IEEE Transactions on Power Delivery, Vol. 4, No.

1, January 1989

735

OPTIMAL SIZING OF CAPACITORS PLACED ON A RADIAL DISTRIBUTION SYSTEM

Mesut E. Baran

Felix F. Wu

Department of Electrical Engineering and Computer Sciences

University of Califomia, Berkeley
Berkeley, CA 94720

Abstract The capacitor sizing problem is a special case of the general

capacitor placement problem. The problem is to determine the optimal
size of capacitors placed on the nodes of a radial distribution system so
that the real power losses will be minimized for a given load profile.
This problem is formulated as a nonliiear programming problem. The
ac power flow model of the system, constraints on the node voltage
magnitudes, and the cost of capacitors are explicitly incorporated in the
formulation A solution algorithm, based on a Phase I - Phase II feasi-,
ble directions approach, is proposed. Also presented are a new formulation of the power flow equations in a radial distribution network and a
numerically robust, computationally efficient solution scheme. This solution scheme is used as a subroutine in the optimization algorithm. Test
results for both the capacitor sizing problem and the power flow solution
schemes are presented.
1. INTRODUCTION
Capacitors are widely used in distribution systems for reactive
power compensation to achieve power and energy loss reduction, voltage
regulation, and system capacity release. The extent of these benefits will
depend upon how the capacitors are placed on the system, i.e., (ii) the
number and location, (i) the type (fixed or switched), (iii) the size, and
(iv) the control scheme of the capacitors.
Therefore, the capacitor placement problem in distribution feeders
consists of determining the place ( number and location), type, size, and
control scheme of capacitors to be installed such that the benefits mentioned above be weighted against the fixed and running costs of the
capacitors and their accessories.
The early analytical methods for capacitor placement are mainly
developed by Neagle and Samson [l]. The problem is defined as determining the location and size of a given number of fixed type capacitors
to minimize the power loss for a given load level. The results, the simplest of which is the celebrated 2/3 rule, were based on a simplified, voltage independent system model obtained by: (i) considering only the
main feeder with uniform load distribution and wire size, (ii) takig into
account only the loss reduction due to the change in the reactive component of the branch cumnts, (iii) ignoring the changes in the node voltages, and (iv) neglecting the cost of capacitors. Later, a more
comprehensive derivation of these rules was given by Bae, [5]. Fawzi
et. al. [ l l ] incorporated the released substation W A and the voltage
raise at light load conditions into the model.
These analytical methods are also integrated with some heuristics
to obtain computer oriented solution methodologies and thus to overcome some of their shortcomings. Cook [2] extended the formulation of
the problem to include peak power and energy loss reduction and pro-

88 WM 065-5
.4 p a p e r recommended and approved
by t h e IEEE T r a n s m i s s i o n and D i s t r i b u t i o n Committee
of t h e IEEE Power E n g i n e e r i n g S o c i e t y for p r e s e n t a t i o n a t t h e IEEEIPES 1988 W i n t e r Meeting, New York,
New York, J a n u a r y 31
F e b r u a r y 5, 1988. Manuscript
s u b m i t t e d September 1, 1987; made a v a i l a b l e o r
p r i n t i n g November 13, 1987.

posed such a solution method to determine the number as well as location and the size of capacitors. Chang [4] considered voltage limits in
his solution scheme.
Duran [3], by exploiting the discrete nature of the capacitor sizes,
used a more realistic model Tor the reeder with many sections of
different wire sizes and concentrated loads and proposed a dynamic programming solution method. Ponnavaiko and Prakaso [12] considered
load growth as well as system capacity release and voltagc raise at light
load conditions and used a local optimization technique called the
method of local variables by treating capacitor sizes as discrete variables.
Recently, interests in Distribution Automation and Control (DAC)
have generated quite a lot of new studies in capacitor placement problem. Grainger and Lee pioneered in refoimulating the problem as a nonh e a r programming problem and in developing computationally efficient
solution methods. In [6], they generalized Cooks formulation for a
radial feeder, regarded the capacitor sizes as continuous variables, and
developed a computationally simple, iterative solution scheme. They
later introduced switched type. capacitors with simultaneous switching [7]
and a voltage dependent model for loss reduction [8]. In [9], together
with El-Kib, they proposed. a solution method to determine the optimal
design and real-time control scheme for switched capacitors with nonsimultaneous switching under certain assumptions. Grainger and Civanlar considered the optimal design and control scheme for continuous
capacitive compensation case [lo] and introduced a voltage dependent
solution scheme [14]. El-Kib et. al., [16] extended the methodologies
developed in [6]-[9] to encompass unbalanced three-phase feeders.
Kaplan [13] extended the formulation for multi lateral feeders and proposed the use of heuristics for the solution. Grainger and Civanlar [151
combined capacitor placement and voltage regulator problems for a general distribution system and proposed a decoupled solution methodology.
The basic idea in our formulation of the general problem is similar
to that of Grainger et. al., i.e., formulating it as a nonlinear programming
problem. We will consider a capacitor placement problem that (i) incorporates directly the ac power flows as system model, (i) enforces the
voltage constraints. We develop a solution methodology by using
decomposition techniques. At the lowest level of the decomposition, we
have special capacitor placement problcms, callcd the buse problems. A
base problem is a sizing type problem which is used to find the optimum
sizes of a set of selected capacitors placed on the nodes of the distribution system and thcrefore, it is of practical intercst in its own right. The
solution algorithm for the sizing problem is a special Phase I - Phase I1
Feasible Directions method which takes into account the structure of the
problem. We present the formulation of the sizing problem and the
solution algorithm developed for it in this paper. The solution of the
general problcm is presented in another papcr [17].
Also in this paper, a new formulation of the ac power flow equations for the radial distribution systems is introduced and a computationally efficient and numerically robust solution methodology, callcd DistFlow, is presented. Such a solution methodology is needed because of
its repeated use in the optimization algorithm.
In section 2, the new power flow equations for radial distribution
systems are presented. The sizing problem is stated in section 3. Section 4 describes the general approach for the solution. The structure of
the problem is studied in section 5 and the proposed solution algorithm
is prcsented in section 6. Section 7 contains the tcst runs and conclusions are given in section 8.

0885-8977/89/0190-0735$01 .WO1989 IEEE

136

II. DISTRIBUTION SYSTEM POWER FLOW

Power flows in a distribution system obey physical laws (Kirchoff
laws and Ohm law), which become part of the constraints in the capacitor placement problem. In our proposed solution algorithm for the capacitor placement problem, the distribution system power Row solution is
to be used as a subroutine in every iteration. Therefore, it is essential to
have a computationally efficient and numerically robust method for solving the distribution system power flow.
In this section, we present the new power flow equations for radial
distribution systems. The formulation is conducive to efficient solution
methods. For pedagogic convenience, we first consider a special case
where there is only one main feeder. The general case for any radial
distribution system is considered next. To simplify the presentation, the
system is assumed to be balanced 3-phase system.
Special Case :Radial Main Feeder
Consider a distribution system consists of a radial main feeder
only. The one-line diagram of such a feeder comprising n branches /
nodes is shown in Fig.1.
Vl

vi

Fig.1 : One line diagram of a main distribution feeder

assumed to be constant. Lines are represented by a series impedance
and loads are treated as constant power sinks,
S, = PL + jQL. Shunt capacitors to be placed at the nodes of the system
will be represented as reactive power injections.
With this representation, the network becomes a ladder network
with nonlinear shunt loads. If the power supplied from the substation,
So = P o + j Q o is known, then the power and the voltage at the receiving
end of the first branch can be calculated as follows.
z, = r, + j x l

- S', = so-

2,

lS,I*/V,2

- s,

Repeating the same process yields the following recursive formula for
each branch on the feeder.
pi+,= pi - ri+,(p:+e:)/v: - p f i + ,
Qi+l

= Qi

(1.i)

- Xi+l(P?+Q?YV? - Q ~ i + t+ Qci+l
(1 .ii)
- 2(ri+,Pi+xi+,Qi)+ ( r z , + x ~ , ) ( P : + Q ~ ) / V(l.iii)
~

Where,
Pi,Qi : real and reactive power flows into the sending end of
branch i+l connecting node i and node i+l,
Vi : bus voltage magnitude at node i ,
Q , : reactive power injection from capacitor at node i.
Eq.(1). called the branch flow equation, has the following form
Xoi+l = foi+l(Xoi*ui+l)
(2)
Where, xgi = [PiQi V:lT and
= Qci+l.
Note that if there is no capacitor at node i, then ui does not appear in
Eq.(2). By abusing notation, we will simply use U as an nc dimensional
vector containing the nc Capacitors to be sized i.e.,

Note also that we have the following terminal conditions:

(i) at the substation; let the given substation voltage be Vsp , then

% = v; = v s P =
(ii) at the end of the main feeder;
Xk,

= P, = 0

V"

..

xkl=Q,=O

Figure 2 : One line diagram of a section of a feeder

For lateral k with nk branches, the same process of formulation
applied to the main feeder can be repeated for the lateral by using the
line flow equations of (1) and the new terminal conditions,
Vko= V, , P,,, = O , Q,,, = 0 . Where, we have used the dummy
variable V ~ for
O notational simplicity. As a result, we have the following 3(nk+l) equations.
i = 0, . . . ,nk-1
xL+, = fb+l(xki
,UL+,)

V , B , = v, - 211, = VO - z , s ; / v ,

Vi:, = V:

(4)

G(xo,u) = 0

Where, 4 = [x&. . . &IT

is the branch variables and U is the capacitor
sizes. For a given load profile (i.e., set of demands P f i , Qfi :
i = 1 , . . . ,n ) and capacitor sizes (connof variables). U , 3(n+I) DistFlow equations can be used to determine the operating point, x,, of the
system. Such a solution algorithm is developed in the appendix.
General Case
The DistFlow equations can be generalized to include laterals.
Consider a lateral branching out from the main feeder as shown in Fig.2.
For notational simplicity, the lateral branching out from node k will be
referred to as the lateral k and the node k will be referred to as the
branching node.

Y----?

In the figure, V o represents the substation bus voltage magnitude and is

s, = so- s,,

as DirtFlow Equations. They are of the form

(3.i)
(3.ii)

The 3n branch flow equations of (2) together with the 3 boundary

conditions of (3) constitute the system equations and will be referred to

Hence. in general, for a distribution network of n branches and 1

laterals, there are 3(n+1+1) DistFlow equations (ac power flow equations) comprising of Eq.(5) for each lateral k = 0, . . . ,l, including the
main feeder as the O'th lateral. Thcy are of the form
G(x,u) = 0

(6)

DistFlow equations can be used to determine the operating point. x

of the system if the capacitor sizes, U are given. We prefer to use DistFlow equations over conventional ac power flow equations because the
special structure of the DistFlow equations can be utilized to develop a
computationally efficient and numerically robust solution algorithm. The
details of the derivation of the solution algorithm are presented in the
appendix. Other distribution systcm load flow methods have been proposed, see [18] and the references cited therein.
III. SIZING PROBLEM IN CAPACITOR PLACEMENT
The capacitor placement problem is to determine the location, type,
and size of the capacitors to be placed on a distribution system. The
objectives are to reduce the losses on the system and to maintain the
desined voltage profile while keeping the cost of the capacitor addition at
a minimum. In this general form, the problem includes the following
factors: (i) the location of the capacitors, (ii) the type of capacitors -fixed
or switched, (iii) load variations, and (iv) the cost of capacitors. A solution scheme for this general problem is proposed in another paper [17].
The proposed approach calls for the solution of a Sizing Problem as a
subroutine.
The sizing problem is a special case of the capacitor placement
problem; it assumes that
the capacitors are placed (thcir number and places are
given),
there is only one load profile,

131

the cost of a capacitor is a differentiable function of its size.

Therefore, the sizing problem determines the optimal sizes of the capacitors once they are placed. This special case was considered by Lee and
Grainger in [8]. However, in this proposed formulation, the voltage
projile of the system (voltages at all nodes) is required to lie within the
acceptable limits.
To formulate the sizing problem as a nonlinear programming problcm, consider a radial distribution system with n branches, J laterals, and
nc capacitors placed at the nodes of the system and a load profile,
Pli , Q, , i = 1, . . . ,n. The objective comprises two terms; the savings due to real power loss reduction in the system as a result of capacitor placement, and the cost of capacitors. The real power losses in the
network can be calculated as the sum of the i2r loss on each branch,
i.e.,

(7)
The other term in the objective function, namely the capacitor cost function, is to be assumed linear with a constant marginal cost r,$/kvur for
simplicity. Then the objective can be written as

Where, kp is the weighting factor for the losses.

There are three sets of constraints.
(i) The power flow equations : G(x,u)= 0
(ii) Voltage limits : vY2 I xk3 = V: I vkw2
(iii) Capacitor Limits : 0 < U 5 U'""
To summarize, the sizing problem in capacitor placement is

PS

min fo = kpp[x)

nc

CrcJuJ

(ii) Objective function of PS does not seem to be flat near the optimum

as in the OPF case.

Based on these considerations, a first order method, called Phase I
- Phase I1 Feasible Direction.@Method, is selected for the solution of the
PS problem. The method obtains the solution in two steps. The first
step, Phase I, 'is applied to get a feasible point if the initial point is
infeasible. In the sccond step, Phase 11, the solution is improved iteratively until the solution converges to the optimal point. This is an
appealing property of the method since it allows one to stop earlier in
the solution process and yet get an approximate and feasible solution.
Also to improve the convergence, the method incorporates the functional
inequality constraints directly in the direction of movement, i.e., search
direction, unlike the reduced gradient-penalty function method in which
these constraints are treated as penalties.
The first step in applying this method to PS involves the transcription of the problem into a nonlinear programming problem with inequality - constraints - only form by eliminating the equality constraints of
DistFlow equations, (6). Since, as shown in the appendix, the Jacobian
is well dcfined, it seems reasonable to
of the system equations,

a
ax

assume that the conditions for the implicit function theorem hold and
there is a solution x(u) for all 0 < U I U'"" such that G(x(u),u) = 0,
i.e., one can solve x in terms of U using Eq.(6). Then the problem
reduces to the following inequality constrainted-type problem.
min ( fo(u) I f J ( u ) 5 0

PC

(%I

C p j V f J 1 2 1 CpI = 1 , pI 2 0 )
I&

SI.

Xh+1

= fh+&

4%
= xm3

&+l)

xkn,= xe, = 0

k=O...I
i = 0 . . . nk-1

v p I X h , = vh' I Vh-2
05

us U'""

This is a discrete optimal control type problem; branch flow equations acts like the difference cquations and the terminal conditions constitute mixed boundary conditions. We first discuss the general solution
approach to this problem and then the special properties of thc problem.
An efficient solution algorithm is proposed by taking advantage of these
propertics.
IV. FEASIBLE DIRECTIONS METHODS
There are basically two types of methods for solving the sizing
problem type nonlinear programming problems [21] - [23]; the ones that
use only first order derivatives - the first order methods - and the ones
that require the computation of the second order derivatives - the second
order methods.
For power system applications, the first order methods are easy to
implement in particular, the power flow solution can be used as a subroutine in the optimization. However, early expericnce with the reduced
gradient-penalty function method for optimal power flow
problems,(OPF) revealed occasional convergence problems due to zigzaging of solution points from iteration to iteration [19]. It is believed
that this is caused by the fact that the objective function is rather flat
near the optimum.
The sclection of the solution method for the sizing problem, PS is
based on the following considerations.
(i) PS has quite a large numbcr of equality constraints due to DistFlow
equations. These equality constraints can be eliminated by noting the
fact that they can be used to solve x in terms of U and we have a very
efficient solution method, DislFlow to implement it. But then the objective and the inequality constraints of PS become implicit functions of U;
therefore, the second order derivatives become difficult to compute.

A Phase I Phase II Feasible Directions Method can be applied to

solve this problem. A comprehensive review of these methods are given
in [20]. The basic factor in determining the computational efficiency of
these methods is the number of gradients of the constraints used in the
calculation of the search direction h ; for example for phase I.

h = -urgmin

J=1

j = 1 , . . . ,n

I&

where, the set M is supposed to contain all the violated constraints.

Computationally, it is desirable to reduce the size of M as much as
possible. The ideal case would be to consider only one constraint at a
time, for example, the constraint which is violated the most. But then it
can be shown that if the solution follows such an approach, it may jam
up around a sharp comer of the feasible region as shown in Fig.3. The
use of only one gradient at a time ( hl in the figure) will lead to a zigzagging solution trajectory. On the other hand, using a convex combination of two gradients, ( h in the figure ), will lead to a much faster convergence. Therefore, the incorporation of only one constraint at a time
in the calculation of the search direction should be used only if sharp
comers will not be encountered in the feasible region defined by the inequality constraints of the problem.

Figure 3 : A feasible region with a "sharp comer".

V. GEOMETRY OF THE FEASIBLE REGION

In this section, we are going to invcstigate whether wc can simplify
the calculations involved in Feasible Directions Methods by selecting
only a few gradients from the violated constraints. As mentioned earlier,
the problem boils down to chccking whether the feasible region has
sharp comers or not.
In order to have an idea about the existence of sharp comers in the
feasible region defined by the voltage constraints, we simplify the branch
flow equations of (1) by neglecting the branch loss terms
r,+l(P,'+Q,')lV,'. Then the branch flow equations become

738
(9.i)
(9.ii)
(9.iii)
Note that the real power flow is decoupled and can be dropped from calculations. This simplified DistFlow model is similar to the current
model adopted in literature [6]. The appealing properties of these equations are that they are linear and can be solved directly. The general
form of the voltage equation is as follows.
It

C c j k Qck
k=l

V', = (VjY + 2

(10)

Where, e l k is the total line reactance of the section of the network which
corresponds to intersection of two paths extending between the nodes 0-j
and 0-k.
Therefore, the upper and the lower voltage constraints become
hyperplanes and the feasible region defined by these constraints will be a
polytope. To illustrate such a feasible region, consider the small system
shown in Fig.4.
VO

V1

vz

v3

v4

v5

QO

Figure 4 : A small main feeder

Some of the voltage equations are
V? = WP)' + b i Q c z + b i Q c 4
V:

= (V,O)'

V,Z = (V,O)'

+ 2 ( ~ 1 +xZ)QCz + W

region, and (ii) sharp comers are created only when constraints of
different types are binding. Therefore, we modify the feasible directions
by including only one or two gradients from the violated constraints. If
only lower voltage limits are violated, we include the gradient of the
most violated constraint (corresponding to the lowest bus voltage) in the
feasible direction. If both lower voltage limit and upper voltage limit
violations are present, we include the gradient of the most violated lower
voltage constraint and the gradient of the most violated upper voltage
constraint in the feasible direction.
To be more specific, the selection of the feasible direction h is as follows.
1. Phase I (to get a feasible point)
Case 1: There are only lower voltage limit violations. Then, we cpnsider only the smallest voltage V, below the lower bound limit V"" ,
i.e.,
V, S V ,

-F j

and

VI <V"

and select the feasible direction as h = V?, .

Case 2 : There are only upper voltage limit violations. Then, we consider only the greatest voltage V,, above the upper bound limit VmPX,
i.e.,
V,,, 2 VI

-F j

V, > Vm"

and

and select the feasible direction as h = - V?,.

Case 3 : There are both lower voltage limit violations and upper voltage
limit violations. We consider only the smallest voltage, V, below Vm
and the greatest voltage, V, above Vmu as defined above and calculate
the feasible direction as
h = - argmin {%I h l z I h = -pL!V9,+p,,,
W,,, ; p/,p, 2 0 ; p,+p,,, = 1 )
P1.k

1 +

xz + x3)Qc4

+ 2 ( ~ 1 +xdQcz + 2 ( ~ 1+ x2 + x3 + x4,QC4

2. Phase II (to get the optimal point)

Case 1:" E,, active constraints" are lower voltage limits only, i.c,

The corresponding feasibility region is shown in Fig.5.

V, <VI

-bL j

and

V, <V"+E,

Then we select the search direction as

h = - argmin (%Ih l z I h = h V f o

- p,V?,

; h , p / 2 0 ; h+p, = 1

h.Pl

Case 2 : " E, active constrain&" arc upper voltage limits only, i.e.,
V, 2 VI

-bL j

V,,, > V"-E,,

and

Then we select the search direction as

+ p,,,V?,

h = argmin {%I h I' I h = h V f

Figure 5 : Feasible region defined by voltage constraints
Now consider the hyperplanes that corresponds to the lower bound
constraints. For the example, they are stacked on top of each other in
almost parallel fashion as depicted in Fig.5. Therefore, the intersection
of two hyperplanes of the lower bound constraints will not create sharp
comers. A sharp comer may be created by h e hyperplanes of conand V? = Vmm2 as shown in the figure.
straints V: = V"'
The foregoing observation on the example leads us to assume that
in general:
the constraints of the same type (either the lower voltage limit constraints or the upper voltage limits constraints) will not create sharp
comers in the feasible region,
sharp comers may be created by the constraints of different types.
This assumption will be used in the next section to simplify the
calculations involved in a Phase I - Phase I1 type solution algorithm.
VI. SOLUTION ALGORITHM
In this section, we will prcscnt an efficient solution algorithm for the
Sizing Problem. The algorithm involves modification of the Phase I Phase I1 feasible directions method specifically for the radial distribution

system capacitor placement problem.

From the analysis of Sec.V, it is reasonable to assume that: (i) the
constraints of the same type do not crcate sharp comers in the feasible

h
b
K
.

; po,pm2 0 ; p,$-@,,, = 1 ]

Case 3 : There are both q active lower and upper voltage limits. We
consider only V, and V,,, as defined above and calculate the search
direction as

The overall solution algorithm is presented below

step 0 :initialization
choose ~y , E >~ O ; a ,P E (0,l) ; uo
step 1 :feasibility check
if E V, <Pin
or E V, >vm"
then phase = 1 else phase = 2
step 2 :search direction
Calculate the search direction h as explained above
step 3 :step size
calculate A i such that ui + Aiih S U"
calculate the step size Xi by Armijo Rule i.e..
if phase = 1 then
yr(u,+hih) - y(ui) 2 - Xialhlz
hi = max X, = p k ~ i i
k=O. X
I
G(xi+l,u;+hih) = 0

'

mere, y(ui) = max{(VmmZ

- ?/(U,)) , (qm((ui)- vm2)]

'

739
if phase = 2 then

A summary of the power flow solutions for the test systems before
the capacitor placement is presented in Table 2. The solutions are
obtained by the DistFlow solution scheme mentioned in the appendix.
They converge in 4 and 3 iterations for TS1 and TS2 respectively.
These results indicate that the convergence of DistFlow is fast. Considering the fact that the feeders have branchcs with diverse r/x ratio (ratio
varies between 3 to 0.02). and TS1 has a low voltage and high loss
profile, the results also indicate the method's numerical robustness. The
run time figures shown in Table 2 are obtained by mnning the p r o g m
on a VAX lln50.
Table 2 : Power Flow results for the test systems by DistFlow

I
step 4 : update
Ui+l

= U, + hih

Step 5 : convergence check

If h, Ihl >
then go to step 1 else stop
Step size calculations in stcp 3 involvcs the solution of system equations, G(xj+l,u,+h, h) = 0 every time Dk is updated to get the
corresponding state variables x,+I(ui+l).Therefore, computationally it is
important to have a quick search. One way is to limit the search intcrval
by putting an upper limit, AQch on the maximum allowable change in
This will speed up the search in Phase 11
conml variables, Aii I hk I-.
since the step size will get smaller as solution approaches to the optimal
point. Also an upper bound K has to be defined on the number of
searches when implementing the search.
Another improvement usually made on the above algorithm, I201
involves the modification of feasible direction for Phase I by combining
phase I direction h, and the phase I1 direction Vf, as follows.

This way the solution will be pushed towards the optimal point in the
feasible region and hence the convergence will improve in phase 11.
The algorithm can start from any initial point, h.However, a good
initial point can easily be obtained by employing the approximate
method of [6] with simplified DistFlow equations of (9).

W.TEST RESULTS
The proposed algorithm has been implemented in Fortran-77 on
both VAX 11/750 and IBM PC-AT. In the implemented algorithm, the
optimization parameters are set as follows:
u = O A , p=O.6, ~ = 0 . 3 &
, , = 0 . 6 , ~ . = 0 . 0 4 ,K = 5 , AQch=3.0

We will present three capacitor sizing test runs and sample power flow
solutions for two test systems.
The first test system, TSl is a 9-branch main feeder developed by
Grainger et. al. [8]. We also follow Grainger et. al. in setting
kp = 168 $lkW, r, = 4.9 $lkvvar and placing three capacitors Q,b,
QClon the nodes 6, 5, 1 of the feedcr respectively. The second test
system, TS2 is a 69 branch, 9-lateral feeder derived from a poltion of
the PG&E distribution system. The network data of this system is given
in [17]. There are three capacitors placed on the system; one on the
, and two on a lateral. Qc47 , Qc52.
main feeder, Qcla

Test results of the proposed method for capacitor sizing are presented
below.
Test System 1 (TS1)
Two test N~IS
were conducted for TSl .
Test run A : In this test run, we (i) neglect the cost of capacitors. (ii)
ignore the voltage constraints. Therefore, the solution is an unconstrainted optimal point and gives the maximum loss reduction. As the
initial starting point, the approximate solution given in [8] is used. The
resulting solution trajectory is given in Fig.7. A summary of the solution and the computation time are presented in Table 3. The following
observations are made from Fig.7 and the comparison of Table 2 and
Table 3.
The convergence is achieved in about 7 iterations (2 times the
number of capacitors to be placed).
The approximate solution is not close to the actual solution as
pointed out in [8].
At the unconstrainted optimal pint. the power loss reduction is
maximum (100 kW and 4890 kvar) and the voltage profile is better
(the minimum voltage raises form .837 p.u. to .882 p.u. and all the
node voltages are still below the substation voltage).
Test run B : In this test run, cost of capacitors are considered and the
lower and upper limits on all bus voltages are set to 0.9 p.u. and 1.1 p.u.
respectively. The resulting solution trajectory is shown in Fig.8. The
trajectory starts with Phase I since the initial point is infeasible (lower
voltage limit violation). In the forth iteration a feasible point is obtained
and the program switches to Phase 11. The convergence is obtained in
the 12'th iteration. A summary of the results and the computation time
are presented in Table 3. We have the following observations about the
test"

a~,

7------

2 700

"7

Figure 7 : Solution Trajectory for TS1 Test A

600

300

--

,0

Figure 8 : Solution Trajectory for TSl Test B

5 i + r

Figure 9 : Solution Trajectory for TS2

740
The convergence in Phase I is fast; the number of iterations is
roughly equal to the number of capacitors.
The convergence in Phase I1 gets slower since the solution trajectory moves along the boundary of a lower limit voltage limit (the
total number of iterations is roughly equal to 2 to 3 times the
number of capacitors.
The incremental changes of capacitors gets smaller as iteration
increases.
The results show the feasibility of capacitor placement for both
loss reduction and voltage regulation (minimum voltage is raiscd to
.901 p.u. while still achieving 64.32 kW loss reduction).
Table 3 : Summary of test results for TSl.

I Test Run I

Substation
I Loss Reduction 1 V-:Qo(kvar) APo(kW) AQo(kvar) V,'
13 050.87 329.62
100.90 4 892.85 .882
13 087.45 279.43
64.32 4 943.04 .901
Po(kW)

Test A
Test B

I Run Time I
CPU I/O
4.38 1.3
6.08 1.9

Test System 2 (TS2)

In this test run, capacitor costs and the voltage constraints are considered also. The initial point is obtained by the application of the
approximate method of [6] with the simplified DistFlow equations
presented in section 5. The corresponding solution trajectory is given in
Fig.9. Since the the system without any capacitors is feasible, the test
run uses Phase I1 only. A summary of the solution and the computation
time are presented in Table 4. The following observations are made.
The convergence is obtained in 4 iterations (less than 2 times the
number of capacitors).
The test result.
= 0 indicates that the capacitor on bus 15 is
not ebnomical to install.
Considerable loss reduction is obtained through optimal capacitor
placemcnt (67.55 kW and 1018.39 kvar).
Table 4 : Summary of test results for TS2.
Loss Reduction
V,,,,,,/Vo Run Time
Substation
CPU VO
QoWr) @o(~W) A Q o ( ~ ~ )
1018.39
.925 10.91 0.92
3 959.55 1778.38 67.55
Po(kW)

MT. CONCLUSIONS
In this paper, a capacitor sizing problem for capacitors placed on a

radial distribution system is formulated as a non-linear programming

problem and a solution algorithm is developed. The problem finds the
optimal size of the capacitors so that the power losses will be minimized
for a given load profile while considering the cost of Qe capacitors. The
formulation also incorporates the ac power Row model for the system
and the voltage constraints.
The solution algorithm developed for the capacitor sizing problem
is based on a Phase I - Phase I1 feasible directions approach. This general method is made computationally more efficient by exploiting the
special structure of the problem.
Another contribution of the paper is the introduction of new power
Bow equations and a solution method, called DistFlow , for radial distribution systems. The method is computationally efficicnt and numerically
robust, especially for distribution systems with large r/x ratio branches.
The DistFlow is used repeatedly as a subroutine in the optimization algorithm for the capacitor sizing problem.
The test results of the proposed algorithm for the capacitor sizing
problem are also presented. They indicate that the method is computationally efficient and has good convergence characteristics. They
demonstrate the feasibility of capacitor placement for both loss reduction
and voltage regulation purposes.
The capacitor sizing problem considered in this paper is a special
case of the general capacitor placement problem. The general problem
determines the place and the type of capacitors to be installed as well as
their sizes. The load variations as a function of time are also incorporated in the general problem in the calculation of the the energy
losses. The general problem is considered in another paper [17]. The
solution algorithm developed in [17] uses the algorithm developed here
as a subroutine.
Acknowledgements
We thank Mr. Wayne Hong and Dr. Dariush Shirmohammadi of
Pacific Gas and Electric for providing the data for the test system and
helpful discussions. This research is supported by TUBITAK-TURKEY
and by National Science Foundation under grant ECS-8715132.
REFERENCES
N. M. Neagle and D. R. Samson, "Loss Reduction from Capacitors
Installed on Primary Feeders", AIEE Transactions part I l l , vol. 75,
pp. 950-959, Oct. 1956.
R. F. Cook, "Optimizing the Application of Shunt Capacitors for
Reactive Volt-Ampere Control and Loss Reduction" , AIEE Trunsuctions part I l l , vol. 80, pp. 430-444, August 1961.
H.Duran, "Optimum Number, Location, and Size of Shunt Capacitors in Radial Distribution Feeders: A Dynamic Programming
Approach, IEEE Trans. on Power Apparatus and Systems, vol.
87, pp. 1769-1774, Sept. 1968.
N. E. Chang, "Locating Shunt Capacitors on Primary Feeders for
Voltage Control and Loss Reduction", IEEE Trans. on Power
Apparatus and Systems, vol. 88, pp. 1574-1577. Oct. 1969.
Y. G. Bae, "Analytical Method of Capacitor Application on Distribution Primary Feeders", IEEE Trans. on Paver Apparatus and
SyStems, vol. 97, pp. 1232-1237, JUly/AUg. 1978.
J. J. Grainger, and S. H. Lee, "Optimum Size and Location of
Shunt Capacitors for Reduction of Losses on Distribution Feeders",
IEEE Trans. on Power Apparatus and Systems, vol. 100, pp.
1105-1118, March 1981.
S. H. Lee and J. J. Grainger, "Optimum Placement of Fixed and
Switched Capacitors on Primary Distribution Feeders", IEEE
Trans. on Power Apparatus and Systems, vol. 100, pp. 345-352,
January 1981.
1. J. Grainger, and S. H. Lee, "Capacity Release by Shunt Capacitor Placcment on Distribution Fecdcrs: a Ncw Voltage Dependent
Model", IEEE Trans. on Power Apparatus and Systems, vol. 101,
pp. 1236-1244, May 1982.

74 1
J. J. Grainger, S.H. Lee. and A. A. El-Kib. "Design of a Real-Time
Switching Control Scheme for Capacitive Compensation of Distribution Feeders", IEEE Trans. on Power Apparatus and Systems. vol.
101. pp. 2420-2428. August 1982.
J. J. Grainger. S.Civanlar, and S. H. Lee, "Optimal Design and Control Scheme for Continuous Capacitive Compensation of Distribution Feeders", IEEE Trans. on Power Apparatus and Systems, vol.
102, pp. 3271-3278, October 1983.
T. H. Fawzi, S. M. El-Sobki, and M. A. Abdel-Halim, "A New
Approach for the Application of Shunt Capacitors to the Primary
Distribution feeders", IEEE Trans. on Power Apparatus and Systems, vol. 102. pp.10-13, Jan. 1983.
M. Ponnavaikko and K. S. hakasa Rao, "Optimal Choice of Fixed
and Switched Shunt Capacitors on Radial Distribution Feeders by
the Method of Local Variations", IEEE Trans. on Power Apparatus
andsystems, vol. 102, pp.1607-1614, June 1983.
M. Kaplan, "Optimization of Number, Location, Size, Control Type,
and Control Setting of Shunt Capacitors on Radial Distribution
Feeders", IEEE Trans. on Power Apparatus and Systems, vol. 103,
pp.2659-2665, Sept. 1984.
J. J. Grainger. S. Civanlar, and K. N. Clinard, L. J. Gale, "Optimal
Voltage Dependent Continuous Time Control of Reactive Power on
Primary Distribution Feeders", IEEE Trans. on Power Apparatus
andSystems. vol. 103, pp. 2714-2723, Sept. 1984.
S . Civanlar and J. J. Grainger, "VolWar Control on Distribution
Systems with Lateral Branches Using Shunt Capacitors and Voltage
Regulators: Part 1. Part 11, Part III", IEEE Trans. on Power
Apparafus andsysKems, vol. 104, pp. 3278-3297, Nov. 1985.
A. A. El-Kib, J. J. Grainger, and K.N. Clinad, L. J. Gale, "Placement of Fixed and/or Non-Simultanmusly Switched Capacitors on
Unbalanced Three-phase Feeders Involving Laterals", IEEE Trans.
on Paver Apparatus and Systems, vol. 104, pp. 3298-3305. NOV.
1985.
M. E. Baran and F. F. Wu, "Optimal Capacitor Placement on Radial
Distribution Systems". submitted to IEEE PES winter meeting, 1988.
D. Shirmohammadi. H. W. Hong. and A. Semlyen, G. X. Luo. "A
Compansation Based Power Flow Method for Weakly Meshed Distribution and Transmission Networks" , IEEE PICA meeting, June
1987. Montreal.
F. F. Wu. G. Gross. and J. F. Luini, P. M. Look. "A Two Stage
Approach to Solving Large Scale Optimal Power Flows", PICA
Proceedings. Cleveland, May 15-18 1979. pp. 126-136.
E. Polak, R. Trahan, and D. Q. Mayne, "Combined Phase I - Phase I1
Methods of Feasible Directions", Mathematical Programming,
~01.17,N0.1, pp.61-73, 1979.
E. Polak, Computational Methodr in Optimization, Academic Press.
New York,1971
D. G. Luenberger, Introduction to Linear and Nonlinear Progrmming, Second Edition, Addison-Wesley Publishing Co., Massachusetts, 1984.
P. E. Gill, W. Murray, and M. H. Wright, Practical Optimization.
Academic Press, New York, 1981.

the beginning of the lateral, xko= [PbQkoV,'O]; since then the rest of the
variables, xy can be calculated by using the branch flow equations successively. To be more specific, let us consider two cases again, namely first
the main feeder and then the general case.
Special Case :Main Feeder Only
Since the substation voltage, Vo is given, the only Variables need to be
determined are z, = [PoQolT,i.e., the power supplied from the substation.
Therefore, z, constitutes the "state" of the system. To eliminate the other
variables, xgi from the Eq.(a.l), we use the terminal conditions at the end
of the feeder, i.e.,
p n =Po,,
(xo,,-l) = 0
;
Q, = 40. (xh-1) = 0
(a.2)
by substituting the branch flow equations, Eq.(a.l.i)
and eliminate %'s
recursively fori = n - l , n - 2 , . . . , 1. As a result, we get two equations of
the following form.

Po,, (zo0.u) = 0
;
(foe (%.U)
=0
(a.3)
We will use these two equations to solve for a given U .
General Case :Feeder with Laterals
For simplicity, it will be assumed that the system consists only of a
main feeder and 1 primary laterals (laterals branching out from the main).
We generalize the procedure of reducing the system equations as follows.
For lateral k, we choose two new variables, Pko and Qko- the real and the
reactive power flows into the lateral respectively - as the extra state variables zb = [PbQbIT . Then the same process of reduction applied to the
main feeder can be repeated for the lateral by using the branch equations
of (a.1) and the associated terminal conditions. Ph = 0 , Qh= 0 This
will give two new equations of the following form.

Ph (210,
ffh(210.

. . . * 2,
f

,zo0,u) = 0

(a.4.i)

. . ,Z b ,z,.u) = 0

(a.4.ii)

Hence, in general, for a distribution network of n branches and 1

laterals, reduced DistFlow equations comprise Q(a.4) for tach lateral
k = 1, . . . ,I together with the equations for the main feeder, Eq(a.3)
which can now be rewritten as,

. . . ,210 ,%*U)
dol, (210, . . . ,zlo .%J4

Pol, (z10.

section 2. comprises branch flow equations and the associated terminal

conditions for each lateral including the main feeder which is treated as the
O'th lateral. They are of the following form.
Xy+1

= fki+l(XY

.UY+l)

xh =xh,=O

(a.55)

'

lZ=d

(a.7)

Step 3 : solution of the following system of equations to update the states

The DistFlow equations of a radial distribution system, as shown in

X,~=Q~,

=0

H(z,u) = 0
(a.6)
'
and can be used to solve for the "state variables" z = [zL . . . ZL 41.
Let's consider solving (a6) for z by Ne,wton-RaphsonMethod (NR)for a
given U. From an estimated value of 9 ,an iteration step of NR involves
three steps:
Step 1 : calculation of the mismatches H(d)
Step 2 :construction of the system Jacobian matrix
a2

.., I

(a.1.i)

i =0,1,.. . ,&-I

(al.ii)

k =0,1,.

(a.5 .i)

Solution of DistFlow Equations

The reduced DistFlow equations are of the form

'
aH
J(z') = -

Appendix :Solution of DistFlow Equations

=0

We can use these equations to determine the operating point,

xy = [PYQYVz]of the system for a given, U. But rather then using the
a b v e equations to solve for xY's directly, we are going to conceptually
reduce the number of equations first and then propose an efficient solution
algorithm.

Reduction of DistFlow Equations

Note that for a given lateral k. we need only to know the variables at

2.

J(z')AzJ = - H(d)

(a.8)

We consider the special case - the main feeder only - first. The
mismatches for this case can be calculated as follows. At iteration j, the
substation bus voltage Vo is given and the updated values of the substation
powers Po,Q o are available. Therefore, the recursive branch Row equations of (a.1.i) can be employed to update the variables
~i = [Pi,Qi .V?lT i = 1, . . . ,n successively along the feeder starting
from the substation and proceeding towards the end of the feeder. When
we reach the end of the feeder, the updated variables P,, ,Q,, will be the
mismatches as indicated by (a.2). We'll refer to this procedure as aforward sweep.
The elements of system Jacobian can be calculated by using the Chain
Rule. For the main feeder case, the Jacobian is a 2x2 matrix of the following form.

742

(a.9)

By using the branch flow equations, (a.1.i) and applying the Chain Rule,

Numerical Properties
The DistFlow method described above is numerically robust and the
solution scheme is insensitive to system parameters, in particular, the
branch r/x ratios. This can be seen as follows.
When the elements of the branch jacobians, J i are calculated in per
unit, PA., J i has the following structure

J(z,-,,) can be constructed as follows.

Ji

1-2ri-

Pi-1

-2ri-

Vi?,

Qi-1

ri

v;2,

( P A +QA)

Ji is the Jacobian of branch flow equation i and will be referred as the

brunch Jacobian. Therefore, calculation of the 2x2 Jacobian in Eq.(a.9)
can easily be conducted by the multiplication of 2x3 and 3x3 matrices in
(a.lO).
The third step of the solution procedure involves the solution of a
2x2 matrix equation.
This calculation procedure can be generalized for general radial distribution feeders. Again, we consider a system having a main feeder and I
primary laterals only.
The mismatches can be calculated by generalizing the forward sweep
procedure as follows. For the lateral k, given the estimated power
injected into lateral, Pko.Qro.and the voltage at the branching node (substation for the main), Vk,the forward sweep method of the main feeder
case can be applied to update the variables xk = [Pk.Q,,V,] along
lateral and get the mismatches corresponding to this lateral: P , and Qh.
To update all the variables and thus obtain all the mismatches, above
procedure is applied to all the laterals in an order such that the main
feeder is updated before the laterals branching out from it.
System Jacobian, J for the general case will be a Z(l+l)x(l+l) matrix
of the following form.

JZl

J =

Jzz

.
.

JII

Jrz

Jir

Jh

JOI

Joz

Jor

JGQ

..

1 AQ;

(a.13)

(a.12)

Where,

and can be calculated by Chain Rule similar to (a.lO).

The third step involves the solution of 2(l+l)xZ(l+l) matrix equation
of the form (a.8).
Note that the mismatches and the Jacobian matrix in the proposed
method involve only the evaluation of simple algebraic expressions and
no trigonometric functiomds opposed to the standard load flow case.
Thus computationally the proposed method is efficient. In the following
tow subsections, we show that the method has very good numerical properties and it can be modified to make it computationally even more
efficient.

(a.15)

and this is true independent of the line parameters.

These results indicate that the NR solution algorithm will be numerically stable and robust. This very desirable property is one of the
advantages of the new formulation over the conventional one where convergence problems has been experienced when there are branches with
high r / x ratios in the system [HI.
Furthermore, these numerical properties can be exploited to modify
the DistFlow solution scheme to make it even more efficient computationally.
Enhancement of Computational Efficiency
We use the numerical properties and the special structure of the Jacobian matrix, J in simplifying its construction and in solving the update
equations of (a.8).
Since by (a.14). the off diagonal terms Jk , i # k k = 1. . . . , l are
much smaller than the diagonal terms, Jk , k = 1 , . . . . I , off diagonal
terms can be dropped in the Jacobian except the ones in the last row,
i.e., d = [Jol . . . Jar] . This basically decomposes the update equations
into I+1 equations as follows.
(a. 16.i)
R = 1, . . . , I
JkAzko = H ~ ( z )
J&

Ju,

.
.

Therefore, the system Jacobian, J is well conditioned, i.e.,

det(J) = 1

JIO

J11

[:I:

Where, IZ;,~I = I E I = 1 ~ ~- V:i21~ 4: 1, APi = P loss on branch i, and

AQi = Q loss on branch i. Therefore, det(Ji) = 1; indicating that Jis
are very well conditioned.
As a result of (a.13), the system Jacobian, J is well conditioned for
the main feeder only case.
For the general case, it can easily be shown that, the diagonal elements, Je of the system Jacobian has the Same properties as that of the
main feeder, J, and the system Jacobian is block diagonally dominant,
i.e.,
det (Je) = 1 > det ( J k ) -V i # k , k # 0
(a. 14)

Where,

= H d z ) - dr,

(a.16.ii)

Where, Hk = &,,(z) &,(z)IT, r, = [AzL. . . &,$IT.

Therefore, the states
that correspond to the laterals, z i o , i = 1.. . . , I can be updated
independently first by Eq(a.16.i). then the state for the main feeder, zm
is updated by (a.16.ii).
Furthermore, since Jes are well conditioned and almost constant, as
indicated by (a.13), the Jacobian needs to be constructed only once and
then it can be used in a l l the iterations. Therefore, updating step of NR
reduces to the solution of (a.16).
This modified solution scheme brings down the overall cost of computation to about 8n+8(1+1) multiplications for each iteration; 8n multiplications to calculate the mismatches and 8(1+1) multiplications to solve
the update equations of (a.16). This amount of computation is comparable to 6n multiplications involved in just solving the update equations
(like Eq4a.16)) of conventional fast decoupled power flow solution
scheme.

Mesut E. Baran received his B.S. and M.S.from Middle East Technical
University, Turkey. He is cumntly a Ph.D.student ai the University of
Califomia, Berkeley.
Felix F. Wu received his B.S. from National Taiwan University, M.S.
from the University of Pittsburgh. and W.D. from the University of California, Berkeley. He is a professor of Electrical Engineering and Computer Sciences at the University of California, Berkeley.

743
Discussion
A. Chandrasekaran and R. P. Broadwater (Center for Electric Power,
Tennessee Technological University, Cookeville, TN): The authors are to
be complimented for bringing under a single mathematical framework all
the various problems of capacitor placement, sizing, time-varying loads,
and other special features of radial power distribution systems. A lot of
rigor has gone into the analysis of transmission systems. However, radial
distribution systems have received very little, and mostly sporadic and
desultory, mathematical treatment. The paper under discussion fulftlls a
long-felt need in distribution analysis. We request the authors to comment
on the following observations.
1) Recognition of the minimum (state?) variables required to solve the
distribution systems as the branch flows in the source branch and the fxst
branches in the radials in the DistFlow algorithm is of special interest. In
[Dl] we proposed a load flow formulation based on branch losses and
voltages and the resulting Jacobean structureis similar. What is the authors
opinion about the convergence characteristics when the variables chosen are
the losses?
2) Even though much is talked about the difficulties of convergence in
radial systems, most of the practical algorithms in use are based on the
simple ladder technique of Kersting [D2] and appear to work very well. Are
there documented cases that the authors are aware of where convergence
was a problem? Is it possible to construct meaningful examples of radial
systems depicting nonconvergence with simple methods?
3) The upper voltage constraints used in Test Run B is 1.1 pu, and if the
source voltage is 1 pu, does this mean leading power factor currents can
flow in some parts of the system? Are there optimal limits for
permissible upper and lower voltages?
4) The cost factors kp and ru used in (8) are oversimplified. What will be
the effect on the solution if more complex factors are used?
We congratulate the authors once again for a very interesting paper.

References
[Dl] A. Chandrasekaran and R. P. Broadwater, A New Formulation of
Load Flow Equations in Balanced Radii Distribution Systems,
C4nadion Electrical Engineering Journal, Vol. 12, No. 4, October
1987, p ~ 147-151.
.
[D2] W. H. Kersting and D. L. Mendive, An Application of Ladder
Network Theory to the Solution of Three-phase Radial Load Flow
Problems, Presented at the IEEE Winter Power Meeting, New
York, 1976.
Manuscript received February 22, 1988.

N. Vempati and R. R. Sboults (The University of Texas at Arlington,

Arlington, TX): The authors are to be commended for their excellent paper.
Their approach of rewriting the branch flow equationsin a recursive format
enables an efficient algorithm. This is under the assumption of a radial
distribution system. When loops are encountered, similar equations can be
formulated. Would the authors comment on the robustness and speed of the
algorithm in the presence of loops.
Loads and capacitors are considered to constant power devices in the
paper. It is now well known that reactive power load is not a constant power
type load, rather it is more like a constant susceptance. Would the authors
comment on the impact, if any, on accuracy if voltage dependent load
models are not used. It has been our experience with various optimization
procedures that significant differences result when different load modeling
assumptions are used.
The authors allude to the fact that second-order methods are also
available in addition to first-order methods for the solution of nonlinear
optimization problems. They also present their rationale for implementing
the first-order method. In optimal power flow (OPF) problems, it has been
shown that first-order methods can give misleading results. The convergence rate can be too slow and sometimes results in a suboptimal solution.
The second-order Newton method has been found to be more reliable. It
would be interesting to see comparison cases between their first-order
method and a second-order Newton type method for their problem
formulation. Would the authors please comment.
We again congratulate the authors for a very fine paper, which is very
characteristic of their proven abilities.
Manuscript received February 23, 1988.

M.E. Baran and F. F. Wu :The authors appreciate the interest shown in

this paper by the discussers and thank them for their comments.
Concerning the first question by Mr. Vempati and Dr. Shoults about
the extension of the DistFlow algorithm to solve networks with loops, two
possible approaches seem workable. The first approach would be to put
DistFlow equations corresponding to the radial system together with additional equations due to looping branches and to apply Newton Raphson
algorithm to solve the resulting equations. In this case,the structure of the
system Jacobian can still be exploited to reduce computational burden and
hence retain the robusmess and the speed of the algorithm. However,
computations for each iteration will increase as the number of loops
increases since the system Jacobian will get fuller. An altemative
approach would be opening of the the loops at certain points of the looping
branches and putting current injections at these points to simulate the
effect of the loops. An iterative scheme can be developed to obtain the
solution as proposed in [18]. In each iteration, currents passing through
the opened points on the looping branches are estimated and then DistFlow equations are solved to determine the power flow in the radial system
with extra injections. The robustness of the method would still be valid
because of the usage of the branch powers as variables. However, the
speed of convergence and hence the computational efficiency of the
method would probably suffer as the number of loops increases.
DistFlow method can also be generalized to represent the voltage
dependent loads and capacitors as shunt susceptances. Modification for
voltage dependent load representation will entail making the real and eactive power components of load,
in branch Bow equations (1 .i)
,
and (1.ii) a function of voltage, (i.e. Pu+l =@fi+l(V,+l), QL+I = &+,(V,+I) ).
Similarly, reactive power injections by capacitors, e,.+, in Eq.(l.ii), can be
Q,4.+lV~l,where Q,4.+1 corresponds to the nominal capachanged as QP.+~=
city of the capacitor, to represent the capacitors as shunt susceptances. We
intend to incorporate voltage dependent loads and capacitors into our program.

Because of the fact that starting point of the algorithm can be

selected as the solution of the approximate method (see comment at the
last paragraph of seC.VI), we have not encountered the problem of slow
convergence and suboptimal solutions that are known to exist in the first
order OPF methods. The success of second order OPF is due largely to the
exploitation of the structure of the OPF equations. We believe the
development of a successful second order solution method for the capacitor sizing problem would be an interesting research project. Here, we
want also to point out another additional advantage of the method which is
not mentioned in the paper. The method used can easily be modified to
include different load/capacitor models as pointed above.
We thank Mr. Chandrasekm and Dr. Broadwater for directing our
attention to reference [Dl]. The use of branch losses and voltages as state
variables is an interesting idea. The convergence characteristics of the
same method (e.g. Newton-Raphson) applied to two different formulations
of the same problem may be different. We do not have experience on the
method using losses as state variables to comment on its convergence
characteristics.
Reference [18] pointed out the convergence problems for distribution
power flows. In addition to convergence, different methods may exhibit
different computational efficiency. The so-called simple methods may be
convergent in most cases, but the convergence is usually slow.
optimal sizing of capacitors may result in leading power factor
currents in the branches dependiig on the load profile. places of capacitors
and the voltage limits. This is more likely the case when capacitors are
used for voltage regulation purposes (like TSl TB in the paper). The permissible upper and lower voltages are set by the load demand requirements and system operating requirements. The optimality seems in practice to be dependent on local condition.
Representing the cost of a capacitor by a fixed cost and a linear capacity cost is an approximation. The capacity cost in practice will be a staircase curve. A linear cost curve is the simplest approximation of the staircase curve as a differentiable function. A quadratic function, or any other
nonlinear function may be used for a better approximation. The increase
in computation as result is minimal. However the effect, if any, on the
overall convergence is not known.
Manuscript received May 2, 1988.