1

Distribution network reconfiguration using an

efficient evolutionary algorithm
E. M. Carreno, Student Member, IEEE, N. Moreira, and R. Romero, Member, IEEE

task, the network structure of distribution feeders is modified

Abstract-- Network reconfiguration is an important tool to changing the open/close status of the sectionalizing (normally
optimize the operating conditions of a distribution system. This is closed) and tie (normally open) switches, attempting to reduce
accomplished modifying the network structure of distribution the general losses of the system relieving the overloading of
feeders by changing the open/close status of sectionalizing
the network components, among other objectives.
switches. This not only reduces the power losses, but also relieves
the overloading of the network components. Network Network reconfiguration for loss reduction is a highly
reconfiguration belongs to a complex family of problems because complex combinatorial, non-differentiable and constrained
of their combinatorial nature and multiple constraints. This optimization problem, due to the high number of switching
paper proposes a solution to this problem, using a specialized elements in a distribution network, and also for the non-linear
evolutionary algorithm, with a novel codification, and a brand characteristics of the constraints used to model the electrical
new way of implement the genetic operators considering the
behavior of the system.
problem characteristics. The algorithm is presented and tested in
a real distribution system, showing excellent results and Different approaches have been used to solve the problem,
computational efficiency. first, there are techniques that used the operators experience
maneuvering the network, with these knowledge were
Index Terms—distribution network reconfiguration, loss developed several algorithms who try to obtain near optimal
reduction, metaheuristics, genetic algorithms, network solutions in large systems, using few computational effort. In
representation. this category are found heuristics like in [1-4].
In another category, techniques using specialist agents and
I. NOMENCLATURE
artificial intelligence have been applied to the problem, like
fc : Loss function
the ones found in [5-7]
NL : Set of branches Finally, there are solutions proposed using evolutionary
Ii : Current in branch i computing, like Simulated Annealing [8], Genetic Algorithms
Ri : Resistance of branch i [9-11], amongst other works in related areas.
ki : Binary variable that represents the topological status of the The important contribution of this work, is the novel
branches representation of the network structure, which is based on tree
Vj : Voltage magnitude at node j representation for network analyses, thus, the network is
V j min : Minimum voltage accepted at node j represented with a vector containing arcs, organized in a way
V j max : Maximum voltage accepted at node j the tree is not connected, without closing loops (radial
configuration).
Ii max : Maximum current accepted at branch i
This representation, allows proposing different ways to
nb : Number of buses in the network
implement the genetic operators, allowing the algorithm to
nl : Number of branches (arcs) in the network
take advantage of the problem characteristics to improve in
efficiency, narrow the search space, and speed up the process.
II. INTRODUCTION
The combination of this representation with an efficient

B ecause of the need of more efficient owned utilities due

to recent advances in regulation for distribution systems,
and fascinating advances in automation, automatic network
evolutionary algorithm, like the Chu Beasley Genetic
Algorithm [12], yield as result an efficient and very fast tool
to network reconfiguration, obtaining optimal or quasi optimal
management is becoming more common in utilities around the solutions for very large systems in seconds with normal
world. One of the tasks involved is this process is the network desktop computers.
reconfiguration of distribution feeders. To accomplish this This paper is organized as follow: Section 3 defines the
network reconfiguration mathematical problem, section 4
This work was supported by Conselho Nacional da Pesquisa do Brasil gives an overview of the specialized evolutionary algorithm
(CNPQ) (Process 141281/2004-9). used, section 5 introduces the problem codification, section 6
The authors are with Universidade Estadual Paulista (UNESP) Ilha Solteira,
SP, Brazil (e-mail: [email protected], [email protected], is devoted to explain the genetic operators proposed, section 7
[email protected]) (web: http://www.dee.feis.unesp.br/lapsee/). presents the results to the test systems, and finally section 8

1-4244-1298-6/07/$25.00 ©2007 IEEE.

 2

presents the final remarks of this work.

The Chu Beasley genetic algorithm can be resumed as:
III. MATHEMATICAL MODEL FOR THE DISTRIBUTION
NETWORK RECONFIGURATION PROBLEM Step 1: Generate Initial Population
The purpose of distribution network reconfiguration is to Step 2: Select 2 parents using tournament selection.
find a radial operating structure that minimizes the system Step 3: Apply genetic operators depending on the problem
power losses while satisfying operating constraints. Thus the (Recombination, Mutation, Shift, etc)
problem can be formulated as follows: Step 4: Try to improve the resulting individual with
information about the nature of the problem

∑I
n
2 (1) Step 5: Include the individual in the population, only if it
min f c = i ki Ri i ∈ NL has a better fitness function than the worst element, and it is
i =1
not already present in it.
Step 6: Repeat from step 2, until the stop criterion is meet.
Subject to:
1. Radial network constraint: V. PROBLEM CODIFICATION
Distribution network should be composed of radial
structure considering operational point of view. The general approach when the network reconfiguration
problem is coded within heuristics algorithm, is to code only
ϕ (k ) = 0 (2)
the switching elements with binary variables, the network is
2. Node voltage constraint: not represented, thus, when different operators in different
algorithms are applied, the resultant topologies generally does
Voltage magnitude Vi at each node must lie within their
not represent a topology in the neighborhood of the
permissible ranges to maintain power quality. generation point.
Vi min ≤ Vi ≤ Vi max (3) For example, if a network is represented by n binary
elements (1,0,...,0), the first element represents a switching
3. Branch current thermal stability constraints: element in feeder 1, the second element represent a switching
Current magnitude of each branch (feeder, laterals and device in feeder k. In any heuristic algorithm, going from
switches) must lie within their permissible ranges. (1,0,…,0) to (0,1,…,0) represents going to a neighborhood
k I ≤I
i i i max
(4) configuration, but in reality, those two configurations are
completely different when viewed topologically.
4. Kirchoff’s current and voltage laws To avoid these situations, another codification is proposed:
gi ( I , k ) = 0 (4) Instead of representing the switching devices, the entire
network configuration resulting from the switching is used as
g v (V , k ) = 0 (5) the individual and considered as a tree graph, represented as a
vector with the arcs in the tree (branches) sequentially
organized, from top to bottom of the network, being the root
IV. CHU BEASLEY GENETIC ALGORITHM node (generally the substation) the top.
The evolutionary algorithm used in this work was first The codification vector stores the arcs (branches) from the
proposed and implemented for Chu Beasley [12] applied to system of nb buses and nl arcs. The first (nb-1) elements are
the Generalized Assignment Problem (GAP). This algorithm the branches of the radial topology (set N1), and the branches
has some very special characteristics and presents excellent between positions nb and nl (set N2) are the connecting
performance. One of the main characteristics is the non- elements (branches outside the radial configuration), see Fig.
generational substitution of the current population, which 1.
means, that in each step of the algorithm only one element of
the current population is replaced.
The substitution of the population is implemented in a way
that two strategies could be efficiently implemented: (1) the
descendent generated can be improved using a local
optimization, which could reduce (or eliminate) the Fig. 1. Codification vector.
infeasibility, and (2) the generated descendent is included in
the population only if it is different and of better quality than In order to represent a radial configuration, the arc in the
any of the individuals in it. This strategy is also more efficient position i must be connected to any of the arcs located
than elitism because the best solutions are always preserved in between position 1 and i-1, and can not form a loop with any
the current population. of these either.
This algorithm has been applied to different problems, An example with the 14 bus test system is shown in Fig. 2,
with good results compared to traditional genetic algorithms. this configuration can be represented as
 3

[1,5,10,2,7,11,3,6,12,4,8,13,9,(14,15,16)], where all the configurations of better quality.

elements between 1 and (nb-1) belong to the network The Objective Function (OF), which is active losses, is
topology (continuous gross line), and the others, between nb calculated with a forward-backward sweep power flow (SPF)
until the end of the vector are the connecting elements out of [14] for each member of the population.
the configuration (dotted thin line).
This representation is not fixed, the same topology can be VI. GENETIC OPERATORS
represented as [1,2,3,4,5,7,6,8,9,10,11,12,13,(14,15,16)], Due to the nature of the proposed codification, the genetics
which can help to diversify the search depending of the type operators can be defined in a different way, the algorithm
of operator employed. proposed only need one operator to be extremely efficient, but
more operators could be developed. The operator used is the
recombination.

A. Recombination:
In traditional genetic algorithms, two individual are
selected and their chromosomes are mixed in some way to
generate one or more individuals. The same concept is applied
with this codification in the next way.
1. Select a recombination point, nr, randomly between 40%
and 80% of the length of the N1 set of the main
generator, so the descendent does not seems almost equal
to one of their parents.
2. Copy all the first nr elements from the main generator to
Fig 2. 14 Bus Test System Configuration. the forming descendent.
3. In the secondary generator, eliminate all the elements
This codification proposal guarantees that if the initial already present in the descendent.
population is composed of radial configurations, then it is 4. For each arc remaining in the secondary generator: If the
possible to develop recombination and mutation operators that arc does not form a loop, include it in N1, otherwise,
preserve the radial topology of the descendant generated. This include it in N2.
codification proposal and the way the recombination is This algorithm always generates radial topologies with some
implemented are one of the main contributions of this paper, arcs of the main generator, some arcs of the secondary
because eliminate the generation of non-radial topologies and generator, and eventually, connecting arcs from one, or both
allow generating descendents in a similar way to the natural topologies.
genetic recombination.
The mutation operator is not used because in the
A. Generation of the Initial Population reconfiguration problem, the most effective way to implement
The next algorithm is used to generate the initial mutation is to swap one connecting arc with an arc inside the
population, using the idea proposed by Prim [13] to generate loop generated if this connecting arc was introduced into the
trees with minimum cost. topology. The local improvement step proposed is a
Being D the set of nodes added to the actual configuration. generalization of this type of mutation, thus, this step is
1. D = φ , N1 = φ , N 2 = φ skipped.
2. Assign the substation node (root node) to D.
3. Search for all the arcs with a vertex in D.
B. Local improvement:
4. Select one arc to enter in the topology (randomly or with
some methodology). After the recombination, the resulting network is analyzed in
5. If the selected arc generates a loop in the topology, the next way:
include it in N2, otherwise, include the opposite vertex in
D and the arc in N1. 1. Identify all the loops formed in the network if the
6. Repeat from 3 until all the arcs are selected. connecting arcs are included.
2. For each path from the connecting arcs to the substation
At the end of the algorithm, (N1,N2) represents the node: Swap the connecting arc with the next arc in the
proposed topology, this algorithm always generate radial sequence, if the losses are inferior to the base topology,
configurations. store the topology and test the next element , otherwise,
go to the next path.
This process can be upgraded to look for initial solutions
with some intelligence instead for complete randomness, like
This procedure is made from bottom to top, in this way, if
assign a bigger probability to the best candidate, or beginning
the OF calculated in a swap operation is worst than the
the tree with part of the original network, which could give
 4

current OF, the rest of the path in analyses can be discarded,

because if the operation continues, the losses are going to be
greater, this fact reduces greatly the computational effort of
this step.

VII. TEST AND RESULTS

A. 14 bus test system:

The 14 bus system data can be found in [1].
This system, in his initial condition, has active losses of
511.43 kW, for this particular system, due to his size, the
algorithm find the optimal solution in just two or three
iterations with a population of 10 individuals, in less than a
second.
Fig. 4. Descendent generated.
The system is used here to demonstrate the use of the
genetic operators proposed 2) Local Improvement
1) Recombination In the topology obtained in the previous step, all the loops
Take for example two configurations, randomly generated, corresponding to include branches 7, 14, and 16 are
and picked from a tournament selection, as the selected identified:
topologies to be combined, the main generator is depicted in Loop 1: [7, 15, 11, 10, 5]: 4 neighbor topologies
Fig. 3, with 492.81 kW of losses. Loop 2: [14, 2, 1, 5, 6, 8]: 5 neighbor topologies
Loop 3: [16, 4, 3, 1, 10, 12, 13]: 6 neighbor topologies
To examine all the possibilities, a maximum of 15 power
flows would be needed.
In the first loop, arcs 7 and 15 are swapped, obtaining a
configuration with 511,43 kW, the losses are worst than the
initial losses of 483,87 kW, this mean that it is not necessary
to examine the rest of the elements that belong to the path
between 7-15 and the substation node, sparing 2 power flows.
Only to complement this information, doing a swap
between arcs 7 and 11 produces a topology with 532.86 kW of
losses, and doing a swap between arcs 7 and 10 produces a
topology with 772.63 kW of losses.
All results from the swap operations are presented in Table
Fig. 3. Main topology to be combined. 1, where all the power flows that are not necessary to
calculate inside the program are marked as NN.
The topology from Fig. 2, is selected as the secondary
generator with 511.43kW of losses. TABLE I
RESULTS FROM SWAP OPERATIONS
In this example the recombination point is randomly
selected as position eight.
In this way, the first part of the new set N1 is, [1, 10, 2, 3,
5, 11, 15, 4].
The elements of the secondary topology are analyzed
sequentially, and those elements that could be in the tree,
without forming loops, are included in N1, otherwise, the
elements are included in N2, the resulting vector N1 is: [1, 10,
2, 3, 5, 11, 15, 4, 12, 6, 13, 8, 9] and N2 is [7, 14, 16].
The resulting network, with 483.87 kW of losses, is
visualized in Fig. 4, where the continuous gross line
represents branches from the main topology, dashed gross
lines belongs to the secondary topology, and the dotted thin
lines are branches out of the configuration (N2).
After 7 SPF in this step, the swap between arcs 8 and 14
produces a better OF. The new topology, with 466.10 kW of
 5

losses, which is the optimum topology for this system, is reconfiguration of distribution network problem. The main
illustrated in Fig. 5. contributions of this work are the codification proposal and a
new way to implement the recombination process.
The codification proposal allows representing a radial
topology, and manipulating it, in a way that is easy to generate
other radial topologies without much effort. This process is
accomplished with a simple but powerful recombination
process that allows two generators configurations share their
characteristics into a radial descendent, and finally, the
diversity of the population is maintained thanks to a local
improvement step
Another characteristic of the algorithm is the few
parameters needed in comparison with a traditional genetic
algorithm.
The result of all this elements together is an efficient
algorithm which can solve easily the distribution network
Fig. 5. Descendent Obtained.
reconfiguration problem in real life networks.

B. 136 buses real system.

The 136 buses and 156 branches system is a real IX. REFERENCES
distribution system located in the city of Três Lagoas-MS
(Brazil), the data for this system can be found in [15]. The [1] S. Civanlar, J. Grainger, S. Lee, “Distribution feeder reconfiguration for loss
Reduction,” IEEE Transaction on Power Delivery, vol. 3, n. 3, pp. 1217–1223,
original configuration has 320.351 kW of losses, and the best
1988.
result is reached in near 200 iterations with a population of
100 individuals, with losses of 280.167 kW, executing around [2] S. K. Goswami, S. K. Basu, “A new algorithm for the reconfiguration of
10000 SPF, after 10 seconds and with minimal memory distribution feeders for loss minimization,” IEEE Transactions on Power
Delivery, vol. 7, n. 3, pp. 1484–1491, 1992.
resources.
The performance of the algorithm with this system is [3] M. E. Baran, F. F. Wu, “Network reconfiguration in distribution systems for
presented in Fig. 6. loss reduction and load balancing,” IEEE Transactions on Power Delivery, vol.
4, n. 2, pp. 1401–1407, 1989.

[4] D. A. Shirmohammadi, H. W. Hong, “Reconfiguration of electric

distribution networks for resistive line losses,” IEEE Transactions on Power
Delivery, vol. 4, n. 2, pp. 1492–1498, 1989.

[5] C. Liu, S. Lee, S. Venkata, “An expert system operational aid for restoration
and loss reduction of distribution systems,” IEEE Transactions on Power
Delivery, vol. 3, n. 2, pp. 619–626, 1988.

[6] D. Bouchard, A. Chikhani, V. I. John, M. M. A. Salama, “Applications of

hopfield neural-networks to distribution feeder reconfiguration,” in Proc.
Applications of Neural Networks to Power Systems, IEEE ANNPS 1993., pp.
311–316.

[7] H. Kim, Y. Ko, K. H. Jung, ”Artificial neural-network based feeder

reconfiguration for loss reduction in distribution systems,” IEEE Transactions
Fig. 6. Performance of the algorithm with the 136 bus test system on Power Delivery, vol. 8, n. 3, pp. 1356–1366, 1993.

[8] K. Nara, M. Kitagawa, “Distribution systems loss minimum re-configuration

The best solution found can be identified with the by simulated annealing method,” in Proc. Advances in Power System Control,
connecting arcs out of the tree. In the 136 test system, the best Operation and Management, International Conference. IEEE APSCOM Hong
solution found presents the next connecting arcs out of the Kong. 1991. pp. 461–466 vol.2.
topology: [7, 35, 51, 90, 96, 106, 118, 126, 135, 137, 138, [9] W. M. Lin, F. S. Cheng, M. Tsay, “Distribution feeder reconfiguration with
141, 142, 144, 145, 146, 147, 148, 150, 151, 155]. refined genetic algorithm,” IEE Proceedings in Generation Transmission and
The algorithm was implemented in Fortran 77, and Distribution, vol. 147, n. 6, pp. 349–354, 2000.
executed in a normal desktop equipped with a Pentium M [10] Y. C. Huang, “Enhanced genetic algorithm-based fuzzy multi-objective
1.60 GHz. approach to distribution network reconfiguration,” IEE Proceedings in
Generation Transmission and Distribution, vol. 149, No 5. pp. 615-620, 2002
VIII. FINAL REMARKS
[11] Radha Bhoomesh, R. T. F. King, H. C. S. Rughooputh, “A Modified
This paper presents a metaheuristic to solve efficiently the Genetic Algorithm for Optimal Electrical Distribution Network
 6

Reconfiguration,” in Proc. 2003 Congress on Evolutionary Computing, vol. 2,

pp. 1472-1479.

[12] P.C. Chu, J.E. Beasley: “A genetic algorithm for the generalized
assignment problem”, Computers Operations Research, vol. 24, No 1, pp 17-
23, 1997.

[13] R. C. Prim, “Shortest connection networks and some generalizations,” Bell

System Technical Journal, 36, pp. 1389–1401, 1957

[14] D. Shirmohammadi, H. W. Hong, A Semlyen, G. X. Luo, “A

Compensation-based power flow method for weakly meshed distribution and
transmission networks,” IEEE Transactions on Power Systems, vol. 3,
n. 2, pp. 753-762, 1988.

[15] UNESP-FEIS Electrical Energy Systems Planning Laboratory, test systems,

available: http://www.dee.feis.unesp.br/lapsee/

X. BIOGRAPHIES

Edgar Manuel Carreno (S/04) Received his B.S. degree from Universidad
Nacional de Colombia in 2000, the M.S. from Universidade Tecnologica de
Pereira in Colombia in 2003. Currently he is working in his Ph.D. Thesis at
UNESP – Ilha Solteira, SP, (Brazil). His main interests are Mathematical
Optimization, Load Forecasting and Network Planning.

Nathália Fernandes Moreira is currently attending her B.S. Studies at

Electrical Engineering Department at UNESP – Ilha Solteira, SP,(Brazil).

Rubén Romero (M´92) received his B.S. degree in 1978 and P.E. degree in
1984 from National University of Engineering, in Lima, Peru, his M.S. and
Ph.D. degrees from UNICAMP (Brazil) in 1990 and 1993. Currently he is a
Professor of Electrical Engineering Department of UNESP/Ilha Solteira-SP
(Brazil).