1902

IEEE Transactions on Power Delivery, Vol. 5 , No. 4, November 1990

Optimal Network Reconfigurations in Distribution Systems: Part 1 : A New Formulation and A Solution Vlethodology
Hsiao-Dong Chiang, RenC Jean-Jumeau School of Electrical Engineering Cornel1 University, Ithaca, NY 14853
Abstract

A new formulation of the network reconfiguration problem for b o t h loss reduction a n d load balancing
taking into considerations load constraints a n d o p e r a tional constraints is presented. T h e number of switchon/switch-off operations involved in network reconfiguration is p u t into a constraint. T h e new formulation is a constrained, multi-objective a n d non-differentiable optimization problem with b o t h equality a n d inequality constraints. A two-stage solution methodology based on a modified simulated annealing technique a n d the -constraint m e t h o d for general multi-objective optimization problems is developed. A salient feature of the solution methodology is t h a t it allows designers to find a desirable, global non-inferior solution for t h e problem. An effective scheme t o speed u p the solution methodology is presented a n d analyzed.

SS 1

0 normaly opened

0rrki-o

node U load pdnt cc substation

ss2

Figure 1: A sample distribution system

Introduction
T h e r e are two types of switches in primary distribution systems: normally closed switches which connect line sections, a n d normally open switches on t h e tie-lines which connect t w o primary feeders, or t w o substations or loop-type laterals. T h e former are termed sectionnlizing switches a n d the latter are referred t o as tie switches (see Fig. 1). T h e s e switches are designed for both protection ( t o isolate a fault) a n d configuration management ( t o reconfigure the network). Network reconfiguration (or feeder reconfiguration) is the process of altering the topological structures of distribution feeders by changing the open/closed status of t h e sectionalizing a n d tie switches [l]. During normal o p e r a t i n g conditions, a n i m p o r t a n t operation problem in configuration management is network reconfiguration. As o p e r a t i n g conditions change,

networks are reconfigurated for two purposes: (1) t o reduce the s y s t e m real power losses and ( 2 ) t o relieve overloads in t h e network. T h e former is referred t o as network reconfiguration for loss reduction and the latter as load balancing. Another configuration management operation involves t h e restoration of service to as m a n y customers as possible during a restorative s t a t e following a fault. T h i s problem is called service restoration. I n this paper, we consider the network reconfiguration problem for b o t h loss reduction a n d load balancing. Conceptually, this problem belongs to the socalled .minimal spanning tree problem: Given a graph (i.e. nodes of the system), find a spanning tree (i.e. a radial configuration ) such t h a t a desired ol>ject,ive function is minimized while certain system constraints are satisfied. Recently, this problem has been formulated as a nonlinear optimization problem with a differentiable objective function. All the solution algor i t h m s proposed in t h e literature for solving the problem employ various techniques belonging to the class of greedy search technique, which accepts only search movements t h a t produce immediate improvement. A s a result, these solution algorithms usually achieve local optimal solutions rather t h a n global optimal solutions. An early work on network reconfiguration for loss reduction is presented by Merlyn a n d Back [2]. Their solution scheme s t a r t s with a meshed distribution syst e m obtained by first setting all switches closed and

9d Y i i 16L+-L+ ?URD A paper recoinmended and a!i:x-ovcci by the I d 3 Transmission and Distribution Cormittee of the I E ,E Power Engineering Society f o r presentation a t the I X X / P C S 1990 Winter ;:ectin&, Atlanta, Georzia, February 1, - 8 , 1990. :lanuscript subnitted Septeinber 1 , 1989; imde available f o r p r i n t i n 2 Tjecembor 27, l9S9.

0885-8977/90/1100-1902$01.00 Q 1990 IEEE

1903

then opening successively the switches to eliminate the loops in the system. Later, the solution scheme is improved by S h i r m o h a m m a d i a n d Hong [3]. In [4], two solution algorithms are presented for network reconfiguration based on some performance indices t o measure t h e degree of cOnstraint violations as well as t o check t h e optimality for loss reduction through network reconfiguration [l]. T h e y derive a simple formula based on some simplifying assumptions t o calculate the loss reduction d u e t o a branch exchange. Aoki e t al. formulate the problem of both service restoration and load balancing taking into account capacity a n d voltage constraints as a mixed-integer non-linear optimization problem a n d converts t h e probiem into a series of continuous quadratic programming subproblems [5,6]. B a r a n a n d Wu formulate the problem for loss reduction a n d load balancing as a n integer programming problem [7]. T w o efficient load flow metho d s with varying degree of accuracy are presented a n d incorporated into t h e solution algorithm which follows the approach proposed in [l]. T h e contributions of this paper are summarized as follows. Firstly, in order to truly reflect t h e objective of load balancing, we propose a system load balancing index which is a Chebyshev norm of each branch load balancing index. T h e purpose of load balancing is t h e n realized via solving a min-max optimization problem. Secondly, owing t o t h e fact t h a t these two objective functions - loss reduction and load balancing are incommensurable, we formulate t h e network reconfiguration problem as a constrained, multi-objective a n d non-differentiable optimization problem with b o t h equality a n d inequality constraints. T h i s is a step toward practical formulation of t h e network reconfiguration problem. Thirdly, we develop a two-stage solution methodology for general multi-objective optimization problems. T h i s new solution methodology allows designers to find a desirable, global non-inferior solution for the problem. Forthly, we present a modified simulat,ed annealing technique t o solve multi-objective optimization problems with the attractive feature t h a t it can yield a global non-inferior solution rather t h a n just a local non-inferior solution. An effective scheme to speed u p t h e simulated annealing technique i s presented a n d analyzed. Lastly, given a desired number of switch-on/switch-off operations involved in network reconfiguration, t h e proposed solution algorithm can identify t h e m o s t effective operations. In contrast, most of existing solution algorithms make use of all of t h e control variables in solving the problem. T h i s kind of solution is impractical because the number of operations could b e too large t o be executed in actual implement a t ion.

" 0 0

vo 1

'ok

%k+l

"Wb

\+...5
Figure 2 : One-line diagram of a main feeder with later als configuration problem is presented. T h i s formulation is a n extension of t h e problem formulations proposed by many researchers via truly reflecting the objective of load balancing.

Load Constraints Consider a main feeder with a lateral branching out from t h e main feeder as shown in Fig. 2 . T o simplify t h e presentation, t h e system is assumed t o b e a balanced 3-phase system. W e summarize the load flow equation formulated in [SI. Please refer t o [SI for details. We refer to node k as a branching node indicating t h a t there is a lateral. I n the figure, lV,l represents t h e substation b u s voltage magnitude a n d is assumed to b e constant. T h e distribution lines are modeled as series impedances q = rl jzl. Load d e m a n d a t bus i is modeled as a c o n s t a n t power sink, S L ~ = P L ~ QL*In general, the real a n d reactive power flowing a t the receiving e n d of b r a n c h i + l , P,+I, Q1+l and the can be voltage magnitude a t the sending end lq+l[ expressed by t h e following recursive set of equations

Equations ( 1 ) - ( 3 ) , called the branch potu equation, can be written in compact form

x,+1=

f,+l(X,)

(41

where X i = [P,, Q;, IKl']. The branch flow equations at branching node k of the main feeder can then be written a s
POk

= = =

POk(x0k-1)

Qok
wok1

Qok(Xok-1)

PkO

(5)
(6)

Qlio

%lOk(XOk--l)

(7)

There are several boundary conditions to be satisfied:

1. at the substation; the voltage magnitude \Voolis specified.
2. at the end of the main feeder; PO, = Po,(XO,~--I) = 0 and Qo, = Qo,(XO,-I) = (1
3. at the end of lateral k; P k n = 0 and
Qlcn

Problem Formulation
In this section, a new formulation of the network re-

=0

1904

For this system (one main feeder with one lateral), since [Link] substation voltage magnitude is specified, knowing t h e variables (Poo, Qoo, P k o , Q k o ) is sufficient t o determine the rest of the system variables. We now consider a general radial distribution system consisting of a main feeder (with n branches) with m laterals (with n1, n 2 ,..., nm branches on laterals 1,2,..., m respectively). T h e branch flow equation (1-3) also holds for every node of the system. For each lateral we choose the real power and reactive power injected from the main feeder to the lateral as variables. For instance, for lateral k, t h e variables are zko := [ P k O , Q t o ] . Introducing these two variables enables us to reduce the total number of branch flow equations by performing a n elimination process and t o derive the following equations:
POn
QOn

arising from line branches, which can be calculated as follows:

where

nb

is the total number of branches in the system

Load balancing

To consider a load balancing index of the entire system, we first use the following function as a branch load balancing
indez, which serves a measure of how much branch i is loaded

= = = =

.i)On(ZOOr

Zlo,...,

Z k O , IvOOl)

= 0 = 0

(8)

QOn(Z00j

~10,. Z ~ O ,

..,

~ V O O= ~) 0

(9)
(10)

where S, is the apparent power at t h e sending end of hrancli i and ,Far is its KVA capacity. We then define the load balancing index of the entire system, termed system load balancing indez as the Chebyshev norm (i.e. 1 1 ) - n o r m of each branch load balancing index
S m a z of [+,i = 1, ...,n b ]

Pknk Qknk
J

p k n k ( Z O O , ZlO,..., &knt(ZOO,

210, IvOOl)

z10!...> Z J O , IvOOl)

= 0 (11)

_<

k , a n d k = 1 , 2 ,...,m

or in compact form

F(z) = 0

(12)

where z := (Pio, Qio, ..., Pmo, Qmo, poo, Q o o ) ~ . T h e distribution power flow equation (13) is a set of nonlinear algebraic equations. T h e question regarding the existence and the uniqueness of solution has been investigated in [8], where i t shows t h a t the load flow solution with feasible voltage magnitude for practical radial distribution networks always exists and is unique. Operational Constraints T h e voltage magnitude of each node and line current of each branch must lie within a permissible range. Here a branch can be a transformer, a line section or a tie-line with a sectionalizing switch.

In comparison with the existing system load balancing index which is defined as the 12-norm of each branch load balancing index, t h e proposed index is more reflective of the objective-load balancing. This is because minimizing the &norm of each branch load balancing index may result in one very small and another very large branch load balancing index even though the total sum of the squares of each index is minimized. On the other hand, minimizing the Chebyshev norm of each branch load balancing index leads t o adjusting each branch load balancing index, thus achieving the objective of load balancing. Of course, from a mathematical point of view, the 12-norm is more desirable because it is differentiable but the Chebyshev norm is not. Now, we discuss the issue of coordinating the above two objective functions; namely (1) loss reduction and ( 2 ) load balancing. Since these two objective functions are incommensurable (or even incompatible), i t is not appropriate t o combine them together into a single objective function such as a linear combination of these two functions. We propose t h e following multi-objective formulation for the cost function of the problem

or in compact form
G(z)

5 0

= 1, ...,TL

(13)

Due to some practical considerations, there could b e a constraint on the number of switch-on/switch-off operations involved in the network reconfiguration. Let t h e desired number be nd and the original network configuration b e expressed as go, then we express this constraint as follows:

N(gO,g*)

nd

(14)

Objective Function T h e objective function of the network reconfiguration problem comprises t w o terms. T h e first term is concerned with loss reduction and the second term is related to load balancing.
Loss reduction This objective is t o minimize the total real power losses

In summary, given a radial distribution system of II nodes, we seek an optimal radial configuration g' (i.e. an optimal spanning tree) among a U possible radial configurations g; by changing the open/closed status of the sectionalizing and tie switches such t h a t both loss [Link] and load balancing are optimized while load constrailits and operational constraints are satisfied. In mathelllatical terms, this problem is expressed as
min
9.

P;" + Q? E T , I2
nb

*=1

IV*

1905
G(r,g,)
N(gO,gv)

5 5

0
nd

(23)
(24)

based on a pre-defined, desired priority. T h e most widely used method of generating non-inferior points is to minimize a non-negative convex combination of the functions f,,i = 1,2,...,m, i.e. minimize a2f,}, where a, 2 0 and a, = 1 . This method suffers from the drawback that it can not generate the entire non-inferior set. Particularly, t,he non-inferior poiot,s whose images are on the non-convex part of the trade-off surface can not be found irrespective to what values a, are used. On the other hand, three methods that can genera t e the entire non-inferior set have been developed: the -constraint method [ll], the weighted minimax method [12] and the shifted minimax function method [13]. Since the -constraint method will be incorporated into a solution methodology to be developed for the network reconfiguration problem, we briefly discuss this method.

T h e above formulation of the network reconfiguration problem is a constrained, multi-objective and nondifferentiable optimization problem. A solution algorithm to solve t h e problem will be presented in a companion paper [lo]. M u l t i - o b j e c t i v e O p t i m i z a t i o n Problems We consider t h e following general multi-objective optimization (MO) problem. min fl(z)

The - C o n s t r a i n t M e t h o d For this method one of the objective functions is chosen

such that
as t h e primary objective function and all other [Link] functions as constraints. To be specific, suppose that f, is the primary objective function, then the method solves the following constrained optimization problem:

A major distinction between a MO problem and a traditional single objective problem is the lack of a complete ordering of t h e fa's, i = 1,2,..., m. In t h e single objective problem, say c(x), a point z* is a global minimum if c(z*) 5 c(z) for all x lying in the region of interest. In the MO problem, a point x such t h a t every component of f simultaneously reach its global minimum is usually nonexistent. This occurs especially when some components of f compete so that one component of f decreases while another increases. In other words, when objectives compete there is n o "optimal solution" t o the MO problem. In this case, the concept of non-inferior (also known as eficiency, Pareto optimality) is used t o characterize the solution t o the MO problem. Definition: T h e feasible region R is the set of state vecl = { x : F(x) = 0, tors x t h a t satisfy the constraints, i.e. f G(x) L 0 1. s a local non-inferior point Definition: A point i E 51 i if there exists a n e > 0 such that in the neighborhood N(?,E) of 5 , there exists no other point x such that (1) f t ( z ) 5 fI(Z), i = 1,2,...,m and (2) f J ( z ) < f J ( ? ) , for some j E {1,2 ,...,m}. In other words, i is a local non-inferior point if there exists a neighborhood N(i.,c) such that for any other point s E N ( i . , c ) , a t least one component o f f will increase its i=1,2, ...,m. value relative t o its value at i or f , ( z ) = f, (i), Definition: A point i E R is a global non-inferior point if i there exists no other point x such t h a t (1) fI(z) 5 fa(?), = 1,2,...,m and (2) f J ( z )< f J ( 5 ) for , some j E {1,2 ,...,m}.
In general there are an infinite number of (global) noninferior points for a given MO problem. T h e collection of such points is called the non-inferior set. We call the image of the non-inferior set trade-ofl (or non-inferior) surface. From a design point of view, a non-inferior point corresponds t o a n optimum trade-off design where attempts to improve any objective will lead to a degradation in at least one of the other objectives. Thus, the availability of the non-inferior set will allow a designer to reach a final design

min f,(z)
t

such that
fJ(z)

se,,

j = 1 , 2 ,..., m ; j # i
iqz) = 0

G(z)

0
E,'S

For different values of , ' S . A systematic variation of will generate the entire non-inferior set.

I t should be pointed out that the computational cost to find the entire non-inferior set is usually very high, if not impossible. Also, current methods (e.g. the above three methods) t o find the entire non-inferior set are based on the presumption that a method which can find the global optimal point for single objective problems is available. From a practical point of view, a more desirable method for the MO problem would be the one that allow the designer to find a desirable, global non-inferior solution in a reasonable computation cost. T o this end, we propose the following two-stage solution methodology. Stage 1: Find a global non-inferior point for the multiobjective optimization problem using the modified simulated annealing technique t o be presented below. (Let the point obtained a t this stage be z * , and c, = f,(z'), j = l , ?,...,m. Also, suppose t h e designer choose fI be the primary objective function). Stage 2: Apply the simulated annealing technique (for single objective function) to the following problem which is obtained by the application of the -constraint method t o (27). min I fz(z) (17) such that
fj(z)

5 CJ

~ C J ,3

= 1,?,...,m;.? #1 F(z) = 0 G(x)

1906

where 6c, designer.

>

0 is a trade-off tolerance made by the

Simulated Annealing An optimization technique based on simulated annealing

for finding a global non-inferior point for an MO problem
(a) Structure of Objective function
c

will be developed in the next section. Therefore, we briefly discuss the simulated annealing in this section. Simulated annealing is a powerful general-purpose technique for solving combinatorial optimization problems. This technique is based on t h e analogy between the simulation and t h e annealing process used for crystallization in physical systems. Annealing is the physical process of heating up a solid to a melting point, followed by cooling it, down until i t crystalizes into a state with a perfect lattice. In this technique, a parameter called temperature is defined, which is of the same dimension as the cost. Just as in a physical system, the temperature is closely related to the freedom with which the entities of the system can move around. T h e system to b e optimized starts a t a high l l the entities of the system move about temperature, and a freely. T h e temperature is then gradually lowered until the l l the entities of the syssystem freezes, a t which point a tem are virtually fixed. This frozen configuration will be close t o t h e lowest energy (or cost) configuration. At each temperature, the Metropolis algorithm derived from statistical physics is employed to simulate the system [14]. A set of moves is selected with which the system can alter from one s t a t e of configuration to another. Moves are chosen randomly. Moves that reduce the system cost (i.e. improve the configuration) are called downhill and moves t h a t increase the system cost are called uphill. In the algorithm, all downhill moves are accepted and u p hill moves are accepted with a probability of e z p ( - % ) , where A C is t h e increase in cost and T is the temperature. More specifically, acceptance for uphill moves is treated probabilistically in the following way: the Boltzman factor e z p ( - q ) is first calculated and a random number r uniformly distributed in the interval [0,1) is then chosen. If r < e z p ( - q ) , the (uphill) move is accepted; otherwise the move is discarded and the configuration before this move is used for the next step. Physically, this means that the system will accept uphill moves with a reasonable probability as long as these moves d o not increase the cost more than T . Thus, at very high temperature the algorithm accepts a l l moves and moves freely in the configuration space (searching broad areas); a t very low temperature the algorithm only accepts downhill moves and behaves like a greedy algorithm (see Fig. 3). I t is due to the probabilistic selection rule t h a t the process can always get out of a local minimum in which i t could get trapped and proceed to the desired global optimum. This feature makes the simulated annealing different from the greedy search approach. T h e simulated annealing at each temperature can be represented in the following pseudo-code repeat
1. move 2 . evaluate AC 3. accept/update until stop criterion = true

(b) High entropy.

I -

Decreasing Temperature

(c) Decreasing entropy.

(d) Low entropy.

Figure 3: Aspects of search p a t t e r n during annealing

process Theoretically, this technique converges asymptotically to the global optimum solution with probability one, provided that certain conditions on the cost function, the moves, and the annealing schedule are satisfied [13]. This technique has been applied to VLSI systems, image processing systems, antenna systems, neural networks and computer system design with successful results [13-151. T h e simulated annealing can b e adapted to constrained optimization problems. One simply has to make sure that all the moves generated from the [Link] continue to satisfy the constraints of the system. This is implemented by adding a feasibility checking function between step 2 (evaluate AC) and step 3 (accept/update) of the pseudo-code. This added function serves to monitor the satisfication of all the equality and inequality constraints. Simulated Annealing for Multi-objective Problems T h e network reconfiguration problem can be casted into the following constrained multi-objective optimization problem (without loss of generality, two objective functions are considered) min CI(ZI,Z~) (28)
21

min
21

C2(~1,~2)

(29)

In this section we present a modified simulat,ed annealing technique to find a global non-inferior point for multiobjective optimization problems T h e key point to find a global non-inferior point using the simulated annealing is the coordination of both [)erturbation scheme and the acceptance criterion. In order to

1907

handle multiple objectives, we propose t o coordinate both perturbation mechanisms and the acceptance criterion in the following way. At each temperature, the following procedure is repeatedly executed.
1. Apply a perturbation mechanism z;
+

z:+'.

2. Compute xi+' from F(z;+',z2) = 0.

3 . Check the feasibility

If G(zf+',zi+')
= z:+'

0,go to 4;

Otherwise go t o 1.
4. Set z;

if one of the following conditions is

satisfied

AC,(z')
min [e-+,

5
AC

0 and ACz(z')
(z')

0;

This error tolerance property can be used to reduce tl1e computation efforts by calculating cost functions approsimately instead of exactly-at least a t high temperature. Another idea of approxiniate cost function calculat,ions is explored in [15] to replace the Boltznian factor by and they report that this scheine may speed I I the solution algorithm by approximately 30 %. The impact of the idea of approximate calculatioiis 011 the network reconfiguration problem can be manifested in the calculation of (18), which itself requires a load flow study. T h e above analysis of error tolerance in cost fiinction therefore suggests that in order t o achieve a fast solution algorithm, an approximate load flow [Link] be used at high temperature and a exact load flow algorithm be used a t low temperature for the calculation of (18).

-9,

exp(-F)
~

AC

( = I )

e-+]

2 random [0,1).

repeat
1. move 2. compute A C

Otherwise, leave it unchanged. Note that t h e search direction of the above procedure does not require that t h e differentiability property b e associated with the objective functions and the constraint functions. Hence, the modified simulated annealing technique is a p plicable t o multi-objective and non-differentiable optimization problems.

approzimotely a t high temperature ezactly a t low temperature 3. accept/update until stop criterion = true

Conclusion
We have formulated the network reconfiguration problem

Computational C o n s i d e r a t i o n s
While the advantages of simulated annealing are its potential to find the global optimal solutions, its general applicability and its flexibility, its main disadvantage is the potentially burdensome amount of computation time required t,o converge to a near-global optimal solution. Admittedly, the amount of computation time required strongly depends on the nature and the size of the problem at hand. T h e situation with respect t o the computational effort becomes worse as problems increase in size. Therefore, i t is important t o investigate possibilities of speeding u p solution algorithms based on simulated annealing in order to keep computation times within desirable limits. In general, speeding up solution algorithms can be achieved by means of the following three approaches:
1. design of a fast (sequential) solution algorithm,

for both loss reduction and load balancing taking into considerations load constraints and operational c0nstraint.s. T h e number of switch-on/switch-off operations involved i n network reconfiguration is also formulated as a constraint. In order to truly reflect the objective of load balancing, we have proposed a new system load balancing index which is a Chebyshev norm of each branch load balancing inclex. T h e new formulation is a constrained multi-objective, 11011differentiable optimization problem with both equality and inequality constraints. We have proposed a two-stage solution methodology for general multi-objective optimization problems. The first stage of the methodology is t o find a global non-inferior point using a modified simulated annealing technique. I n the second stage, we apply the e-constraint method and t,he simulated annealing technique to find an acceptable noiiinferior solution by adjusting the 6,'s according to (1) [Link] global non-inferior point obtained a t the first stage aud ( 2 ) a trade-off tolerance made by the designer. A salient feature of the solution methodology is that it allows designers to find a desirable, global non-inferior solution for the problem. An effective scheme to speed up the simulated annealing technique have been presented and analyzed. Based on the two-stage solution methodology, a comprehensive solution algorithm for the network reconfiguration problem is presented and evaluated in a companion paper [lG]. Acknowledgements This research was partly supported by the NSF under grant ECS-8810544 and grant ECS-8957878. M r . Jean-Jumeau thanks the support from the Organization of American States in Washington D.C.

2. hardware acceleration, and

3 . design of parallel solution algorithms.

Hardware acceleration can be achieved by employing dedicated hardware for evaluating time-consuming parts in the algorithm. Design of a parallel simulated annealing algorithm is a promising approach t o speeding u p the execution of the algorithm. B u t it is by no means a trivial task t.0 design a parallel version of a simulated annealing algorithm mainly due to the intrinsic sequential nature of the algorithm. In this paper we focus on the design of a fast (sequential) solution algorithm. iFrom the acceptance criterion of the simulated annealing, one would intuitively expect t h a t as long as the error i n the cost function calculation is much less than the temperature, the solution algorithm would still work. This is because the probability of accepting an uphill move, which is e z p ( - Y ) , would not change significantly if AC were to 6 << T ) . This shows that be replaced by ( A C f 6 ) , where ( the simulated annealing algorithm can tolerate errors in cost function calculations of t h e order of the temperature.

References
[l] S. Civanlar, J. J. Grainger, and S . H. Lee, "Distribution Feeder Reconfiguration for Loss Reductiou",

1908

IEEE Transactions on Power Dehvery, vol. 3, No. 3, July 1988 , pp. 1217-1223.
A. hferlin and H. Back. Search for a MinimalLoss Operating Spanning Tree Configuration in Urban Power Distribution Systems, Proc. of 5th Power Systems Comp. Con., Cambridge, U.K., Sept. 1-5, 1975. [J] D. Shirmohammadi and H. W. Wong, Reconfiguration of Electric distribution Networks for Resistive Line Losses Reduction, IEEE Transactions on Power Delivery, vol. 4 , No. 2, April 1989, pp.1492-1498. [4] D. W. ROSS, M. Carson, A. Cohen et al., Development of Advanced Methods for Planning Electric Energy Distribution Systems, D E 0 final report no SCI5263, Feb 1980. [5] K. Aoki, H. Kuwabara, T. Satoh, and M. Kanezashi, An Effecient Algorithm for Load Balancing of Transformers and Feeders by Switch Operation in Large Scale Distribution Systems, I E E E P E S Summer Meeting, 1987, paper no: 87SM 543-2. [6] K. Aoki, T. Satoh , M. Itoh e t al., Voltage Drop Constrained Restoration of Supply by Switch Operation in distribution Systems, IEEE PES Summer Meeting, 1987, paper no: 87SM544-0. [7] M. E. Baran, F. F. Wu, Network Reconfiguration in Distribution Systems for Loss Reduction and Load Balancing, IEEE Transactions on Power Delivery, vol. 4, No. 2 , April 1989, pp. 1401-1407. [8] H.D. Chiang and M. E. Baran, On the Existence and Uniqueness of Load Flow Solution for Radial Distribution Power Networks, to appear in IEEE Trans. o n Circuits and Systems, Vol. CAS-37, 1990. [9] M.R. Lightner and S.W. Director, Multiple Criterion Optimization for the Design of Electronic Circuits IEEE Trans. on Circuits and Systems Vol. CAS-28, Mar. 1981, pp. 169-179. [lo] Y.Y. Haimes, W.A. Hall and H.T. Freedman, Multiobjective Optimization in Water Resources Systems, New York: Elsevier Scientific, 1975. [ll] D. Mitra, F. Romeo and A. Sangiovanni-Vincentelli, Convergence and Finite-Time Behavior of Simulated Annealing, Proceeding of 24th Conference on Decision and Control, Dec. 1985, pp. 761-767. I131 U. Faigle and R. Schrader, On t h e Convergence of Stationary Distributions in Simulated Annealing Algorithms, Information Processing Letters, Vol. 27, 1988, pp. 189-194. [13] P.J.M. Laarhoven and E.H.L. Aarts, Simulated Annealing: Theory and Applications, Reidel, Dordrecht, 1987. [14] S. Kirkpatrick, C. D. Gelatt Jr., M. P. Vecchi, Optimization by Simulated Annealing, Science, vol. 220, 1983, pp. 671. [15] D.S. Johnson, C.R. Aragon, L.A. McGeoch and C. Schevon, Optimization by Simulated Annealing: an Experimental Evaluation: P a r t I, AT&T Bell Laboratories, Murray Hill, preprint.

I[

Hsiao-Dong Chiang received the B.S. and M.S. degrees in electrical engineering from National Taiwan University, Taipei, Taiwan. He spent 1982-1983 as a faculty member at Chung-Yuan University i n Taiwan. He received the PI1.D. degree in electrical engineering and computer sciences from the University of California a t Berkeley in 1986, and the11 worked a t the Pacific Gas and Electricity Company 011 a special project. In 1987, he joined the faculty of Cornell University, where he is currently an Assistant Professor of School of Electrical Engineering. In 1988 the National Science Foundation named Chiang a recipient of an Engineering Research Initiation Award. In 1989, Dr. Chiang received a Presidential Young Investigator Award from the National Science Foundation. He is a member of the technical committee on Nonlinear Circuits and Systems, a member of t h e technical committee on Power Systems and Power Electronics and Circuits, IEEE Circuits and Systems Society. His research interests include nonlinear systems, power systems, control systems and optimization theory.

Red Jean-Jumeau studied Electromechanical Engineering a t the State University of Haiti, Port-au-Prince, Haiti, where he obtained his B.S. He was awarded a graduate fellowship by the Organization of American States which enabled him t o attend Cornell University. He is currently in the Ph.D. program in Electrical Engineering at Cornel1 University and studying with Professor Chiang. His research interests involve power systems, optimization theory and nonlinear systems.

[IG! H.D. Chiang and R.M. Jean-Jumeau, Optimal Network Reconfigurations in Distribution Systems: Part 2: A Solution Algorithm and Numerical Results , submitted t o I E E E PES Winter Meeting 1990.

1909

Discussion Ross Baldick, University of California, Berkeley, CA 94720. This paper applies simulated annealing to the network reconfiguration problem, for which there does not seem to be any fast deterministic algorithms. This is a practical approach to an otherwise intractable problem and can potentially integrate many aspects of distribution system control into a single framework. On the other hand, reconfiguration without consideration of, for example, switched capacitor settings may result in unnecessary losses: consider the elementary distribution system shown in figure 1. It consists of a load that can be fed from either of two resistive lines and two switched capacitors connected to the ends of the lines. The losses will be calculated under various switch configurations; the initial configuration is indicated in figure 1. For ease of calculation, the loads are shown as current sinks, although similar results could be found with any sort of load-voltage dependence. Furthermore, assume that load balancing is not a limiting factor.

ing framework just as simply as branch interchange, but would significantly increase the dimension of the problem if there are many switched capacitors in the system. Instead, I suggest that the configuration space of the simulated annealing algorithm be limited to actual branch exchanges and that capacitor optimizaf the losses for each tion be performed as part of the evaluation o configuration: fast algorithms are available to perform the capacitor optimization and evaluate the minimum losses, given a network configuration [A]. In this way, fast algorithms can be used on those parts of the problem that are tractable, while still treating the difficult parts of the problem using the general simulated annealing algorithm presented in this paper.

Reference
[A] Ross Baldick and Felix F. Wu. Efficient Integer Optimization Algorithms for Optimal Coordination of Capacitors and Reg- ulators. Paper 90 WM 175-0 PWRS presented at the IEEE Power Engineering Society 1990 Winter Meeting, Atlanta, Georgia, February 4-8, 1990.
Manuscript received March 5, 1990.

Substation

Figure 1: Initial configuration of system. The initial losses are 2 kW. First suppose that capacitor switching is not included in the reconfiguration optimization: then the only feasible alternative is to close switch Sz and open switch S3, yielding losses of 4 kW. However, a superior solution including capacitor switching is to close switches S I and Sz and open switches S3 and S4, yielding losses of only 1 kW. This example indicates the importance of capacitor coordination in distribution system loss minimization. Capacitor switching is treatable within the simulated anneal-

Hsiao-Dong Chiang and RenC Jean-Jumeau: We would like to thank Mr. Baldick for an interesting discussion. First, it should be pointed out that much of the research in the optimal operation and design of electric power systems consists of two stages of efforts: (1) problem formulation and (2) development of solution algorithms. These two are inter-related. It is quite common to approximate problem formulations to make the problems under study solvable with available optimization techniques. This can be exemplified by the work on the optimal power flow problem as well as some of the work on distribution systems. The generality of the simulated annealing technique makes it possible to decouple these two inter-related process. The technique allows us in the process of formulating problems to adhere to normal principles of power system operation so that the problem is expressed in a realistic way. In this paper we have presented a new formulation of the network reconfiguration problem for both loss reduction and load balancing with load and operational constraints. The new formulation is a multi-objective, non-differentiable optimization problem with equality and inequality constraints. Second, it is strongly believed that multiple local optimal solutions are likely to exist in many power system problems. A solution algorithm based on simulated annealing can achieve global optimal solutions rather than just local optimal solutions. In practical implementation, the solution algorithm starts with any initial condition and converge to a near-global optimal solution. In this paper we have developed a two-stage solution methodology based on a modified simulated annealing technique for general multiobjective optimization problems and applied it to the network reconfiguration problem. It is always desirable to be able to optimally coordinate all the available devices, such as network reconfiguration, capacitor switching, voltage regulators adjustment, to reduce real power losses in distribution systems. Mr. Baldicks suggestion of using the simulated annealing technique to determine network reconfigurations and using fast algorithms developed by Baldick and Wu in [A] to determine capacitor switchings appears attractive. Further assessments of the effectiveness of both this suggestion and the scheme that uses the simulated annealing technique to determine both network reconfigurations and capacitor switchings on practical distribution systems may prove rewarding.
Manuscript received April 6, 1990.