Electric Power Systems Research 62 (2002) 37
/42

www.elsevier.com/locate/epsr

Optimal reconfiguration of electrical distribution network using the

refined genetic algorithm
J.Z. Zhu
Alstom ESCA Corporation, 11120 NE 33rd Place, Bellevue, WA 98004, USA

Received 25 June 2001; received in revised form 17 January 2002; accepted 18 January 2002

Abstract

This paper proposes an improved method to study distribution network reconfiguration (DNRC) based on a refined genetic
algorithm (GA). The DNRC model, in which the objective is to minimize the system power loss, is set up. In order to get the precise
branch current and system power loss, a radiation distribution network load flow (RDNLF) method is presented in the study. The
refined genetic algorithm is also set up, in which some improvements are made on chromosome coding, fitness function and
mutation pattern. As a result, premature convergence is avoided. The proposed approach is tested on 16-bus and 33-bus distribution
networks. Study results are given in the paper. # 2002 Elsevier Science B.V. All rights reserved.

Keywords: Optimization; Artificial intelligence; Genetic algorithms; Distribution systems; Network reconfiguration

1. Introduction switches), and the total number of trees is extremely

large. Ordinary optimization methods have shown to be
It is known that distribution networks are built as ineffective and impractical to this problem [1].
interconnected meshed networks, while in the operation Merlin et al. proposed a heuristic approach to the
they are arranged into a radial tree structure. This DNRC problem [2]. It begins with all branches closed (a
means that distribution systems are divided into sub- complete graph) and performs a procedure of opening
systems of radial feeders, which contain a number of branches which carry the least current. This method is a
normally-closed switches and a number of normally- greedy algorithm that does not necessarily guarantee
open switches. From graph theory, a distribution net- feasibility of the final solution.
work can be represented with a graph of G (N , B ) that Nahman et al. presented another heuristic approach
contains a set of nodes N and a set of branches B . Every in [3,4]. The algorithm starts from a completely empty
node represents either a source node (supply transfor- network, with all switches open and all loads discon-
mer) or a sink node (customer load point), while a nected. Load points are connected one by one by
branch represents a feeder section that can either be switching branches onto the current subtree. The search
loaded (switch closed) or unloaded (switch open). The technique also does not necessarily guarantee global
network is radial, so that feeder sections form a set of optima.
trees where each sink node is supplied from exactly one Recently, the branch exchange approach has been
source node. Therefore, the distribution network recon- used in research on DNRC [5
/7]. In fact, this is a
figuration (DNRC) problem is to find a radial operating gradient method in the space of a graph structure. In
structure that minimizes the system power loss while
this method, one normally-open switch is closed, which
satisfying operating constraints. In fact, this problem
forms a loop and violates the topological constraints.
can be viewed as a problem of determining an optimal
When a closed branch (or switch) in the loop opens, a
tree of the given graph. However, real distribution
new topology is produced. However, the existing branch
systems contain many nodes and branches (and
exchange based algorithms are capable of finding only
local optima, where the final solution heavily depends
E-mail address: [email protected] (J.Z. Zhu). on the starting configuration.
0378-7796/02/$ - see front matter # 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 8 - 7 7 9 6 ( 0 2 ) 0 0 0 4 1 - X
38 J.Z. Zhu / Electric Power Systems Research 62 (2002) 37
/42

At present, new methods based on artificial intelli- 1) GAs work with coding of parameters rather than
gence have been used in DNRC [8,9]. Chiang et al. [8] the parameters themselves.
presented a simulated annealing (SA) method to solve 2) GAs search from a population of points rather than
the DNRC problem, in which the SA was very time- a single point.
consuming. It is needed to apply the improved SA with 3) GAs use only objective functions rather than
high speed to handle the DNRC problem. For the first additional information such as their derivatives.
time, genetic algorithm (GA) was applied to the global 4) GAs use probabilistic transition rules, and not
optimal solution of DNRC in [9], which has shown a deterministic rules.
better performance over the SA approach. In this paper,
These properties make GAs more robust, more
the GA method is further refined by modifying the
powerful and less data-independent than many other
string structure and fitness function:
conventional techniques.
1) In Ref. [9], the string used in GA describes all the The theoretical foundation for GAs was first de-
switch positions and their ‘on/off’ states. The string scribed by Holland [11], and was presented tutorially by
can be very long and it grows in proportion with the Goldberg [12]. GAs provide a solution to a problem by
number of switches. For large distribution systems, working with a population of individuals each repre-
such long strings can not be effectively searched by senting a possible solution. Each possible solution is
GA. In this paper, the string will be shortened. termed a ‘chromosome’. New points of the search space
2) To reduce computational burden, approximate are generated through GA operations, known as repro-
fitness functions was used in GA to represent the duction, crossover and mutation. These operations
system power loss [9]. It may affect the accuracy and consistently produce fitter offsprings through successive
effectiveness of GA. GA is essentially unconstrained generations, which rapidly lead the search towards
search procedures within the given represented global optima [12].
space [10
/12]. All information should be fully
represented in the fitness function. Over-approxi-
mated fitness function would lead directly to 3. Mathematical model of DNRC
unreliable solution.
The purpose of DNRC is to find a radial operating
In this paper, the DNRC model, in which the structure that minimizes the system power loss while
objective is to minimize the system power loss, is set satisfying operating constraints. Thus, the following
up. Since the distribution network is a simple radial tree model can represent the DNRC problem.
structure, in which the ratio of R /X is relatively big, Min f Sb ½Ib ½ 2 kb Rb b
NL (1)
even bigger than 1.0 for some transmission lines, neither
P
/Q decoupled method nor Newton
/Raphson method such that
is suited to compute the distribution network loadflow. kb ½Ib ½ 5Ibmax b
NL (2)
Therefore, a radiation distribution network loadflow
Vimin 5Vi 5Vimax i
N (3)
(RDNLF) method is presented in the study. In order to
enhance performance of GA, some improvements are gi (I; k)0 (4)
made on chromosome coding, fitness function and gv (V ; k)0 (5)
mutation pattern. Among these improved features, an 8 (k)0 (6)
adaptive process of mutation is developed not only to
prevent premature convergence, but also to produce where Ib is the current in branch b; Rb is the resistance
smooth convergence. The proposed approach is tested of branch b. Vi is the node voltage at node i. kb
with satisfactory results on 16-bus and 33-bus distribu- represents the topological status of the branches. kb /1
tion networks. if the branch b is closed, and kb /0 if the branch b is
open. N is the set of nodes, and NL is the set of
branches. Subscripts ‘min’ and ‘max’ represent the lower
and upper bounds of the constraint.
In the above model, Eq. (2) stands for the branch
2. Brief description of GA current thermal stability constraints. Eq. (3) stands for
the node voltage constraints. Eqs. (4) and (5) represent
GAs are effective parameter search techniques. They Kirchhoff’s current and voltage laws. Eq. (6) stands for
are considered when conventional techniques have not topological constraints which ensure radial structure of
achieved the desired speed, accuracy or efficiency [10]. each candidate topology.
GAs are different from conventional optimization and In order to get the precise expression of system power
search procedures in the following ways [12]. loss, the current branch Ib , will be computed through a
 J.Z. Zhu / Electric Power Systems Research 62 (2002) 37
/42 39

Fig. 1. A simple distribution network. Source transformer busbars */; closed switches
/; open switches ---; sink nodes (load nodes) /.

radiation distribution network loadflow (RDNLF) networks, it is not efficient to represent every arc in the
method in the study. It is well known that in the string, since its length will be very long. In fact, the
distribution network, the ratio of R /X (resistance/ number of open switch positions is identical to keep the
reactance) is relatively big, even bigger than 1.0 for system radial once the topology of the distribution
some transmission lines. In this case, P
/Q decoupled networks is fixed, even if the open switch positions are
loadflow is invalid for distribution network load flow changed. Therefore, to memorize the radial configura-
calculation. It will also be complicated and time- tion, it is enough to number only the open switch
consuming to use the Newton
/Raphson loadflow positions. Fig. 1 shows a simple distribution network
because the distribution network is only a simple radial with five switches that are normally open.
tree structure. Therefore, RDNLF method is presented In Fig. 1(a), positions of five initially-open switches 5,
in the paper. RDNLF calculation consists of two parts. 8, 10, 13 and 14 determine a radial topology. In Fig.
One is calculation of branch current from the ‘top of a 1(b), positions of five initially-open switches 1, 4, 7, 9
tree’ node to the ‘root of a tree’ node. Another is the and 10 determine another radial topology. Therefore, in
calculation of node voltage from the ‘root of a tree’ node order to represent a network topology, only positions of
to the ‘top of a tree node. The initial conditions are the the open switches in the distribution network need to be
given voltage vectors at root nodes as well as real and known. Suppose the number of normally open switches
reactive power at load nodes. In final, the deviation of is No, the length of a genetic string depends on the
injection power at all nodes can be computed. The number of open switches No. Genetic strings for Fig.
iteration calculation will be ceased if the deviation is less 1(a) and (b) are represented as follows, respectively.
than the given permissive error.

0101 1000 1010 1101 1110

4. Refined GA approach to DNRC problem Switch 5 Switch 8 Switch 10 Switch 13 Switch 14

Model M-1 may be solved by first generating all Genetic string for Fig. 1(a).
graph trees and subsequently by performing evaluation.
However, real distribution systems contain many nodes
0001 0100 0111 1001 1010
and branches, and many trees. Conventional optimiza-
tion methods have shown to be ineffective and imprac- Switch 1 Switch 4 Switch 7 Switch 9 Switch 10
tical, because of dimensionality [1]. GAs has shown to
be an effective and useful approach for the DNRC Genetic string for Fig. 1(b).
problem [9]. Some refinements of the approach are
described in this paper. 4.2. Fitness function

4.1. Genetic string GAs are essentially unconstrained search procedures

within a given represented space. Therefore, it is very
In Ref. [9], the string structure is expressed by ‘Arc important to construct an accurate fitness function as its
No.(i)’ and ‘SW. No.(i)’ for each switch i. ‘Arc No.(i)’ value is the only information available to guide the
identifies the arc (branch) number that contains the i-th search. In this paper, the fitness function is formed by
open switch, and ‘SW. No.(i)’ identifies the switch that combining the object function and the penalty function,
is normally open on Arc No.(i ). For large distribution i.e.
40 J.Z. Zhu / Electric Power Systems Research 62 (2002) 37
/42

Max f 1=L (7)

where
L Si ½Ii ½ 2 ki Ri b1 maxf0; (½Ii ½Iimax )2 g
b2 maxf0; (Vimin Vi )2 g
b3 maxf0; (Vi Vimax )2 g (8)
where bi (i/1, 2, 3) is a large constant.
Suppose m is the population size, the values of the
maximum fitness, the minimum fitness, sum of fitness Fig. 2. A 16-bus distribution system.
and average fitness of a generation are calculated as
follows. The mutation scale will decrease as the process
continues. The minimum mutation probability in this
fmax  ffi ½fi ]fjfj ; j 1; : : :; mg (9)
study is given as 0.05. This adaptive mutation not only
fmin ffi ½fi 5fjfj ; j 1; : : :; mg (10) prevents premature convergence, but also leads to a
fS Si fi ; i 1; : : :; m (11) smooth convergence.

fav fS =m (12)

The strings are sorted according to their fitness which 5. Numerical examples
are then ranked accordingly.
The proposed approach for distribution network
4.3. Selection reconfiguration is tested on 16-bus and 33-bus distribu-
tion systems as shown in Figs. 2 and 3, respectively.
In order to obtain and maintain good performance of System data and parameters are listed on Tables 1 and 2.
the fittest individuals, it is important to keep the The 16-bus test system contains 3 source transformers
selection competitive enough. It is no doubt that the and 13 load nodes. The three initially-open switches are
fittest individuals have higher chances to be selected. In ‘4’, ‘11’ and ‘13’. The total system load is 23.7 MW,
this paper, the ‘roulette wheel selection’ scheme is used, while the initial system power loss is 0.5114 MW. The
in which each string occupies an area of the wheel that is 33-bus test system consists of one source transformer
equal to the string’s share of the total fitness, i.e. fi /fS.

4.4. Crossover and mutation

Crossover takes random pairs from the mating pool

and produces two new strings, each being made of one
part of the parent string. Mutation provides a way to
introduce new information into the knowledge base.
With this operator, individual genetic representations
are changed according to some probabilistic rules. In
general, the GA mutation probability is fixed through-
out the whole search process. However, in practical
application of DNRC, a small fixed mutation prob-
ability can only result in a premature convergence. In
this paper, an adaptive mutation process is used to
change the mutation probability, i.e.
p(k)pstep if fmin (k) unchanged
p(k 1) p(k) if fmin (k) decreased
pfinal if p(k)pstep Bpfinal
p(0)pinit 1:0
pstep 0:001
pfinal 0:05
where k is the generation number; and p is the mutation
probability. Fig. 3. A 33-bus distribution system.
 J.Z. Zhu / Electric Power Systems Research 62 (2002) 37
/42 41

Table 1
System data and parameters for 16-bus distribution network

Line no. Node i Node j Resistance Reactance Receiving node j Receiving node j

R (V) X (V) P (MW) Q (MVAr) Voltage (p.u.)

1 1 4 0.0750 0.1000 2.0 1.6 0.9907//0.3968

3 4 5 0.0800 0.1100 3.0 0.4 0.9878//0.5443
2 4 6 0.0900 0.1800 2.0 /0.4 0.9860//0.6972
5 6 7 0.0400 0.0400 1.5 1.2 0.9849//0.7043
7 2 8 0.1100 0.1100 4.0 2.7 0.9791//0.7635
8 8 9 0.0800 0.1100 5.0 1.8 0.9711//1.452
9 8 10 0.1100 0.1100 1.0 0.9 0.9769//0.7701
6 9 11 0.1100 0.1100 0.6 /0.5 0.9710//1.526
10 9 12 0.0800 0.1100 4.5 /1.7 0.9693//1.837
15 3 13 0.1100 0.1100 1.0 0.9 0.9944//0.3293
14 13 14 0.0900 0.1200 1.0 /1.1 0.9948//0.4562
16 13 15 0.0800 0.1100 1.0 0.9 0.9918//0.5228
12 15 16 0.0400 0.0400 2.1 /0.8 0.9913//0.5904
4 5 11 0.0400 0.0400
13 10 14 0.0400 0.0400
11 7 16 0.0900 0.1200

Table 2
System data and parameters for 33-bus distribution network

Line no. Node i Node j Resistance Reactance Receiving node Receiving node

R (V) X (V) P (MW) Q (MVAr) Voltage (p.u.)

1 1 2 0.0922 0.0470 100.0 60.0 0.9970/0.0145

2 2 3 0.4930 0.2512 90.0 40.0 0.9829/0.0960
3 3 4 0.3661 0.1864 120.0 80.0 0.9755/0.1617
4 4 5 0.3811 0.1941 60.0 30.0 0.9681/0.2283
5 5 6 0.8190 0.7070 60.0 20.0 0.9497/0.1339
6 6 7 0.1872 0.6188 200.0 100.0 0.9462//0.0964
7 7 8 0.7115 0.2351 200.0 100.0 0.9413//0.0603
8 8 9 1.0299 0.7400 60.0 20.0 0.9351//0.1334
9 9 10 1.0440 0.7400 60.0 20.0 0.9292//0.1959
10 10 11 0.1967 0.0651 45.0 30.0 0.9284//0.1887
11 11 12 0.3744 0.1298 60.0 35.0 0.9269//0.1785
12 12 13 1.4680 1.1549 60.0 35.0 0.9208//0.2698
13 13 14 0.5416 0.7129 120.0 80.0 0.9185//0.3485
14 14 15 0.5909 0.5260 60.0 10.0 0.9171//0.3862
15 15 16 0.7462 0.5449 60.0 20.0 0.9157//0.4094
16 16 17 1.2889 1.7210 60.0 20.0 0.9137//0.4868
17 17 18 0.7320 0.5739 90.0 40.0 0.9131//0.4963
18 2 19 0.1640 0.1565 90.0 40.0 0.9965//0.0037
19 19 20 1.5042 1.3555 90.0 40.0 0.9929//0.0633
20 20 21 0.4095 0.4784 90.0 40.0 0.9922//0.0827
21 21 22 0.7089 0.9373 90.0 40.0 0.9916//0.1030
22 3 23 0.4512 0.3084 90.0 50.0 0.9794//0.0650
23 23 24 0.8980 0.7091 420.0 200.0 0.9727//0.0237
24 24 25 0.8959 0.7071 420.0 200.0 0.9694//0.0674
25 6 26 0.2031 0.1034 60.0 25.0 0.9477/0.1734
26 26 27 0.2842 0.1447 60.0 25.0 0.9452/0.2295
27 27 28 1.0589 0.9338 60.0 20.0 0.9337/0.3124
28 28 29 0.8043 0.7006 120.0 70.0 0.9255/0.3904
29 29 30 0.5074 0.2585 200.0 100.0 0.9219/0.4956
30 30 31 0.9745 0.9629 150.0 70.0 0.9178/0.4112
31 31 32 0.3105 0.3619 210.0 100.0 0.9169/0.3882
32 32 33 0.3411 0.5302 60.0 40.0 0.9166/0.3805
34 8 21 2.0000 2.0000
36 9 15 2.0000 2.0000
35 12 22 2.0000 2.0000
37 18 33 0.5000 0.5000
33 25 29 0.5000 0.5000
42 J.Z. Zhu / Electric Power Systems Research 62 (2002) 37
/42

Table 3 shortened to minimize the required memories and to

DNRC results for 16-bus test system ensure search efficiency. The proposed process of
Radial network Initial network Proposed method adaptive mutation not only prevents premature conver-
gence, but also leads to a smooth convergence. From
Open switches Switch 4 Switch 6 several case studies and comparison with other methods,
Switch 11 Switch 9
Switch 13 Switch 11
it can be concluded that the global optima have been
Power loss (MW) 0.5114 0.4661 found by the proposed algorithm. The validity and
effectiveness of the proposed methodology have also
been demonstrated.
Table 4
Comparison of DNRC results for 33-bus test system

Radial network Initial Method in Proposed References

network Ref. [7] method

Open switches Switch 33 Switch 7 Switch 7 [1] J.M. Wojciechowski, An approach formula for counting trees in a
Switch 34 Switch 10 Switch 9 graph, IEEE Trans. Circuit Syst. 32 (4) (1985) 382
/385.
Switch 35 Switch 14 Switch 14 [2] A. Merlin, H. Back, Search for minimum-loss operating spanning
Switch 36 Switch 33 Switch 32 tree configuration in an urban power distribution system, Proc.
Switch 37 Switch 37 Switch 33 5th Power System Computation Conference, Cambridge, 1975,
Power loss (MW) 0.202674 0.141541 0.139532 Paper 1.2/6.
[3] J. Nahman, G. Strbac, A new algorithm for service restoration in
large-scale urban distribution systems, Elect. Power Syst. Res. 29
(1994) 181
/192.
and 32 load points. The five initially-open switches are [4] V. Glamocanin, Optimal loss reduction of distribution networks,
‘33’, ‘34’, ‘35’, ‘36’ and ‘37’. The total system load is IEEE Trans. Power Syst. 5 (3) (1990) 774
/782.
3.715 MW, while the initial system power loss is [5] D.W. Ross, J. Patton, A.I. Cohen, M. Carson, New methods for
0.202674 MW. The system base is V /12.66 kV and evaluating distribution automation and control system benefits,
IEEE Trans. PAS 100 (1981) 2978
/2986.
S /10 MVA. [6] G. Strbac, J. Nahman, Reliability aspects in structuring of large
Results on the two study systems are listed in Tables 3 scale urban distribution systems, IEE Conf. on Reliability of
and 4. The test results are also compared with those in Transmission and Distribution Equipment, March 1995, pp. 151
/
Ref. [7], in which the solution was stated global optima. 156.
[7] D. Shirmohammadi, H.W. Hong, Reconfiguration of electric
It can be observed that the results in this paper are even
distribution networks for resistive line losses reduction, IEEE
better than those in Ref. [7]. Thus, we can say that the Trans. PWRD 4 (2) (1989) 1492
/1498.
global optima have been found in this paper. [8] H.D. Chiang, R.J. Jumeau, Optimal network reconfiguration in
distribution systems, part 1: a new formulation and a solution
methodology, IEEE Trans. Power Deliv. 5 (4) (1990) 1902
/1909.
6. Conclusions [9] K. Nara, A. Shiose, M. Kitagawa, T. Ishihara, Implementation of
genetic algorithm for distribution system loss minimum reconfi-
guration, IEEE Trans. Power Systems 7 (3) (1992) 1044
/1051.
An improved method to study distribution network [10] J.Z. Zhu, C.S. Chang, G.Y. Xu, X.F. Xiong, Optimal load
reconfiguration using GA is presented in the paper. The frequency control using genetic algorithm, Proc. 1996 Int. Conf.
DNRC model, in which the objective is to minimize the Elect. Eng., ICEE’96, Beijing, China, August 12
/15, 1996, pp.
distribution system loss, is formed. In the application of 1103
/1107.
[11] J.H. Holland, Adaptation in Nature and Artificial Systems, The
GA to the DNRC, some improvements of algorithms University of Michigan Press, Michigan 1975.
are made on chromosome coding, fitness function and [12] D.E. Goldberg, Genetic Algorithms in Search, Optimization and
mutation pattern. The genetic string used in the paper is Machine Learning, Addision-Wesley, Reading 1989.