See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/265786476

Power Distribution Network Reconfiguration by Evolutionary Integer

Programming

Conference Paper · September 2014

DOI: 10.1007/978-3-319-10762-2_2

CITATIONS READS

4 930

5 authors, including:

Kaifeng Yang Michael Emmerich

Leiden University Leiden University
25 PUBLICATIONS   123 CITATIONS    278 PUBLICATIONS   4,021 CITATIONS   

SEE PROFILE SEE PROFILE

Thomas Bäck
Leiden University
374 PUBLICATIONS   17,332 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Process mining for multi-objective online control View project

Decision Making in Business, Industry, Politics, and Healthcare View project

All content following this page was uploaded by Kaifeng Yang on 18 September 2014.

The user has requested enhancement of the downloaded file.

 Power Distribution Network Reconfiguration
by Evolutionary Integer Programming

Kaifeng Yang1 , Michael T.M. Emmerich1 , Rui Li1 ,

Ji Wang2 , and Thomas Bäck1
1
LIACS, Leiden University, Niels Bohrweg 1, 2333 CA Leiden, The Netherlands
2
Central South University, 410083 Changsha, Hunan Province, China

Abstract. This paper presents and analyses new metaheuristics for

solving the multiobjective (power) distribution network reconfiguration
problem (DNRP). The purpose of DNRP is to minimize active power
loss for single objective optimization, minimize active power loss and
minimize voltage deviation for multi-objective optimization.
A non-redundant integer programming representation for the prob-
lem will be used to reduce the search space size as compared to a binary
representation by several orders of magnitudes and represent exactly
the feasible (cycle free, non-isolated node) networks. Two algorithmic
schemes, a Hybrid Particle Swarm Optimization - Clonal Genetic Algo-
rithm (HPCGA) and an Integer Programming Evolution Strategy (IES),
will be developed for this representation and tested empirically.
Conventional algorithms for solving multi-objective DNRP are con-
verting the multiple objective functions into a single objective function
by adding weights. However, this method cannot capture the trade-offs
and might fail in case of a concave Pareto front. Therefore, we extend the
HPCGA and IES in order to compute Pareto fronts using selection proce-
dures from NSGA-II and SMS-EMOA. The performance of the methods
is assessed on large scale DNRPs.

Keywords: Power Distribution Network Reconfiguration, Integer Pro-

gramming, Particle Swarm Optimization, Clonal Genetic Algorithm, Evo-
lution Strategies, Multiobjective Optimization.

1 Introduction
With the sustainable development of economy, there is an increasingly high de-
mand for the quality and reliability of the electricity supply in every industry.
Power distribution network reconfiguration is an important method of optimiz-
ing the distribution system, which is significant to enhancing the security, the
efficiency, and the reliability of the system. There are two types of switches in a
power network system: normally closed switches and normally open switches. See
Figure 1, for an example of a power distribution network configuration, the 119
bus system [1], where the black solid lines represent normally closed switches,
and red dashed lines represent normally open switches. Network reconfiguration
is the process of changing the topology of the power network by operating these
switches for the purpose of minimization of the power loss. Since each switch has

T. Bartz-Beielstein et al. (Eds.): PPSN XIII 2014, LNCS 8672, pp. 11–23, 2014.

c Springer International Publishing Switzerland 2014
12 K. Yang et al.

Fig. 1. Initial configuration of the 119-bus test system

two conditions, for a system which has N nodes, there should be 2N −1 potential
switch configurations. To make sure that all the customers can get electricity
and no short circuit exists in the system, there are, however, two constraints for
network reconfiguration: no cycles (the radial structure of the network must be
maintained in each new structure) and no islands (all the loads must be served).
The problem of finding an optimal distribution network reconfiguration is
known to be NP hard, and larger instances with more than 100 nodes cannot yet
be solved exactly, such that several metaheuristics were proposed: Zhang et al.
[1] developed a tabu search algorithm for real power loss minimal reconfiguration
in large-scale distributed systems. Aman et al. [2] proposed to use evolutionary
programming (EP) to find the optimal topology of the distributed network for
minimizing the real power loss. Rao et al. [3] presented artificial bee colony
algorithms for determining the sectionalizing switch to be operated to solve the
real power loss minimization problem. Niknam et al. [4] presented an interactive
fuzzy satisfying method based on hybrid modified honey bee mating optimization
for aggregation-based multiple objective optimization. Kasaei [5] used an ant
colony algorithm to solve the optimal network reconfiguration and capacitor
placement problem for power loss reduction and voltage profile enhancement in
distribution networks.
The contribution of this paper is to discuss concise representations of the
problem, discuss fast and reliable evolutionary integer programming solvers,
 Power Network Reconfiguration 13

and extend them for Pareto-based (set-based) multiobjective optimization. The

objective functions considered will be power loss and voltage profile enhance-
ment, but as opposed to the study in [5], we are interested in visualizing the
trade-off and the robustness by using multiobjective optimization techniques.

2 Problem Description
The network reconfiguration problem in a power distribution system is to find
the best configuration of a radial network. The objective functions are the mini-
mization of power loss and the maximization of the network’s reliability. Besides,
the network has to satisfy certain operating conditions [3].

2.1 Objective Function

The objective function for the minimization of power loss can be described
as [6]:
b
P 2 + Q2 b
min floss = ki Ri i 2 i = ki Ri | Ii |2 (1)
i=1
Vi i=1

subject to:
Vimin ≤ Vi ≤ Vimax (2)

Ii ≤ Iimax , i = 1, ..., b (3)

Here b is the number of branches and for each branch i ∈ {1, ..., b}, Ri is the
branch resistance, Pi and Qi are the active power and the inactive power of a
branch terminal i, Vi is the terminal node voltage of branch i, Vimin and Vimax
are the minimum and maximum bus voltage of branch i, respectively, ki is the
status variable of i-th switch. If ki is 0, then switch i is open and if ki is 1, then
switch i is closed. Ii is the branch current and Iimax is the maximum current in
branch i.
The objective function for minimization of voltage deviation can be ex-
pressed as follows [7][8]:

min fV DI = max{| 1 − Umin |, | 1 − Umax |} (4)

where Umin and Umax are respectively the minimum and maximum values of
bus voltage divided by rated voltage to normalize them to value in [0, 1].

2.2 Feasible Constraint on Network Topology

Recall that, in order to ensure the supply to all nodes and avoid short circuits,
the network must be cycle free and not contain isolated nodes. The previous
method for verifying whether an individual is feasible or not is based on the
topology checks of the structure [9]. This method can be computed fast, but
requires some manual parameter settings beforehand.
14 K. Yang et al.

A design for an automatic feasibility check is proposed next. It rests upon

the following idea of cycle free network construction: suppose there are N nodes
in a system, and there should be at least (N −1) line segments to connect all these
nodes. Then the first line segment can connect 2 arbitrary nodes to the system.
Any new line segment, which is not the first line segment, can only connect a
new node to the existing system, because otherwise a cycle would be introduced.

The feasibility check algorithm is based on the relationship between the num-
ber of the nodes and line segments. Suppose there are N bus buses and N line
branches in the system. Based on the aforementioned, we can easily conclude
that: if N line > N bus, there must be at least one cycle in the structure.
Clearly, if N line < N bus − 1, there must be at least two separated compo-
nents in the structure. Therefore, N line must equal N bus − 1, if the structure
is feasible. Given N line = N bus − 1, feasibility is implied by the non-existence
of isolated nodes, due to the following: if there is a cycle and only N bus − 1
edges can be used, at least one of the nodes cannot be connected and it will be
an isolated node. As detecting isolated nodes is simple, we can now state the
following fast and correct algorithm for checking whether a network is cycle free
and does not contain isolated nodes:

Algorithm 1. Topology Feasibility Check

Step 1 Verify N line equals N bus − 1; if not equal, the individual is not feasible,
otherwise go to step 2;
Step 2 Verify whether there is a separated component or not. If there exists a sepa-
rated component, the individual is not feasible, otherwise it is feasible.

3 Power Loss Minimization Algorithms

Two algorithms for minimizing power loss in DNRP will be discussed next, and
extended in section 6 to bi-objective optimization methods. All algorithms are
based on a concise sequence encoding of the feasible search space.

3.1 Encoding Strategy

The encoding strategy is a main part which influences the efficiency of the al-
gorithm in distribution network reconfiguration. For a genetic algorithm, it is
common to use binary encoding, and it is also straightforward in DNRP as
each switch can be associated with a binary variable. However, considering the
topology constraint, many solutions of the search space induced by binary en-
coding would be infeasible. Instead, we therefore propose to use sequence en-
coding [10], which has the following advantages: 1. It is easy to realize; 2. The
probability of generating a feasible solution is high.
The sequence encoding system regards each loop (potential cycle) as a gene.
In each loop exactly one switch has to be open, as otherwise either there would
 Power Network Reconfiguration 15

be isolated components (two open switches) or there would be short circuits (no
open switch). Each loop of the power network is encoded by a natural number,
and all the switches are coded sequentially by natural numbers starting from 1.
For each gene, the value of this gene is the position of the switch that is open in
the loop represented by the gene. Sequence coding strategy can eliminate most
of the infeasible individuals from the search space.
Example 1. Figure 2, is an example of the IEEE-16 distribution system [6][11].
Here a connection of two bus nodes is also a loop. There are three loops in this
system. Loop 1 is composed by branches 5, 6, 8 and 9; loop 2 is composed by
branches 4, 7, 13, 14 and 15; loop 3 is composed by branches 10, 11 and 12.
To make sure the structure satisfying the constraints, one branch in each loop
should be opened. Therefore, the search space size is 4 × 5 × 3 = 60 using the
sequence encoding system. Compared to a search space size of 212 = 4096 using
binary encoding system the search space can be dramatically reduced.

Fig. 2. IEEE-16 distribution system after coding

However, for a real world problem, there could be hundreds or thousands of

transformers in one city, and the number of solutions in the search space is very
large, even using this sequence coding. For example, in a IEEE-69 system, the
binary encoding generates a search space of size 268 ≈ 2.95 × 1020 and sequence
encoding approximately generates a search space size of 1.78 × 106 . For the 119
system (see Figure 1), the search space size is 2119 ≈ 6.65 × 1035 and 1.44 × 1018 ,
respectively, using binary coding and sequence encoding strategy.

3.2 Hybrid Particle Swarm/Clonal Genetic Algorithm

The Hybrid Particle Swarm Optimization/Clonal Genetic Algorithm (PSO-CGA)
is based on two generational transitions (variation and selection steps) that are
applied in an alternating manner. The PSO is fast for local optimization, and the
CGA is mainly integrated to the PSO in order to prevent premature convergence
and increase diversity, e.g., by a special mutation-shift operator. The flowchart of
PSO-CGA can be seen in Figure 3.
16 K. Yang et al.

Fig. 3. Flowchart of hybrid PSO-CGA

In PSO it is assumed that the solution space has dimension D, and the popu-
lation is composed by N particles X = {x1 , ..., xi , ..., xn }, the position of the i-th
particle is xi = (xi1 , xi2 , ..., xiD )T , Pibest = (Pi1best best
, Pi2 best T
,...,PiD ) stands for
best best best best T
the best known position of particle i, and g = (g1 , g2 , ..., gD ) stands
for the best known position of the entire swarm, the velocity of particle i is
Vi = (vi1 , vi2 , ..., vid )T . Particle xiD updates its velocity and position informa-
tion by [12]:

Vidk+1 = ω × Vidk + c1 × r1 × (Pid

best
− Xid
k
) + c2 × r2 × (gdbest − Xid
k
) (5)
k+1
Xid = k
Xid + Vidk+1 i = 1, 2, ..., N ; d = 1, 2, ..., D (6)
where ω is the inertia factor, c1 and c2 are the learning factors, r1 and r2 are
random real numbers within [0, 1].
Traditionally, clonal genetic algorithms (CGA) [13] use roulette wheel selec-
tion. This paper uses another selection mechanism, and this selection mecha-
nism in CGA partitions the parent population P in three subpopulations, say
PB , PM , PW , the size of which is fixed by |PB | = u1 |P |, |PM | = u2 |P |, and
|PM | = u3 |P | with u1 + u2 + u3 = 1. The new generation is generated in three
steps:
Step 1: Select u1 ×100% best offspring for mutation, if there is no improvement
after mutation, choose the offspring before the mutation; the general idea is to
search for improvements nearby current best solutions.
Step 2: Select u3 × 100% worst offspring for initialization; the general idea
behind this mechanism is to maintain the population’s diversity.
Step 3: Apply mutation operator for the remaining offspring, the difference
between this step 3 and step 1, is that after step 3, the mutated individuals will
always be kept, no matter whether there is an improvement or not. This allows
slow diffusion away from a local optimum.
 Power Network Reconfiguration 17

3.3 CGA Shift and Mutate Operator

The CGA variation operator is composed by shift and mutation operators [13]
[14]. The multi-shift operator adds (or subtracts) the same random offset Num-
ber Shift to all genes in a random subset of genes gene shif t. A random direction
flag Direction F lag is used to decide whether to add or subtract. If the interval
boundary gets exceeded the value of the gene is set to the interval boundary.
Example 2. An example for the multi-shift operator is shown in the table below.
Where gene shif t = [2, 4, 5], and the current individual is [× × 7 11 21].
Parent N umber Shif t Direction F lag Offspring
[× 4 × 11 21] 5 0 [× 9 × 16 26]
[× 4 × 11 21] 5 1 [× 1 × 6 16]
The mutation operator is another main part in CGA. It determines first the
genes to be mutated and then sets them to a new random value. A single-shift
(single-mutation) operator shifts (mutates) only a single gene.

4 Integer Programming Evolutionary Strategy

We studied evolution strategies for integer programming (IES) by Rudolph [15],
as an alternative search algorithm approach. It features a (μ + λ)-selection
scheme, that is λ offspring Q are generated based on μ parents P and the best μ
individuals out of P ∪ Q form the next parent population. Each individual is cre-
ated by selecting randomly two parents (sexual case) or more than two parents
(panmictic case), applying discrete recombination (choose each gene randomly
from one of the parents) or intermediate recombination (averaging) to create a
single offspring, which then is mutated. The mutation operator perturbs each
gene by the difference of two geometrically distributed pseudo-random numbers.
For each gene a step-size ςi is maintained and undergoes a mutation, too, which
makes it possible for the step-size to adapt. The mutation maps an individual
(X, ς) ∈ (Zn × (R+ )n ) to its mutant (X , ς  ) ∈ (Zn × (R+ )n ) as follows:
ςi = max{1, ςi exp(τ Nc + τ  Ni )}, Ni ∼ Normal(0, 1), Nc ∼ Normal(0, 1)
  
ln(1 − uij )
zij = , ϕi = 1 − ζi (1 + 1 + ζi2 )−1 , uij ∼ Uniform(0, 1), j = 1, 2
ln(1 − ϕi )

Xi = Xi + zi1 − zi2 , i = 1, . . . , n (7)

Since the original geometric distribution is single tailed, Rudolph proposed

the use of the difference zi1 − zi2 and could show that the resulting multivariate
distribution has 1 symmetry, maximal entropy, and infinite support. It features
(multiple) self-adaptive mutation step sizes and, for a minimal stepsize greater
than zero, global convergence for t → ∞. In case interval boundaries are exceeded
reflection at the interval boundary is used [16]. To prevent stagnation the stepsize
is bounded from below by 1. Learning rates τ and τ  determine the speed of step-
size adaptation. See also [16] for a detailed description, default parameters and
analysis.
18 K. Yang et al.

5 Single-objective Optimization Result

The simulation results1 are based on distribution system 119 [1]. The test sys-
tem is a 11 kV distribution system with 118 sectionalizing switches and 15 tie
switches, and the total power loads are 22709.7 kW and 17041.1 kVAr. The
topology of the distribution system 119 is shown in figure 1. All simulations were
performed in MATLAB 8.2, CPU: Intel 2 Core 3.16GHz, 2.0 GB DDR RAM (800
MHz). The power flow calculation (to calculate power loss and voltage deviation)
is using Newton Method based on MATPOWER [17], and maximum number of
iterations is 20, termination tolerance on per unit is 1e-8.

Fig. 4. The best structure

PSO-CGA Hybrid: The PSO-CGA hybrid was run with a population size of
30, and the PSO parameters were ω = 0.8, c1 = 2, c2 = 2, v max = 4 (Max
speed factor), v min = −4 (Min speed factor). For each run, 50 PSO steps (7
generations per step) and 50 CGA steps (7 generations per step) were conducted
in alternation. One run takes 15-30 minutes. In the CGA one of the operators,
single-shift, single-mutation, multi-shift, or multi-mutation is chosen randomly.
In case of multi-shift and multi-mutation the number of genes to be mutated
is chosen randomly, too, between 1 and the number of genes. An experiment
was designed to find optimal settings for parameters u1 , u2 , and u3 within their
bounds [0, 1] and respecting the constraint u1 + u2 + u3 = 1. A design of
experiments for mixtures was applied following parameter setting in [18].
1
The MATLAB source code of the numerical experiments is available on request by
the authors and on http://natcomp.liacs.nl/index.php?page=code.
 Power Network Reconfiguration 19

The results (shown in Table 1) of the experimental design are visualized as

ternary diagrams (Figure 5). The best results are achieved with a low rate of
reinitialization (u3 ) and a relatively high rate for u1 .

Fig. 5. Mixture design (left) and interpolated results (middle, right) of u1 (selection),
u2 (diffusion), and u3 (reinitialization) for CGA-PSO hybrid on Test System 119

Table 1. Experiments on Hybrid PSO-CGA

u1 1 0 0 0.5 0.5 0 0.333 0.167 0.167 0.666 0.333
u2 0 1 0 0.5 0 0.5 0.333 0.167 0.666 0.167 0.333
u3 0 0 1 0 0.5 0.5 0.333 0.666 0.167 0.167 0.333
Exp.1 875.155 1037.9 1329.96 883.828 875.155 1325.33 891.033 891.864 875.155 890.918 874.86
Exp.2 875.155 1085.33 1478.89 875.155 875.155 1120.87 875.155 990.197 875.155 874.86 874.86
Exp.3 875.539 1110.29 1487.06 874.86 890.918 1057.02 875.155 875.155 888.966 875.155 874.86
Exp.4 870.856 1209.17 1138.57 874.86 874.86 1412.98 883.858 887.39 875.155 875.155 875.155
Exp.5 875.155 1068.33 1408.56 875.155 886.437 1366.81 903.49 876.178 876.729 883.641 874.86
Exp.6 875.155 1130.18 1492.56 875.155 875.155 1405.2 875.155 874.86 874.86 874.86 875.155
Exp.7 875.155 1195.5 1492.28 883.858 875.155 1363.57 875.155 869.727 874.86 891.57 875.155
Exp.8 887.505 1081.05 1390.6 877.34 878.96 1186.35 874.86 890.918 874.86 878.209 889.16
Exp.9 874.86 1014.13 1281.5 874.86 874.86 1277.85 878.334 890.918 887.39 874.86 875.155
Exp.10 875.155 1041.54 1384.4 874.86 874.86 1177.81 875.155 875.155 1013.09 869.727 874.86
Min 870.856 1014.13 1138.57 874.86 874.86 1057.02 874.86 869.727 874.86 869.727 874.86
Max 887.505 1209.17 1492.56 883.858 890.918 1412.98 903.49 990.197 1013.09 891.57 889.16
Mean 875.969 1097.34 1388.438 876.993 878.152 1269.38 880.735 892.236 891.622 878.896 876.408
Deviation 4.27671 65.1574 114.1382 3.68630 5.78028 126.08 9.60066 35.3921 43.0215 7.37168 4.482993

Integer Programming Evolution Strategy: The parameters for the IES are as fol-
lows: μ = 15, λ = 100, σinitial = (ub − lb)/8, ubsigma = (ub − lb)/3, lbsigma = lb,
τ = √2n1
and τ  = √ 1√ , where ub and lb is the upper bound and lower bound
z 2 nz
vector of the variables, N = 15 in case of the 119 system. Each run had 100
iterations and took about 5-10 minutes. For the IES preliminar experimentation
showed that the recombination type was a crucial choice. Table 2 shows results of
different recombination methods based on the Integer ES. We tested four recom-
bination types: none (1), discrete (2), panmictic discrete (3), and intermediate
(4); see [19]. The best result that was found was with a discrete recombination
on the object variables, and a panmictic discrete recombination on the step size.
20 K. Yang et al.

Table 2. Experimental results for the 119 system based on ES

s. X = 1 s. X = 2 s. X = 3 s. X = 4
Min 939.6829 893.6304 935.478 930.895
s. ς = 1
Max 1302.274 1687.879 1556.437 1904.853
Min 881.8714 874.8604 869.7271 878.3646
s. ς = 2
Max 1083.146 978.1069 974.6982 1048.195
Min 875.155 894.2617 878.3646 876.0975
s. ς = 3
Max 1009.264 1088.795 1120.487 1090.32
Min 912.3288 1103.351 950.3428 1059.776
s. ς = 4
Max 3666.716 7029.378 3071.888 10456

Both Hybrid PSO-CGA and IES can find the same optimal result, see Table 3.
The open switches in best result, whose structure is shown in Figure 4, are 42-43,
26-27, 23-24, 51-52, 62-49, 58-59, 39-40, 91-96, 71-72, 74-75, 97-98, 108-83, 105-
86, 109-110 and 34-35. Comparing PSO-CGA and IES, it is found that the former
was more robust (better average values) while the latter found better results and
converged faster (IES 5-10 min, PSO-CGA 15-30 min). All strategies perform
significantly better than PSO-CGA with settings u1 = 0, u2 = 0, u3 = 1, which
is essentially a trial and error strategy (within the sequence representation). The
improvement is by a factor of 1298.0861/869.7271 , i.e. the power loss found by
PSO-CGA or IES metaheuristics is only ca. 67% of the power loss of a solution
found by trial and error for the same number of evaluations.

Table 3. Best results for 119 system

Before reconfiguration After reconfiguration based on MATPOWER

Refference [1] [3] Matpower [1] [3] Hybrid PSO-CAG & IES
Unit: kW 1294.3 1298.09 1298.0861 887.5055 869.7271 869.7271

6 Multiobjective Optimization
We extended IES to a multiobjective optimization algorithm by replacing the
selection scheme by that of a multiobjective algorithm, namely the (μ + μ) se-
lection of NSGA-II [20] and the (μ + 1) selection of SMS-EMOA [21] (which has
earlier been used also in Pareto archivers [22]). As a second objective voltage
deviation was minimized (see equation 4).
As an adaptation, we introduced a variant of SMS-EMOA and NSGA-II with a
self-adaptive single step size. Whenever more than five mutations per individual
were unsuccessful, the step size was multiplied by a constant factor of 1/1.2
(following the 1/5th success rule). Success of a generation was registered if a new
non-dominated solution entered the archive of non-dominated solutions among
all solutions encountered so far.
The results (attainment curves) for a population size of μ = 30 and 11 runs per
algorithm are shown in Figure 6, where f 1 and f 2 represent voltage deviation and
 Power Network Reconfiguration 21



  





 

 
 
!" 




























 






 



 



  





 

 
 
!"# 




























 






 



 

Fig. 6. Best, worst, and average attainment curves for the multiobjective optimization
of 119 DNRP

power loss. In the following discussion by Pareto front, we mean the archive of all
non-dominated solutions encountered in a single run. SMS-EMOA with a single
step size provides the best Pareto fronts in the best case and also in the average
case. Interestingly, all strategies find a Pareto front with a concave part. The
interpretation of this is that locally there is a strong conflict between power loss
minimization and voltage deviation minimization for this problem. However, the
range of voltage deviation is relatively small, so that ’from a distance’ the Pareto
front has a clear knee point region. Solutions in this region can be recommended
as good compromise solutions, whereas points located on the flanks of the Pareto
front are not recommended as small improvements in one objective will cause
large deterioration of the other objective.

7 Conclusions
In this paper two well performing optimization strategies for solving the power
distribution network reconfiguration problem have been described and tested on
a challenging problem with more than 100 switches. A concise integer repre-
sentation was chosen and it was demonstrated that it reduces the search space
size by many orders of magnitude as opposed to the binary representation used
in genetic algorithms so far. In the experiments we focused on finding a good
ratio between exploration and exploitation in the PSO-CGA and on choosing a
good recombination operator in the IES with self-adaptive mutation. This turned
out to be discrete recombination on object and step-size variables. The results
clearly show that it is much better to use metaheuristics instead of trial and
22 K. Yang et al.

error strategies when minimizing power loss. Also multiobjective optimization

algorithms, NSGA-II and SMS-EMOA, were applied to compute a Pareto front
between voltage deviation and power loss objectives. These objectives turned out
to be conflicting in the knee point region but globally, when zooming out, they
appear to be complementary. For the future work, it is recommended to test the
strategies on a broader set of benchmarks and further investigate multiobjective
problem formulations, for instance including objective functions on reliability.

Acknowledgment. Kaifeng Yang acknowledges financial support from China

Scholarship Council (CSC), CSC No.201306370037.

References

[1] Zhang, D., Fu, Z.C., Zhang, L.C.: An improved ts algorithm for loss-minimum
reconfiguration in large-scale distribution systems. Electric Power Systems Re-
search 77(5), 685–694 (2007)
[2] Aman, M.M., Jasmon, G.B., Naidu, K., Bakar, A.H.A., Mokhlis, H.: Discrete evo-
lutionary programming to solve network reconfiguration problem. In: TENCON
Spring 2013 Conference, pp. 505–509. IEEE, Sydney (2013)
[3] Rao, R.S., Narasimham, S.V.L., Ramalingaraju, M.: Optimization of distribution
network configuration for loss reduction using artificial bee colony algorithm. In-
ternational Journal of Electrical Power and Energy Systems Engineering 1(2),
116–122 (2008)
[4] Niknam, T., Meymand, H.Z., Mojarrad, H.D.: An efficient algorithm for multi-
objective optimal operation management of distribution network considering fuel
cell power plants. Energy 36(1), 119–132 (2011)
[5] Kasaei, M.J., Gandomkar, M.: Loss reduction in distribution network using si-
multaneous capacitor placement and reconfiguration with ant colony algorithm.
In: Asia-Pacific Power and Energy Engineering Conference, APPEEC 2010, pp.
1–4. IEEE, Chengdu (2010)
[6] Qin, Y.M., Wang, J.: Distribution network reconfiguration based on particle clonal
genetic algorithm. Journal of Computers 4(9), 813–820 (2009)
[7] Chiou, J.P., Chang, C.F., Su, C.T.: Variable scaling hybrid differential evolution
for solving network reconfiguration of distribution systems. IEEE Transactions on
Power Systems 20(2), 668–674 (2005)
[8] Niknam, T.: A new hybrid algorithm for multi-objective distribution feeder re-
configuration. Cybernetics and Systems: An International Journal 40(6), 508–527
(2009)
[9] Wang, S.X., Wang, C.S.: A novel network reconfiguration algorithm implicitly
including parallel searching for large-scale unbalanced distribution systems. Au-
tomation of Electric Power Systems 24(19), 34–38 (2000)
[10] Ma, X.F., Zhang, L.Z.: Distribution network reconfiguration based on genetic
algorithm using decimal encoding. Transactions of China Electrotechnical Soci-
ety 19(10), 65–69 (2005)
[11] Bi, P.X., Liu, J., Liu, C.X., Zhang, W.Y.: A refined genetic algorithm for
power distribution network reconfiguration. Automation of Electric Power Sys-
tems 26(2), 57–61 (2002)
 Power Network Reconfiguration 23

[12] Poli, R., Kennedy, J., Blackwell, T.: Particle swarm optimization. Swarm Intelli-
gence 1(1), 33–57 (2007)
[13] Li, M.J., Tong, T.S.: A partheno genetic algorithm and analysis on its global
convergence. Automatization 25(1), 68–72 (1999)
[14] Wang, J., Luo, A., Qi, M.J., Li, M.J.: The improved clonal genetic algorithm
& its application in reconfiguration of distribution networks. In: Power Systems
Conference and Exposition, PSCE 2004, pp. 1423–1428. IEEE, New York (2004)
[15] Rudolph, G.: An evolutionary algorithm for integer programming. In: Davidor,
Y., Männer, R., Schwefel, H.-P. (eds.) PPSN III. LNCS, vol. 866, pp. 139–148.
Springer, Heidelberg (1994)
[16] Li, R., Emmerich, M., Eggermont, J., Bäck, T., Schütz, M., Dijkstra, J., Reiber,
J.H.: Mixed integer evolution strategies for parameter optimization. Evolutionary
computation 21(1), 29–64 (2013)
[17] Zimmerman, R.D., Murillo-Sánchez, C.E., Thomas, R.J.: Matpower: Steady-state
operations, planning, and analysis tools for power systems research and education.
IEEE Transactions on Power Systems 26(1), 12–19 (2011)
[18] Anderson, M., Whitcomb, P.: Find the optimal formulation for mixtures (2002),
http://www.statease.com/pubs/chem-2.pdf
[19] Bäck, T.: Evolutionary algorithms in theory and practice: evolution strategies,
evolutionary programming, genetic algorithms. Oxford University Press, New
York (1996)
[20] Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.A.M.T.: A fast and elitist multi-
objective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Com-
putation 6(2), 182–197 (2002)
[21] Beume, N., Naujoks, B., Emmerich, M.: SMS-EMOA: Multiobjective selec-
tion based on dominated hypervolume. European Journal of Operational Re-
search 181(3), 1653–1669 (2007)
[22] Knowles, J., Corne, D.: Properties of an adaptive archiving algorithm for storing
nondominated vectors. IEEE Transactions on Evolutionary Computation 7(2),
100–116 (2003)

View publication stats