinds of operations.
The method clearly allows dynamic dispatch, which is
a very important quality and is certainly the main reason
for its extensive use. Various improvements and extensions
have been carried out, such as:

Pi + Ci : 0 i physical load flow or

Ki(O , V)

which gives:

l -

V.2.1 Injections method. In tile early sixties, the following

problem was presented by the author:

<s,

vF' < v, v

0i

for every generating bus

0/~< T0

(9)
(10)

(Jl)

v i
l\)r any couple i, / of
neighbouring buses

(]2)

This is one general statement of the n security real reactive

complete optimal power flow problem.
Tile optimality conditions were found using the Kuhn and
Tucker theorem. These conditions are well known: the
reader may find them in English in ref 65.

extension to multiarea economic dispatch 3,24,6s

improvements in the computat,ons of tile loss quadratic

formula 10
direct computation of the 'first order differential !osses'
OP/~Pi as proposed by Happ 48 by the transposed
Jacobian:

a#
Op _ j T I __
OP
O0

(5)

The main drawback of this simple method is the absence

of security: this concept is now introduced.

The numerical solution was obtained by attempting

to meet the optimality conditions by a try and check
algorithm based upon the values of the dual variables
appearing in the conditions. As this numerical nrethod was
elaborated in 1961, a relaxation process was used (matrix
sparsity, triangularization and Newton's method were used
in France only one year later "7,s). Compared to present-day
algorithms, file process was slow and had convergence
difficulties in some heavily constrained cases.
Interesting by-products were found in the l'ornl of dual
variables, e.g. ~.i = OF/OCi and ,I/i aF/~Di, incremental
values of the real and reactive delivered powers at the bus i.
This is the reason (in addition to the dual character of the
algorithm) why one of the corresponding programs is still
in use in France for capacitor localization planning studies
in large scale networks.
=

V. General static optimization methods I1: noncompact secure optimal power flow methods
v. 1 General characteristics

The problem statement includes inequalities, so that

Lagrange's method is theoretically no longer applicable;
the conditions for optimality must come either from the
application of the theorem of Kuhn and Tucker I or from
putting the projected gradient of a convenient function
equal to zero. The noncompact methods apply these processes directly to a system model including the whole set
of variables, without building an intermediate 'reduced
model' limited to the control variables u. They are in
general relatively easy to program and may exhibit high
performance often due to the sparsity of the physical load
flow equations. However, they are generally not so well

V.2.2 Other methods. The main quality of the algorithm

associated with the Injections method was to exist. Progress
occurred later, Shen and Laughton iv being most successful
when they generalized the statement introducing transformer
tap settings (voltage magnitude and phase shifts) and apparent
power flows through the lines, and used Newton's method
in the algorithm.
It seems to be a general property of naethods of this family

that even frmn a purely practical point of view, it is always

difficult to guarantee convergence a priori.

Electrical Power & Energy Systems

V.3 Sasson Hessian methods

V.3.1 Principle. In these methods, which are able to handle
the complete real reactive problem, network equations
g(X) = 0 and the inequality constraints h(X) <~ 0 are taken
into account through penalty functions. A new objective
function C(X) is built up, such as:

C(X) = F(X) + rTg2(X) + s r h 2 ( X )

(13)

h(x, u) <~ 0 including x ('functional constraints') are taken

into account through penalty functions so that an artificial
cost function f(x, u) is produced such as:
f i x , u) = F(x, u) + rrh2(x, u)

Then,f(x, u) must be minimized subject to the load flow

equations g(x, u) = 0. Applying Lagrange's method gives
dL = 0 with L(x, u) = f ( x , u) + kTg(x, U) (XT row vector):

r T and ST are penalty coefficient row vectors, s K being 0 if

h K ( X ) < --eK, with eK small and positive.

~3L_0f+[~3g']

ax

Then, let G(X) be the gradient d C / d X and H(X) the Hessian

dZG/dX 2 of C(X). With convenient values for r and s, the
constrained optimum is reached for G(hr) = 0. This point
is approached with a quadratic-type accuracy changing X
by AX, such as

C(X)

G is sparse and H is still sparse, which may allow relatively

fast computations.
V.3.2 Practical aspects and improvements. The first research
in this field was performed by Sasson 13,18,19 and Cory 29,33
using the Fletcher and Powell method without computing
the Hessian but building it more and more accurately by
iteration.

The real reactive problem with n security was solved

initially; Cory introduced transformer tap ratios 29 and
n - 1 (and even n - 2 ) security 33 using a simplified sensitivity
matrix built-up using the DC approximation. However, in
the latter case, the problem size seems limited, and the
Fletcher and Powell method fails for problems involving
more than 100 variables.
Sasson and others 26 therefore improved the algorithm very
satisfactorily, computing the Hessian directly and using
its sparsity.
More recently, Barcelo and others 66 built a program for
the real problem only with ( n - l ) security. They use a
simplified Hessian with the so-called 'Diflex Hessian'
approximation, and implement the inequality constraints
in a new way, including slack variables. The algorithm was
initially designed for real time despatch and they started
front a point not too far from the optimum and obtained
fast results with small storage location for a 1200 bus
network. This enjoyed considerable success, but it should
be noticed there were only 22 generation buses, the use
of slack variables is not at all equivalent to that of linear
programming, and there were at most five constraints.
With a compact method this would correspond to a very
large physical load flow but to a small optimization problena, which would certainly be at least as fast and easier
for an online implementation.

V.4 Dommel- Tinney reduced gradient

V.4.1 Principle. This method, initiated by Dommel and
Tinney 12 is certainly the most popular in the United States
and may solve the complete real-reactive problem. It is an
extension of Newton's method and uses matrix sparsity
intensively. The variables X are partitioned into control
variables u and state variables x. The constraints on u are
nothing but bounds and are trivial. The constraints

Vol 1 No 1 April 1979

0u

(16)

~, = 0

(17)

7-

--=g(x,u) = 0

(14)

k=0

ax l a x l r
-

0u
0L

H(X) A X =

(is)

(18)

0k
Differentiating g(x, u) = 0, it is easy to show that OL/Ou
is the reduced gradient V f o f f versus u considered as independent variables, i.e. d f = Vfrdu. So, the basic minimization algorithm, for n security only, is as follows:
(1)
(2)
(3)
(4)
(5)

assume an initial value o f u

solve the load flow problem (18), which gives x
compute k from the linear set of sparse equations (16)
compute the reduced gradient V f = OL/Ou from (17)
project 7 f o n t o the feasible region, i.e. cancel the
components V~. which would give an unfeasible
direction for u l, taking account of the bounds on u i.
This replaces V f b y V~f
(6) ifV~t "= 0 stop;if not go to (7)
(7) move u by Au = - c V f a n d go to (2)
Once Newton's method is programmed, this algorithm is
extraordinarily simple to implement and may profit intensively by matrix sparsity. The total process does not generally require many iterations. The major drawback lies in the
choice of c which may require trials for a new system (an
approximation for the diagonal terms of the Hessian
allowing computation of c). The use of penalty functions
to handle the inequality constraints does not seem to be
troublesome, and this was not clear a priori.
V.4.2 Practical aspects and improvements. From the
beginning, the algorithm had a high performance, solving
the real-reactive problem with variable tap settings for
networks with more than 300 buses and 80 control
variables. The original algorithm 12 was limited to n security.

An alternate algorithm was written by Velghe and Peterson, 37 including area interchange equations. Alsac and
Stott 49 ingeniously generalized the algorithm to n - 1
security. They considered not only the state variable x
corresponding to .the intact system but also all the variables
x l, x 2, . . . , x k corresponding to all the studied contingencies, as well as the related dual vectors M, k 2, . . . , k k,
u having a single value except in some special cases. The
numerical example presented was limited to 30 buses and
15 control variables: as could be foreseen from the number
of variables considered, it seems that the n - I security
requirement cuts down the performance of the method.
Other refinements were performed sl,67 essentially using a
Hessian to define c, or real-reactive decoupling. 36

Concerning tire use of penalty functions to handle constraints, it is true for this reduced gradient method and for
the previous Hessian method that constraints are nret in a
'soft' manner, with a relatively large tolerance zone; this is
certainly acceptable in many cases but causes problems for
some difficult security problems.

(6) 'Nonlinearities in the constraints are taken into account

and Z" is conrputed again to meet the nonlinear constrainls.
which gives a final value:
Z =7~2,2 j

(23)

such that
V.5 Peschon GRG m e t h o d
V.5.1 Principle. A number of other noncompact methods
were proposed and it is impossible to be exhaustive.
A special mention must be made of the application of the
generalized reduced gradient (GRG) method 9,16 of
Peschon 27 and others. Peschon, whose previous research
with Tinney and others (among whom CuOnod had originated the reduced gradient method just studied) presented
GRG in 1971 and compared it with the Dommel Tinney
method.
The GRG is a general convex programming method. Several
versions exist. Peschon used the general program, which
may roughly be summed up as follows.
(1) At each step, the constraints C(Z) = b are linearized,
such as:
A dZ

db

(19)

(2) The linearized constraints are handled exactly as in

linear programming: the variables Z are partitioned into
basic (or dependent) variables Z and nonbasic (or independent) variables Z: A dZ + A d 2 = 0. The (square) basis
A is inverted by recurrent formulae, e.g. the product form
of the inverse, and the reduced gradient G(Z) = d F ( Z ) / d 2
is computed:
OF

C(2) = - - z

bF
- A-'~

az

-:-

az

(2o)

(3) G ( Z ) is projected onto the feasible region to take an

account of the bounds on Z, which gives G'(TZ)
[f G'(') = 0, stop. Otherwise go to (4)

{21)

and
(22)
which defines the variation
A Z = ph

meeting the linearized constraints.

(5) p is taken as the minimum of:

P l minimizing F

P2 such that a variable Zj reaches one of its bounds

P3 such that a variable Zk reaches one o f its bounds
This gives a value of Z: Zj = ZI, ZI

(24)

This step uses the general Newton method to solve a set

of nonlinear equations. It is called 'RDD' ('Rentrde dans
le domaine' which means 'reentering tile feasible region')
(7) l f p = pj or P2 go to (2). I f p = ,P3, ,~k has reached one
of its bounds. Change the partition Z, 2" and consequently
the basis A , exchanging Zk for a convenient nonbasic
variable 2 / a n d go to (2).
Moreover, to get more accuracy and speed in (2) tile
reduced gradient G(2') is computed replacing F by F f,S,
tt being the dual variable row vector and S the nonlinear
part of C in the neighbourhood of Z. This is called Beale's
correction.
This is tile genuine GRG algorithm, the principles of which
were used for optimal power flows by Peschon and the
present author, 14"31,3s,ss,s6,77,78 and should not be confused with some language extensions referring lo completely
difl'erent algorithms. 37
A series of tests on convex programming carried out by
Colville around 1968 showed that GRG is one of the mosl
secure, high performance, general methods in convex
programming.
Peschon applied tile general GRG program to the complete
real-reactive optimal power flow with variable transformer
tap ratios. He applied GRG to the whole problem, Z being
the vector X including all tile variables. This work was done
essentially fl)r demonstration, and he did not write a special
GRG code but used the general form.
Results and discussion. P e s c h o n c a r r i e d o u t tests o n
networks of up to 30 buses. For these cases, despite the
fact that he used a nonspecialized GRG program, tie found
that GRG was well suited to optimal power flows, giving
the same performance as tile Domnrel Tinney reduced
gradient. He also mentioned that with a specially written
GRG code, results should be much better.
V.5.2

(4) Compute

= -c'(2)

C(Z) = C(22, Z l ) = b

(23)

This experiment is of high interest for two reasons first

its results and second the history of GRG which is rather
entertaining. GRG was conceived by the present author in
1964 (the reader may, by the way, recognize here the use
of Newton's method), not in order to solve a general mathematical program but the optimal power flow problem, the
first internal name of GRG being 'differential injections';
(optimal power flow method which will be presented in the
following section). In 1965, it was perceived that differential injections could give a good by-product for general
mathematical programming which the author called GRG
and which operated at the end of 1965. At that time, Abadie
joined the author and from 1966 worked alone on GRG,
improving the general code, while the author, appointed in
Greece, kept working on differential injections only. This

Electrical Power & Energy Systems

explains why the GRG general code fitted so well with optimal power flows in Peschon's experiment, GRG being a
by-product of optimal power flow and not the reverse as
may be believed. This also corroborates Peschon's opinion
about the efficiency of a specialized program for two
reasons: on the one hand, in the GRG-differential injections algorithm, special additions and short-cuts exist
which make it much more efficient; on the other hand it
is applied only to the physical control variables of the
problem. This shows that our field of study may sometimes
be useful to more general applied mathematics; the most
unfortunate thing being that circumstances prevented
Peschon and the author having direct contact.
Peschon did not limit himself to the usual optimal power
flows but altered the program to perform optimal load
curtailment under emergency conditions, which is an application which deserves the greatest attention.

VI. General static optimization methods II1:

compact secure optimal power flow methods
Vl. 1 General characteristics
Compact methods are characterized by the existence of an
intermediate 'reduced model' of the system expressed
versus the (physical) control variables, all the constraints
being expressed versus the latter variables. This reduced
model is a compact representation of the whole system,
valuable over a wide region, and allows optimization to be
performed. The algorithms include two completely different
parts: constructing the reduced model through sensitivity
and optilnizing the reduced model subproblem thanks to
mathematical programming. Calling these two jobs a 'step',
we may distinguish three families of methods, in increasing
order of sophistication: one step linear, sequential linear
and nonlinear, the two latter methods involving several
iterative steps. These methods are in general not so easy
to program as the noncompact methods and need more
storage location, the reduced problem matrix not being
sparse. Computation times may be very short depending
on the sophistication either of the method or of the desired
results within a given method and in any case increase more
than linearly with the number of effective constraints.
Consequently, performances are unique for real optimization
where the number of effective constraints is always small
and comparable with noncompact methods for complete
or only reactive optimizations, where the number of
effective constraints is greater. By nature, they fit well with
n - 1 or n - k security and meet constraints 'hardly', i.e.
exactly, which is favourable for security. Moreover, the
use of the reduced model alone may always be fast; as the
complete system behaviour is represented, the reduced
model may easily be used for real time operation and
control, which is a very important property.

VI.2 One step linear methods

VI.2.1 Principle. The principle is very simple: first make
the objective function linear or piecewise linear if it was
not so already. Then, linearize the network equations. Use
these linearized equations to eliminate the state variables
(corresponding to the intact system or to a contingency
case) from the constraint inequalities and from the objective
function if necessary. Add the linearized real power balance
relationship: the result is a linear program which is solved
by usual dual or primal simplex algorithms.

Vol 1 No 1 April 1979

To linearize, DC load flows or more accurate sensitivity

computation techniques (through the Jacobian) may be
used: DC load flow neglects the losses whereas sensitivity
techniques allow them to be taken into account to some
extent.
For the real problem the method is particularly fast and
has been used for years. A program for the real-reactive
problem has recently been written for use online on a minicomputer.
VI.2.2 Practical aspects and developments. The real problem was initiated by Wells ll and improved by Shen and
Laughton. 21 Both may handle n - 1 and n - 2 security constraints. Wells considers a global system with a spinning
generating capacity constraint which is linear in P; kaughton
and Shen do the same and, moreover, meet spinning spare
capacity constraints for each generator modelled as a set
of linear inequalities in Pik. They use the DC load flow
approximation to build the reduced model and the revised
dual simplex method to solve the linear program. They are
able to solve very large problems very quickly.

To solve a similar problem to Laughton's, Sterling 63

noticed that the constraints have a horizontal basis block
angular structure, the satellite elements being the generator
spinning spare capacity constraints. So, alter building the
reduced model through the DC approximation, he used
the Dantzig and Wolfe decomposition algorithm to solve
the problem in a primal simplex manner.
Very recently, Stott and Hobson 69 published a paper with
a number of improvements:
general n - k security constraints
phase shift angles and HDVC line settings
losses are taken into account approximately, including
them in loads
they distinguished between preventive and corrective
control, noticing that the normal n - 1 (or n - k ) insecure
operating state may be subdivided into 'insecure correctable' and 'insecure uncorrectable', the latter being
the only one to be considered in the preventive control
for corrective control, they propose different possible
objectives and introduce emergency start-up and load
shedding
Concerning computational aspects, they use a symmetric
Jacobian to model the intact system state and a DC approximation for contingency. Construction of the" reduced model
uses sensitivity and matrix sparsity. Optimization is performed with a revised dual simplex method. Numerical
performances are excellent for large scale systems (462
buses, 70 controlled generator units, 1 s CPU in IBM370/
158). Accuracy is better for the intact system than for
contingency constraints and the need for an extension of
linear programming to convex programming is indicated.
Aschmoneit, Ruhose and Wagner 71 recently solved the
complete linearized real-reactive problem with a program
written for a minicomputer (up to 150 buses, 50 control

variables, 50 constraints; an example with 109 buses, 33

control variables, 19 constraints lasted about 1 min). They
used sensitivity and sparsity to build the model and the
simplex method. The validity of the linearization is certainly good for the real variables and constraints but seems
fairly doubtful for the reactive parl of the problem.

VI.3 Sequential linear methods

VI.3.1 Principle. The previous methods may be improved
by performing several steps, each step consisting of building
the linear reduced model then optimizing it. A typical
solution of this kind is the 'Maya' model developed by
Merlin 34 for real optimization only.

In Maya, model building uses a simplified load flow with

approximated losses p. In phase I, a feasible solution is
computed by the dual simplex algorithm, introducing the
effective inequalities only one after the other and improving
the value o f ~ used at each iteration.
In phase II, the objective function, which is piecewise linear,
is optimized. The power balance equation is written so as
to keep the values of the losses limited to a first order
Taylor series development. Primal simplex is used to optimize. But, due to the limitation of the first order of the
Taylor development of the losses, the coefficients of the
real power balance change from the beginning to the end of
the linear program. A sequence of linear programs, associated with a dichotomic process, is therefore necessary to
converge towards the solution of the optimization: using
this process, a nonlinear program is solved by a sequence of
linear programs.

the possibility of using the reduced model separately.

This is valuable over a wide region and 1nay be used in
real time with high performance
The earliest and most typical among these methods is
certainly the differential injections method, described now
in detail, as an example.
VI.4.2 Differential injections method14,31,35,42,ss, s6, 77,78

Principle. 14,42 The differential injections method is a

general method for optimal power flows with various
options, the basic one being the complete real-reactive
problem with variable transformer tap settings for voltage
magnitude, n - 1 security for currents in lines, and, which is
unique, voltage level n - 1 security.
Its main characteristics are:
(a) The nature ofahe reduced model where
the real power balance relationship is developed by a
Taylor series up to the second degree (in exactly the
same way as in the classic economic dispatch)
the security inequality constraints are linearized
~APi.
/

, z S d / 7Id2p
' [ ~ ] ' A/t= 0

(25)

for every selected constraint k

(26)

l~-~oI7 ,Art

Au ~< 3,k
[ du o ]

(b) Tile use of the generalized reduced gradient method

(GRG) to optimize the reduced problem

VI.3.2 Practical aspects. Maya may rapidly solve large scale

(500 bus) real problems with n - 1 security. It is currently
applied for planning studies where it often plays the part of
a subroutine. Real time implementation is foreseen for
1981.
Khan and Pal s9 realized a program very similar to Maya,
also with n - 1 security. Duran 41 proposed a solution of the
real-reactive problem with n security, by sequences of
equality-constrained problems, which appears intermediate
between one step and sequential linear compact methods.
VI. 4 Nonlinear compact methods
VIA. 1 General characteristics. In the nonlinear compact
methods, the reduced model is nonlinear and is solved
using a general nonlinear programming method, sometimes
specially adjusted to the particular nature of the problem
being solved.

(c) The computation of first and second order sensitivities

to build the reduced model
(d) The intensive use of matrix sparsity properties in
Newton's physical load flow and sensitivity computations.
The quadratic character of equation (25) is fundalnental
for it gives validity to lhe reduced model over a large
region and practically guarantees convergence. Philosophically, this reduced model, as well as the whole method,
appears as a direct extension, in the presence of constraints,
of the classical economic dispatch, to which it is reduced
(with exact B coefficients) in the real power only option if
no constraints are effective.
The basic algorithm is as follows:
(1) settle the initial values of the control variables u

This gives the general following properties:

(2) by a Newton's load flow compute the state variables x
a sophisticated program
often a lower speed and greater storage location than for
linear programming techniques
an outstanding accuracy and an exact solution concerning
the optimum as well as constraints meeting
a secure convergence, due to the use of a general strong
nonlinear programming technique

10

(3) analyse the constraints and select those which are

outside, on or near their limits, which gives the set of
the 'selected constraints' to be included in the reduced
problem. This includes n- 1 security analysis
(4) build the reduced problem by sensitivity computations,
including first dp/duo and second d2p/du2o 'differential
losses', the latter being computed only at steps (1) and
(2) [only at step (1) for the real only option].

Electrical Power & Energy Systems

(5) from the second step, if the new reduced problem is

identical with the previous one, stop; otherwise go
to (6)
(6) optimize the reduced problem by GRG and go to (2)
In (3) the degree of accuracy of the n - 1 security analysis
for line currents may be chosen by the user, ss from a
linearized option up to an exact computation. The latter
is not so time consuming as might be thought, because
effective constraint 'angling' is first performed with the
linearized option and accurate computations performed
only for selected constraints.
The n - 1 voltage level security ss guarantees after a generator
trip changing the reactive power balance, that voltage at
each generator bus will remain within acceptable limits.
In any case, the after-trip variables are expressed versus
the control variables before tripping, such coefficients as
'distribution factors' for line currents and similar coefficients
for voltages being computed and stored to be used in the
constraints, which appeared sufficient for accuracy since
at the last step Au ~ 0
In (4), sensitivities are computed using the transposed
Jacobian, as shown for the first differential losses as early
as 1962 (reference 5)
In (6), GRG is specialized to the reduced model the
constraints of which are linear except for one which is
quadratic. Compared with the general version, presented in
paragraph 5.5.1, changes are as follows:
Point (5): Pl minimizing G is computed directly, which
avoids a long search. Point (6): RDD is carried out at each
GRG iteration only by solving a second degree equation.
One, two or three times per GRG, a Newton load flow
allows computation of the actual value of the constraint
slack variables and an RRD process is performed to meet
the actual nonlinear constraints, i.e. to compensate all the
approximations due to the truncated Taylor developments
of the reduced model: tile latter process is equivalent to
having an exact convex reduced model, which makes the
method exact. Moreover, as in the general GRG code, the
optimization itself is preceded by a phase 1 finding a
feasible point in a dual simplex manner.
Practical aspects and extensions. The method has been
improved for 10 years, and four main options now exist :
real-reactive (ID), real only (IDA), reactive only (IDR)
and decomposed real-reactive (IDD).
The existence of ID, IDA and IDR allowed a study of real
reactive decoupling possibilities s6 to be carried out. It was
shown that only the reduced model was worth decoupling.
This gave the IDD option, in which one builds two reduced
models, one for the real and one for the reactive variables
and constraints. This option needs less storage location
and tends to replace ID.
The average number of steps is 3 to 4, except for the real
only option for which it is 2 to 3.
The convergence is very secure; there is no empirical
coefficient to adjust and the method is straightforward-

Vol 1 No 1 April 1979

portable. The necessary storage locations are relatively

large for reactive and complete options.
Computation durations, although not so short as for linear
methods, remain acceptable: in 370/168, from 2 s for the
real option of a 100 bus and 30 control variable system
to a few minutes for a 480 bus system with 220 control
variables in a real-reactive option in a nearly overconstrained case. As yet, the programs have been used mainly
for studies concerning five different countries.
A reactive only option was run on a minicomputer prepared for the Algerian system; it has been running with
simulated data for one year and is just being installed now
in Alger for real-time use: for a system with about 100
buses and 35 control variables, the response time is 1 min
when alone and may reach 3 min due to other simultaneous
tasks.
Mainly for data and computer reasons, the programs have
not yet been used for scheduling or in real time in France,
the present-day objective being real time implementation
for 1981.
Promising extensions are possible towards combination with
secondary voltage regulation, 78 discussion of the n - 1
security,77 overconstrained cases and the use of the reduced
models for real time 35,78 are in view.

VI.4.3 Other compact nonlinear methods. As it seems to

be the most advanced and is well known by the author,

the differential injections method has been presented in
detail, but many other searchers found excellent algorithms,
especially Sjvelgren and Bubenko, s3,62 lnnorta, Marannino
and Macenigo, s8 all of whom as in IDD solve the complete
real-reactive problem with n - 1 (line) security using
separate real-reactive reduced models.
Sjvelgren uses a linear real reduced model solved by dual
Simplex; the reactive reduced model is nonlinear; Sjvelgren
tried two methods to solve it and Zoutendijk successive
linear programs and quadratic programming: he found
quadratic programming the better method.
In Innorta's real-reactive decoupling, the reactive model
has the interesting special property of handling the current
limits. Sophisticated mathematical programming methods
are used: Griffith and Stuart algorithms, Dantzig and Wolfe
decomposition and Rosen projected gradient methods.
Both Sjvelgren and Innorta methods handle large scale
systems.
Wollenberg and Stadlin, s2 in an excellent paper, solve the
real problem with n security by Dantzig and Wolfe decomposition for nonlinear programming. They also compute
constrained participation factors for online operation and
give very interesting solutions for the overconstrained case.
Nabona and Freris 61 solve the real-reactive problem with
n security and may also perform optimal allocation of
spinning reserve by 'quadratic approximation programming', the idea of which is the same as Beale's correction
in GRG, and Podmore 47 uses the Rosen projected gradient
method to solve the real problem with n security.

11

Makala and Laiho v3 and Kopflnan 7s proposed a 'direct

search' algorithm. The author is not sure that they always
reached tire optimum: Nielsen and Poulsen s7 compared
this algorithm with the Dommel and Tinney method and
found the latter definitely better.
Finally, a very interesting method but difficult to classify
was proposed by Glavitsch 4s for real-reactive online
applications: it solves the reduced problem directly from
Kuhn and Tucker theorem.

VII. Particular aspects of static optimization

methods
VII. 1 Special purpose extensions
Global spinning reserve and sparse generating capacity 11,21,61,63 may naturally be introduced into the problem statement, adding linear or quadratic constraints.
HDVC lines may be introduced with the same formulae
as phase shifters. 69 Coronna losses 73 correspond to ground
inductances varying linearly with voltages in the normal
voltage range, which may give optimal voltage levels not
as high as usual. Valve point loading was performed by
Fink 2 with classic economic dispatch. It is combinatorial
and does not seem to have been generalized with security
constraints. Environmental considerations 23,46, so arise
from the necessity to limit some thermal plant effluents
such as nox and to limit thermal pollution. Both are
increasing functions of the generation power produced and
the problem may be restated so as either to minimize
pollution or limit it or better, combine it with fuel cost in
the objective function, e.g. through corresponding tax
values. Finally, multiarea operation has been solved in an
excellent way for classic economic dispatch 3,24,6s but
remains a challenge when security is consideredfl 8

VII.2 Discussion of the concept of n - 1 security

If n - 1 security is met, then the sudden trip o f one element

will not provoke a general failure of the system, n 1
security constraints prevent the trip of one element from
causing the trip of a second element. Sometimes, it may
happen that the trip of this second element has no other
effect on the network, and thus tire trip cascade stops.
In the latter case, the n - 1 security constraint was not useful
and moreover prejudicial because it increased the generation
cost. This was recently perceived 69,77 and a theoretical
solution proposed by the author in reference 77. When a
n - 1 constraint is effective, on one hand the corresponding
operation overcost is estimated, e.g. through the corresponding dual variable, or on the other hand the trip cascade is
simulated, e.g. by n- k contingency and the resulting nondelivered power computed. Multiplying the cost of this nondelivered power by the probability that the first considered
element trips gives the mathematical expectation of the
possible damage if the n - 1 constraint is not met (enormous
for a general breakdown, comparable to the overcost for a
small amount of nondelivered power, zero if the cascade
effect stops). The comparison of the operation overcost and
of the damage-expected-cost then allows a choice of keeping
or dropping the considered n - 1 constraint. This solution is
still at a theoretical stage but may be implemented easily
with a fast n - k contingency analysis method: corresponding
recurrent fornmlae, allowing computation of n k contin-

12

gency quickly once n 1 has been carried out are given in

reference 77, and should be compared to those given in
reference 69.
Finally, this analysis allows the abandonment of nonusefid
n 1 constraints, giving rise to savings and also avoiding
overconstrained cases.
V l 1.30verconstroined cases
Overconstrained cases are of prime importance. They
happen .when the constraints are such that the feasible
region is empty and no solution exists. Then the problem is
what to do.
As in the Wollenberg and Stadlin s2 analysis, one may try to
reduce the 'overloads el', the amount by which the constraints are surpassed. For this purpose, they minmrized
functkms of e:
~ a i e i, Maxei, ~,aie 2
i

and found the latter objective led to the most reasonable

solution. Peschon 2v searched for the optimal load curtailment meeting the constraints. Stott and l lobson 69 searched
for enrergency start-up and load sheddings meeting the
constraints. Front the author's point of view, the solution
should be as follows: drop the n - I non-useful constraints
as in section VII.2, then meet all other n-- 1 (and of course
n) constraints by load shedding, emergency start-up, and
even structure changes whenever possible.
This problem is certainly the most important in optimal
power flows today;often, in such emergency situations,
dispatchers have not the same scientific attitude as model
builders: they usually gamble to avoid little load sheddings
and usually win (but the corresponding probability is not
unity), Fast and secure automatic load shedding (and incidentally start-upt computation programs and control
devices would certainly be the solution to this problem.

VIII. Dynamic optimal power flows and control

As stated in the problem statement in 11.6, the final objective
is real time secure and optimal operation of the system.
Corresponding to point (1) of this statement, if the
optimal power flows program is fast enough, this may be
carried out periodically and the usual control devices
employed. This gives an approximate solution and is better
than nothing. For this case as well as for cases 2 and 3, real reactive decoupling is valuable, ()wing to the system feedback. Real power implementation is straightforward. An
example of reactive power implementation of this kind is
given by the Alger dispatch centre, already mentioned.
An important improvement concerning real power consists,
as a by-product of an optimal power flow, in computing
'participation factors' sharing forecasted load variations
between the various plants of the s y s t e m . 32,44, .':;2,64
In this direction, Patton 44 in particular takes an account
of 'change-related costs' due to power changes in the
plants. Wollenberg and Stadlin s2 compute 'constrained
participation factors' taking account of the line current
security constraints, which is a very important improve-

Electrical Power & Energy Systems

ment. Adler and Fisch164 base the computation of these

constrained participation factors on the worst case bus
load variation forecasts.
To guarantee the matching of constraints, the latter solution
is good as far as the actual load variations remains the same
as forecasted. To take account of the actual load variations,
the classic economic dispatch might be extended to use
the reduced models of compact optimal power flows on
the one hand, and data coming from state estimation on
the other, to perform a secure control meeting the security
constraints. An example of such a possible application was
proposed by the author for real powers as using the real
power reduced model of the differential injections method.
It was even possible to introduce dynamics into this
process, which thus became a secure generalized frequency
control process.
The use of the reactive reduced model of compact optimal
power flows makes it possible to build a voltage control
regulation. The author is studying such an application 78
leading to a tertiary voltage control: primary control is
given by the usual local generator voltage control; secondary
voltage control will, as in France, incorporate five to ten
generation buses, the reactive powers of which are linearly
dependent (which cuts down computation durations for
the differential injections method, changing only the
variable R i in the physical load flows; for secondary control
see also reference 74); the reduced model would give a
tertiary secure and optimal control.

References
Kuhn, H W and Tucker, A W 'Non linear programming' Proc. 2nd Berkeley Symposium on Mathematics,
Statistics and Probability University of California
Press, Berkeley, California (1951)
2

Kirchmayer, L K Economic operation of power

systems Wiley, New York (1958)

Kirchmayer, L K Economic control of interconnected

systems Wiley, New York (1959)
Squires, R B 'Economic dispatch of generation directly
from power system;voltages and admittances' AIEE
Trans. PAS Vol 52,.Part III (1961), pp 1235-1244

Carpentier, J 'Contribution ~ I'~tude du dispatching

~conomique' Bulletin de la Soci~t~ Frangaise des
Electriciens Ser 8, Vol 3 (August 1962)

Carpentier, J and Siroux, G 'L'optimisation de la production ~ I'EIectricit~ de France' Bulletin de la Socidt~

Francaise des Electriciens, Ser. 8, Tome IV, No 39
(March 1963)
7 Carpentier, J and Canal, M 'Ordered eliminations'
Proc PSCC 1, London (1963)
8 Carpentier, J 'Application of Newton's method to
load flow computations' Proc. PSCC 1 London (1963)

Carpentier, J and Abadie, J 'Generalisation de la

IX. Conclusion
Since their introduction in the early 60s, optimal power flows
have developed and even seem to provide savings 31,6s,66 all
the more interesting as fuel costs increase.
But their existence was due to a need for security, and in
this field, even if considerable work has been performed,
progress is still necessary,-especially for real time applications: developments based upon applications of the reduced
models of the compact methods will certainly be very
helpful, as well as a deeper analysis of the security constraints discussion and of overconstrained cases.
Finally, better contacts between researchers and a better
understanding of others' methods would be very profitable
ou the world scale: the author cannot help claiming his
strong faith in the usefulness of this Journal to meet this
objective.

m~thode du gradient r~duit de Wolfe au cas de contraintes non lin&aires' Proc. 4th IFORS Meeting
Boston (1966)
10 Dopazo, J F, Klitin, O A, Stagg, G W and Watson, M
'An optimization technique for real and reactive power
allocation' Proc. IEEE Vol 65 (1967), pp 1877-1885
11 Wells, D W 'Method for economic secure loading of a
power system' Proc. lEE Vol 115, No 8 (August 1968}
pp 1190-1194
12 Dommel, H W and Tinney, W F 'Optimal power flow
solutions' IEEE Trans. PAS Vol 87 (1968) pp 18661876
13 Sasson, A M 'Nonlinear programming applications to
power systems' Proc. Symposium Helors-IFORS
Athens (1968)

14 Carpentier, J, Cassapoglou, C and Hensgen, C

Acknowledgments
The author thanks his wife Rosine whose kindness and
understanding allowed him to write this paper, especially
for the most studious Christmas Day of her life. Thanks
are also due to his friends from Electricit6 de France who
helped him in a period of 'overconstrained' health, especially Mrs Ferrier, Mrs Bruneau, Mr Girard and Mr Kalil for
their Saturday working, and all his foreign friends whose
friendship considerably helped him. He also wants to thank
in advance all those who will write him to mention an
error, an omission or a new development, or to ask for any
information related to this review.

Voi 1 No 1 April 1979

'Injections diff&entielles, une m~thode de r~solution

g~n&ale des probl~mes de dispatching economique
sans variables enti~res utilisant le proc~d~ du gradient
r~duit g~n~ralise' Proc. Symposium Helors-IFORS
Athens (1968)
15 Carpentier, J, Cassapoglou, C, Bellon, J L, Bonneau, P
and Xirokostas, D 'Injections totales, une m~thode
g~n~rale de r~solution des probl~mes de dispatching
economique avec variables enti~res, tenant compte
exactement des co~ts de marche ~ vide des groupes'
Proc. Symposium Helors-IFORS Athens (1968)

13

16 Abadie, J and Carpentier, J 'Generalization of the

Wolfe Reduced Gradient method to the case of nonlinear constraints' in Optimization Academic Press
(1969), pp 37-47
17 Shen, C M and Laughton, M A 'Determination of
optimum power system operating conditions under
constraints' Proc. lEE Vol 116, No 2 (February 1969),
pp 225-239
18 Sasson, A M 'Nonlinear programming solution for the
load flow, minimum loss and economic dispatch
problems' IEEE Trans. Vol PAS-88 (April 1969),
pp 399-406
19 Sasson, A M 'Combined use of the Powell and Fletcher
Powell nonlinear programming methods for optimal
load flow' IEEE Trans. Vol PAS-88 (October 1969),
pp 1530-1537
20

Fink, L H 'Economic dispatch of generation via valvepoint loading' IEEE Trans. Vol PAS-88, No 6 (June
1969), pp 805-811

21 Shen, C M and Laughton, M A 'Power system load

scheduling with security constraints using dual linear
programming' Proc. lEE Vol 117, No 11 (November
1970)

31

Carpentier, J, Cassapoglou, C, Brellas, I~1and Margaritidis, P 'Elaboration des incidences de nouvelles

methodes de calcul elaborees en Grace sur la securite
et I'~conomie de I'exploitation ainsi que sur I'~quipement d'un dispatching automatique' Proc. Symposium
on the automation of dispatch centers, CEE paper
GE 71-16364 (October 1971)

32 Bechert, T E and Kwatny, H G 'On the optimal

dynamic dispatch of real power IEEE Trans. PAS
Vol PAS-91 (May-June 1972), pp 889-898
33 Cory, B and Henser, P 'Economic dispatch with security
using nonlinear programming' Proc. PSCC 4 (1972)
34 Merlin, A 'On the optimal generation planning in a
large transmission system (The Maya model)' Proc.
PSCC 4 (1972)
35 Carpentier, J 'Results and extensions of the methods
of differential and total injections' Proc. PSCC 4
(1972)
36 Adielson, T 'Determination of an optimal power flow
by iterative suboptimizations' Proc. PSCC 4 (1972)
37 Velghe, J 'Optimal control of real and reactive power
flow under constraints' Proc. PSCC 4 (1972)

22 Despotoric, S T, Babic, B S and Mastilovic, V P

'A rapid and reliable method for solving load flow
problems' IEEE Trans. PAS-90 (1971), pp 120-130

38 Dy Liacco, T 'Design considerations for an advanced

system operation center' Proc. PSCC 4 (1972)

23 Gent, M R and Lamont, J W 'Minimum emission

dispatch' Proc. PICA (1971), pp 27-38

39 Stott, B and Alsac, O 'Fast decoupled load flow',

presented at the IEEE PES summer meeting Vancouver,
Canada (July 1973)

24 Aldrich, J F, Leuer, J F and Happ, H H 'Multi-area

dispatch' Proc. PICA (1971 ), pp 3 9 - 4 7
25 Dy Liacco, T E, Wirtz, B F and Wheeler, D A 'Automation of the CEI system for security' Proc. PICA
(1971), pp 93-102
26 Sasson, A M, Viloria, F and Aboytes, F 'Optimal load
flow solution using the Hessian matrix' Proc. PICA
(1971), pp 203-209
27 Peschon, J, Bree, D W and Hajdu, L P 'Optimal solutions involving system security' Proc. PICA (1971),
pp 210-218
28 Jaimes, F J and El Abiad, A H 'Optimization by a
sequence of equality constrained problems' Proc.
PICA (1971), pp 219-227
29 Henser, P B and Cory, B J 'Solution of the minimum
loss and economic dispatch problems including real and
imaginary transformer tap ratio' Proc. PICA (1971),
pp 228-236
30 Sasson, A M, Aboytes, F, Cardenas, R, Gomez, F and
Viloria, F 'A comparison of power system static
optimization techniques' Proc. PICA (1971), pp 3 2 8 340

14

40 Takahashi, K, Fagan, J and Chen, M 'Formation of a

sparse bus impedance matrix and its application to
short circuit study' Proc. PICA (1973), pp 63-69
41

Duran, H 'A simplex like method for solving the

optimum power flow problem' Proc. PICA (1973),
pp 162-167

42 Carpentier, J 'Differential injections method, a general

method for secure and optimal load flows' Proc. PICA
(1973), pp 255-262
43 Carpentier, J 'Total injections method, a method for
the solution of the unit commitment problem including
secure and optimal load flow' Proc. PICA (1973),
pp 263-268
44 Patton, A 'Dynamic optimal dispatch of real power
for thermal generating units' Proc. PICA (1973)
pp 403-411
45 Glavitsch, H 'Economic load dispatching and corrective
rescheduling using on-line information on the system
state' Proc. PICA (1973), pp 412-420
46

Lamont, J W and Gent, M R 'Environmentally-oriented

dispatching techniques' Proc. PICA (1973), pp 4 2 1 427

Electrical Power & Energy Systems

47 Podmore, R 'Economic power dispatch with line

security limits' IEEE Trans. PAS (January-February
1974), pp 289-295
48

Happ, H H 'Optimal power dispatch' iEEE Trans. PAS

(May-June 1974), pp 8 2 0 - 8 3 0

49 Alsac, O and Stott, B 'Optimal flow with steady state

security' IEEE Trans. PAS (May-June 1974), pp
745-751
50 Delson, J K 'Controlled emission dispatch' IEEE Trans.
PAS (September-October 1974), pp 1359-1366
51

Rashed, A M and Kelly, D H 'Optimal load flow

solution using Lagrangian multipliers and the Hessian
matrix' IEEE Trans. PAS (September-October 1974),
pp 1292-1297

52 Wollenberg, B F and Stadlin, W O 'A real time optimizer for security dispatch' IEEE Trans. PAS
(September -October 1974), pp 1640-1649
53 Sjelvgren, D V and Bubenko, J A 'Decomposition
techniclue in a security related optimal power flow'
Proc. PSCC 5 (1975)
54

64 Adler, R B and Fischl R, 'Security constrained economic dispatch with participation factors based on
worst case bus load variations' IEEE Trans. PAS-96
(March-April 1977), pp 347-356

Despotovic, S T 'Optimal power dispatch with line

security limits' Proc. PSCC 5 (1975)

65 Happ, H H 'Optimal power dispatch. A comprehensive

survey' IEEE Trans. PAS-96 (May-June 1977), pp
841-853
66 Barcelo, W R, Lemmon, W W and Koen, H R 'Optimization of the real-time dispatch with constraints for
secure operation of bulk power systems' IEEE Trans.
PAS-96 (May-June 1977), pp 741-757
67

68 Mamandur, K R C and Berg, G J 'Economic shift in

electric power generation with line flow constraints'
IEEE Trans. PAS-97 (September-October 1978),
pp 1618-1626
69 Stott, B and Hobson, E 'Power system security control
calculations using linear programming' IEEE Trans.
PAS-97 (September-October 1978), pp 1713-1731
70

55 Carpentier, J 'System security in the differential

injections method for optimal load flows' Proc. PSCC 5
(1975)
56 Carpentier, J, Saminaden, V, Boull6, D, Girard, R and
Nguyen, V T 'Real and reactive decoupling possibilities
in optimal load flows: a compact version of the
differential injections method' Proc. PSCC 5 (1975)
57

58

59

Nielson, J and Poulsen, W 'Comparison of different

numerical methods for economic load dispatch under
security related constraints' Proc. PSCC 5 (1975)

Innorta, M, Marannino, P and Mocenigo, IV] 'Active and

reactive power scheduling with security and voltage
constraints' Proc. PSCC 5 (1975)
Khan, M A and Pal, M A 'Security constrained optimization of power systems' Proc. PICA (1975), pp
61-67

60 Vuong, G T and Robichaud, Y 'Emergency control of

a power system using on-line security constrained
optimization' Proc. PICA (1975), pp 6 8 - 7 0
61

Nabona, N and Freris, L L 'Optimum allocation of

spinning reserve by quadratic programming' Proc. lEE
Vol 122, No 11 (November 1975)

62 Sjelvgren, D 'Application of mathematical programming to electric power system expansion planning'

PhD Thesis Dept of Electric Power System Engineering,
the Royal Institute of Technology, Stockholm (1976)
63 Sterling, M J H and Irving, M R 'Constrained dispatch
of active power by linear decomposition' Proc. lEE
Vol 124, No 3 (March 1977)

Vol 1 No 1 April 1979

Bala, J L and Thanikachalam, A 'An improved second

order method for optimal load flow' IEEE Trans.
PAS-97 (July-August 1978), pp 1239-1244

Fernandez, R A, Happ, H H and Wirgau, K A 'The

application of optimal power flow for improved
systems operations' Proc. PSCC 6 (1978), pp 465-472

71 Aschmoneit, F C, Ruhose, K H and Wagner, G G

'Steady state sensitivity analysis for security enhancement' Proc. PSCC 6 (1978) pp 4 8 1 - 4 8 4
72 Wallach, Y 'Remarks on parallelisation of economic
dispatch calculations' Proc PSCC 6 (1978) pp 4 8 5 - 4 8 8
73 Makela, L and Laiho, Y 'Optimal voltage level and
reactive power control using an approximated sensitivity matrix' Proc. PSCC 6 (1978), pp 489-492
74 Van Meteeren, H P and Maas, J 'Secondary voltagereactive power control by constrained least square
minimization' Proc. PSCC 6 (1978), pp 493-496
75

Kopfman, G 'Management of electrical power systems

at minimum cost using the sensitivity matrix' Proc.
PSCC 6 (1978), pp 497-501

76

Fletcher, R and Ramsay, B 'Optimal electric power

system scheduling by use of a multiplier penalty
function method' Proc. PSCC 6 (1978), pp 502-508

77 Carpentier, J 'Discussion of the security concept in

optimal power flows taking into account economical
and stochastic aspects' Proc. PSCC 6 (1978), pp 5 0 9 516
78 Carpentier, J 'Optimal voltage scheduling and control
in large scale power systems', proposed paper to the
IFA C Symposium on computer applications in large
scale power systems, New Delhi ( 16-18 August 1979)

15