Computers & Operations Research 40 (2013) 24–32

Contents lists available at SciVerse ScienceDirect

Computers & Operations Research

journal homepage: www.elsevier.com/locate/caor

An iterated sample construction with path relinking method: Application

to switch allocation in electrical distribution networks
Alexander J. Benavides a, Marcus Ritt a,n, Luciana S. Buriol a, Paulo M. Franc- a b
a
Instituto de Informática, Universidade Federal do Rio Grande do Sul – UFRGS, Brazil
b
Faculdade de Ciências e Tecnologia, Universidade Estadual Paulista – UNESP, Brazil

a r t i c l e i n f o a b s t r a c t

Available online 24 May 2012 We present a metaheuristic approach which combines constructive heuristics and local searches based
Keywords: on sampling with path relinking. Its effectiveness is demonstrated by an application to the problem of
Metaheuristics allocating switches in electrical distribution networks to improve their reliability. Our approach also
Sample algorithms treats the service restoration problem, which has to be solved as a subproblem, to evaluate the
Path relinking reliability benefit of a given switch allocation proposal. Comparisons with other metaheuristics and
Power systems reliability with a branch-and-bound procedure evaluate its performance.
Switch allocation & 2012 Published by Elsevier Ltd.

1. Introduction In Section 2 we give a formal description of the problem. In

Most of the faults of an electrical power system take place in
the distribution network [1,2]. The installation of redundant lines
and switches is a common method to improve the reliability of a
distribution network. In this way, in case of failures, service
providers can modify the network topology and reduce the areas
without energy. The installation of switches in every line is
impracticable due to high costs. For this reason, companies must
carefully choose the lines to install switches. This combinatorial
optimization problem is called the switch allocation problem. To
assess the reliability of an allocation proposal for a set of switches,
we have to determine the expected non-supplied energy for a
given distribution of failures. Since the actual non-supplied
energy depends on how much of the service area can be restored
by reconfiguring the network, the service restoration with elec-
trical constraints occurs as a subproblem of the switch allocation
problem.
In this paper, we propose a sample construction and a sample
local search for the switch allocation problem, and, based on these
algorithms, a new iterated sample construction with path relink-
ing (ISCPR) to solve the switch allocation problem. Another
contribution is that our treatment of the service restoration
problem makes no assumption on the topology of the network
and includes electrical restrictions.
Section 3 we describe the construction and local search algo-
rithms based on sampling. In Section 4 we introduce the ISCPR
algorithm which uses the sample construction to modify the
current solution and the path relinking to perform a guided local
search. Section 5 first compares the performance of sample based
construction and local search methods to other such methods,
and then compares several heuristics for the switch allocation
problem. When compared to a heuristic based on Greedy Rando-
mized Adaptive Search Procedure (GRASP) [3], ISCPR shows the
best performance. Another comparison with a branch-and-bound
procedure attests the quality of the solutions yielded by our
approach.
2. Description of the problems
Fig. 1a shows an example of an electric power distribution
network in normal operation. The basic circuit of an operational
distribution network has no cycles. It is composed of substations
(square nodes), consumers (round nodes), and feeder lines (solid
lines). Additionally, redundant lines (dotted lines) can restore the
energy in areas affected by failures. Switches control the power
flow. In normal operation, switches of redundant lines are open,
and switches in the basic circuit are closed.

2.1. Graph model of distribution networks

n
We use an undirected graph G ¼(N,E) to model a distribution
Corresponding author. Tel.: þ55 5133086818.
E-mail addresses: ajbenavides@inf.ufrgs.br (A.J. Benavides),
network. The set of nodes N represents substations and consumer
marcus.ritt@inf.ufrgs.br (M. Ritt), buriol@inf.ufrgs.br (L.S. Buriol), load points, and the set of edges E represents the feeder lines.
paulo.morelato@fct.unesp.br (P.M. Franc- a). A Boolean value Be indicates the presence of a switch on a line

0305-0548/$ - see front matter & 2012 Published by Elsevier Ltd.

http://dx.doi.org/10.1016/j.cor.2012.05.006
 A.J. Benavides et al. / Computers & Operations Research 40 (2013) 24–32
1 2
8 13
4 10 14
9 9
5 11 12
6 7 16
Fig. 1. Three-feeder example of a distribution network by Civanlar et al. [4]. (a) Normal operating conditions and (b) sectors affected by a failure.
e A E. We represent a solution for the switch allocation problem by
a set EB ¼ feA E9Be ¼ 1g of lines that are selected to install new
switches. The sector SðnÞ of a node n A N is the largest connected
subgraph which contains n and is connected only with basic
circuit feeder lines (normally closed) that have no switch installed
(Be ¼0). Ellipses in Fig. 1b represent sectors. The sector of a line
e A E is defined as the union of e with the sectors corresponding
to the two incident nodes. The frontier F ðG0 Þ of a subgraph
(or sector) G0 is the set of edges e A E which are incident to exactly
one node in G0 .
2.2. Network reliability estimation
In this paper, we use the expected energy non-supplied (EENS)
[5] to measure the network reliability. It is defined as
X X
EENS ¼ lf rf P n ðMW h=yearÞ,
f A Enc n A Nf
where Enc is the set of lines that can fail (normally closed), Nf is
the set of nodes affected by a failure f, lf is the average failure
rate, rf is the average outage time, and Pn is the average energy
consumption of node n.
Our reliability estimation algorithm simulates a failure f in
each sector S of the network, and opens the switches of the
corresponding frontier F ðSðf ÞÞ to isolate the failure. Then, it
determines the set Nf of consumers affected by the failure with
a service restoration algorithm, and calculates their partial EENS.
Evaluating sector by sector, the partial reliability of every feeder
line of a sector is evaluated at once, reducing the number of
service restoration problems that have to be solved. The same
process is applied to the remaining lines that are not part of any
sector.
2.3. The service restoration problem
Network reconfiguration is the process of opening and closing
some switches to change the network topology. When a power
failure is detected, the network is reconfigured to isolate the
failure and to restore energy by alternate lines.
Consider a failure in line f of Fig. 1b. Without switches, the
whole tree under substation 2 would be unattended. When the
frontier switches F ðSðf ÞÞ ¼ fa,b,cg are opened, the failure is iso-
lated in sector Sðf Þ. Consequently, sector Sð9Þ has neither failure
nor power supply. The service on sector Sð9Þ is restored when the
switch on line d is closed.
The service restoration problem consists in choosing which
switches must be opened or closed to minimize the unattended
25
3 1 2 3
a
S(f)
8 13
4 b 10 14
c
S(9)
5 11 12
d
15 6 7 16 15
area after the isolation of a failure. A solution has to satisfy electrical
constraints. The electrical constraints considered in this paper are
the radiality of the reconfigured network, the capacities of the lines
and substations, and the maximum allowed voltage drop.
Since we are interested in solving the service restoration
problem as a subproblem of the switch allocation problem, we
do not consider secondary objective functions, such as minimiz-
ing the number of switching operations. For the same reason, we
do not consider additional constraints such as the priority of
customers, switching times, or restoration cost.
For a single substation, the service restoration problem can be
modeled as finding a spanning tree minimizing the non-supplied
energy subject to electrical constraints [6]. Different from the
minimum weight spanning tree, this version is NP-hard [7,8].
Our service restoration algorithm calculates the affected con-
sumers after a failure isolation. It expands the attended area
starting from the substations sector by sector, when this is
possible without violating the electrical constraints. The electrical
simulation is computationally expensive, but important to obtain
a realistic approximation of the attended area.
A detailed description of the reliability estimation and the
service restoration algorithm can be found in [9]. For a survey of
approaches to service restoration we refer the reader to Ćurčić
et al. [10] and Sudhakar and Srinivas [11].
2.4. The switch allocation problem
The number and position of the switches in the network
influences the amount of non-supplied energy in case of con-
tingencies. The switch allocation problem consists in selecting a set
of feeder lines to install a given number of new switches in a
distribution network. The objective is to maximize the reliability,
and is subject to electrical and topological constraints.
The switch allocation problem is NP-hard, since its solution
requires to solve the NP-hard service restoration problem. For this
reason the literature concentrates on heuristic solution methods.
Exact methods for the switch allocation problem based on
integer programming have been proposed by Zambon et al. [12],
Bupasiri et al. [13], Soudi and Tomsovic [14,15], Sohn et al. [16].
These methods apply only to small-sized distribution networks.
Among the heuristics proposed are simulated annealing
[17], immune algorithms [18,19], genetic algorithms [20–22],
divide-and-conquer [23], reactive tabu search [24], evolutionary
algorithms [25], particle swarm optimization [26], and ant colony
optimization [5]. Some of these methods address the more
general problem of network reinforcement planning, allowing to
26 A.J. Benavides et al. / Computers & Operations Research 40 (2013) 24–32
install besides the sectionalizing switches other devices, e.g.
protective devices such as reclosers or fuses, or capacitor banks.
Comparing the efficiency of these methods directly is not
possible, since the authors use different test instances, which
are often not completely specified or from private real networks.
Further, reliability is measured differently, for example by SAIDI
[20], SAIFI [21], EENS [23,5], or ECOST [17,26]. Some authors use
combined costs for non-supplied energy and installation and
maintenance [17–19,22–24,26,5], or multiple objectives, e.g., the
number of installed switches versus the non-supplied energy
[25], or the installation costs versus the EENS [5].
Many of these methods considerably simplify the underlying
service restoration problem by assuming that any redundant line
can restore energy supply and that the affected area or the overall
cost are known or easy to compute.
3. Construction and local search
This section describes four construction algorithms (random,
sample, greedy and semi-greedy) and three local search strategies
(sample search, first and best improvement) for the switch
allocation problem.
3.1. Construction algorithms
Random construction selects k switches randomly and evalu-
ates the resulting solution. Greedy, semi-greedy and sample
constructions build a feasible solution by adding one element at
each iteration. Greedy construction evaluates all the elements to
select the best each time. Semi-greedy construction (Fig. 2a)
evaluates every possible element to build a restricted candidate
list which contains a fraction of a elements with the highest
reliability. Finally, one element is selected randomly from the
restricted candidate list. (Setting a ¼ 0 selects the best element,
while a ¼ 1 selects randomly between all the elements.) Sample
construction (Fig. 2b) first selects randomly a sample of b
elements. Then, it evaluates the sample candidate list to choose
the best. (A value of b ¼ 0% corresponds to a random construc-
tion, while b ¼ 100% corresponds to a greedy construction).
The most expensive operation is the reliability estimation.
When installing k switches, the number of reliability estimations
is Oðk9E9Þ for the greedy and semi-greedy constructions, Oðk9E9bÞ
for sample construction, and O(1) for random construction.
Fig. 2. Construction algorithms. (a) Semi-greedy and (b) sample.
3.2. Local search algorithms
A local search algorithm iteratively replaces the current solu-
tion with a better neighbour. The entire neighbourhood for our
local search strategies is defined by the relocation of one switch at
a time from its current position e A EB to a feeder line e0 A E\EB .
First improvement local search explores the neighbourhood until
a better neighbour is found. This neighbour becomes the current
solution and the search continues with the next iteration. The
number of evaluated neighbours varies with the iterations. Best
improvement local search explores all the neighbourhood to
select the best neighbour for the next iteration. Sample local
search (Fig. 3) reduces the size of the explored neighbourhood
and the number of reliability estimations. It samples b percent of
the lines with switches ða A EB Þ and b percent of lines without
switches ðbA E\EB Þ to relocate one switch. If the algorithm finds a
better solution in the sample, it becomes the current solution for
the next iteration. Sample and best improvement evaluate a
constant number of neighbours in each iteration, but sample
search reduces the neighbourhood size.
First and best improvement stop when there are no better
solutions in the neighbourhood. Sample neighbourhood explora-
tion is not exhaustive and does not guarantee to find a local
minimum. Thus, the stop criterion may be a maximum number of
iterations or a number of iterations without improvement. To
guarantee that a local minimum is reached, a best improvement
local search can be performed after the sample search.
When installing k switches, the number of reliability estima-
2
tions of each iteration is Oðk9E9b Þ for the sample local search and
Oðk9E9Þ for the other two strategies.
4. Iterated sample construction with path relinking (ISCPR)
ISCPR is a variant of iterated local search (ILS) [27]. At each
iteration, ISCPR perturbs the last solution with a partial random
destruction followed by a sample construction, and uses path
relinking as a guided local search. Path relinking [28] explores the
neighbourhood between two solutions. It repeatedly relocates
one switch of the initial solution, to a position defined by the
guiding solution. It can be seen as a local search where the
neighbourhood is limited to the differences between initial and
guiding solutions. By sampling and searching only in a limited
neighbourhood, ISCPR reduces the number of reliability estima-
tions during construction and local search phases.
 A.J. Benavides et al. / Computers & Operations Research 40 (2013) 24–32 27

Table 1
Instances.

Inst. SS Cns. Lns Pw.D. Op.V.

BR 1 32 39 3.70 12.66
BU 1 32 39 3.70 12.66
R3 1 32 37 4.55 12.66
B4 3 38 72 24.58 11.00
R4 11 83 96 35.20 12.66
AR 5 80 109 28.60 20.00
AU 5 80 109 28.60 20.00
R5 8 135 156 19.96 12.66
R6 3 201 216 32.44 33.60
R7 7 873 900 148.99 126.60

SS, Cns, Lns: Number of substations, consumers, and lines, respectively.

Op.V.: Operation voltage (kV).
Pw.D.: Total power demand (MW).

Consumer demand for instances BR and AR are random values,

Fig. 3. Sample local search algorithm.
and for BU and AU are constant values.
Instance B4 is the RBTS Bus 4 from Billinton and Jonnavithula
[17]. Instances R3, R4, R5, R6, and R7 are from the REpository of
Distribution Networks (REDS) [29]. To complete the necessary
information, we assume an outage time of r ¼ 2 h, a resistance
R ¼ 0:257 O=km, a reactance X ¼ 0:087 O=km, a failure rate
l ¼ 0:065 f=km=year, and a maximum current IMAX ¼ 500 A for
every line. The problem instances can be found at [30].

5.2. Choice of parameters

The proposed algorithms depend on five parameters, as shown

Fig. 4. ISCPR algorithm.
The ISCPR algorithm (Fig. 4) starts generating a solution with
the sample construction algorithm. After initialization, it iterates
over three steps. First, some elements are removed from the
previous solution. Then, a solution is reconstructed using the
sample construction algorithm. Once a new solution is complete,
path relinking guides a local search from the current solution to
the best solution. Finally, the best solution found is returned.
5. Computational experiments
All tests have been executed on a PC with an Intel Core i7
processor (2.8 GHz, 12 GB RAM). The algorithms have been
implemented in Cþþ, and compiled with GNU Cþþ with optimiza-
tion level 2 (-O2).
in Table 2.
To find a good setting we conducted experiments with the
construction and local search procedures on instances B4 and R6.
In some preliminary tests we found that the sample local search
rarely improved after 10 iterations without improvement,
so we fixed the stop criterion at I¼10. We tested values of
a A f0:1,0:2, . . . ,0:9g and b A f10%,20%, . . . ,90%g and compared
the averages of 100 replications with different random seeds of
the results after construction and after local search.
We observe that the goal of tuning the construction proce-
dures is not only to provide solutions of good quality, but also to
maintain a good variation of the constructed solutions that can be
exploited in the subsequent optimization by a local search or
another heuristic. We found that for both instances the solution
quality of the semi-greedy construction increased with a and the
solution quality of the sample construction increased with b.
After the local search, the variation of the solution quality did not
depend significantly on a for semi-greedily constructed solutions,
so we chose to fix a ¼ 0:5 at the midpoint of its domain. The
solutions constructed by sampling produced the largest variation
for b ¼ 10%. Since the main function of the sample construction is
to provide diverse solutions we fixed b at this value. Another
advantage of a smaller sample size is that the construction is
faster.
Finally, for the parameter d in ISCPR we found that d ¼ bs=2c,
where s is the number of switches to install to be a good
compromise between perturbing the solution sufficiently to
escape a local minimum, but not too much to produce a random
solution.

5.1. Problem instances 5.3. Construction and local search methods

Table 1 presents details of the problem instances that were used To compare construction and local search methods, we tested
in our tests. Instances AU and AR are from Augugliaro et al. [7], all combinations C–L of construction algorithms and local search
and BU and BR are from Baran and Wu [6]. The total power strategies. The construction algorithm C can be greedy (Gr), semi-
demand was divided between the consumers load points. greedy (SGr), sample (Spl), or random (Rnd). The local search L
28 A.J. Benavides et al. / Computers & Operations Research 40 (2013) 24–32

Table 2
Parameters and values used in the experiments.

Par. Description Value

a Greediness in semi-greedy 0.5

construction
b Sample size in sample 10%
construction
I Maximum iterations without 10
improvement in SLS and SBI
d Number of elements destroyed bs=2c
and reconstructed in ISCPR of a
total of s

70
19 Construction
Rnd

Relative deviation from best known solution (11.262 MWh/year)

SGr
Spl
18 Gr 60

Local search
17 BI 50
FI
EENS (MWh/year)

Spl
16
40

15
30

14
20

13

10
12

0
0 2 4 6 8 10
Time (seconds)

240
6 Construction
Rnd
Relative deviation from best known solution (1.793 MWh/year)

SGr
5.5 Spl
Gr 200

5 Local search
BI
FI 160
4.5 Spl
EENS (MWh/year)

4 120

3.5

80
3

2.5 40

2
0
0 20 40 60 80 100
Time (seconds)

Fig. 5. Average performance for construction and local search combinations. (a) Instance B4 and (b) instance R6.
 A.J. Benavides et al. / Computers & Operations Research 40 (2013) 24–32 29

can be first improvement (FI), best improvement (BI), sample
(SLS), or sample followed by best improvement (SBI). For exam-
ple, Spl–SBI is the combination of a sample construction and a
sample local search followed by a best improvement local search.
We repeated each experiment 1000 times for B4, and 100 times
for R6 (except Gr–BI which is deterministic) allocating 20
switches. Fig. 5 and Table 3 report average results.
Table 3 shows the average relative deviation from the best
known solution after the construction (c) and after the local
search (l), the average runtime after construction (tc) and local
search (tl), and the percentage of tests that reach the best known
solution (lBK), or the relative deviation of the best found solution
in parentheses when the best known solution is not found. The
best solution found in our tests for 20 switches has an EENS of
11.262 MW h/year for B4 and 1.793 MW h/year for R6.
Fig. 5 shows the average EENS and the relative deviation (in
percent) from the best known solution over time. Four points
show the average result of the construction algorithms and curves
depict the average performance of the local search methods.
In the following we refer to values for instance B4 and give the
values for R6 in parentheses. Initial solutions created by the semi-
greedy algorithm are better than random solutions by 2.2
(1.2) MW h/year, and require 0.5 (4.7) s while random solutions
require less than 4 ms. Greedy construction generates always the
best initial solution at almost the same computational effort than
semi-greedy construction, but a following local search does not
find the best known solution. Solutions created by the sample
algorithm are better than random solutions by 5.0 (3.4) MW
h/year. Sample construction creates better solutions than the
semi-greedy algorithm in about 10% of the corresponding time,
having the best cost/benefit among all construction algorithms.
The average final solutions after local search shown in Table 3
are less than 3.5 (4.6)% from the best known solution, except after
sample local search with 6.9 (12.8)%. All search strategies find the
best known solution at least once, except sample search, which
yields solutions 0.4 (0.9)% above the best known value. The search
strategies differ in their performance over time. The curves in
Fig. 5 show a quick initial progress for FI, but BI becomes better
later. FI finds better neighbours in early iterations, but spends
most of its runtime in later iterations with small improvements.
The runtime of sample search is very small, due to the restricted
neighbourhood. Performing BI after sample search (SBI) to reach a
local minimum increases the runtime, but it remains about half of
the runtime of BI.
Rnd-FI is the most expensive combination. The cheapest
combination is Spl–Spl. The fastest combination that finds the
best known solution in our experiments is Spl–SBI.
5.4. Comparison of metaheuristics
In this section, we compare ISCPR with a Greedy Randomized
Adaptive Search Procedure (GRASP) [3]. GRASP is a metaheuristic
which repeatedly constructs a solution and applies a local search.
Both the semi-greedy and the sample construction algorithms
present the greedy, randomized and adaptive characteristics of
GRASP as described below. The semi-greedy construction restricts
greedily the candidate list and selects randomly one of them,
while the sample construction restricts randomly the candidate
list and selects greedily one of them. Both constructions are
adaptive when selecting a new element because they take into
account the benefits of the previously selected elements. Four
variants of GRASP were implemented combining the construction
and local search algorithms tested before (SGr–BI, SGr–SBI, Spl–BI,
Spl–SBI). We also tested variants of GRASP with path relinking to
the current best solution after the local search (GRASP-PR), since
often this can improve GRASP [31].
An iterated semi-greedy construction with path relinking, ICPR
(SGr-PR) is also compared to ISCPR (Spl-PR). Fig. 6 and Table 4
present results for the installation of 20 switches. Experiments
were repeated 100 times, except for instance R7 with 10 repeti-
tions). The stop criterion is a fixed time tmax given in Table 4, and
10 min for results in Fig. 6.
Fig. 6 presents a time-to-target plot [32,33] which shows the
percentage of tests that reach a given target value within a given
execution time. It uses instance B4. The target is the best known
solution for 20 switches.
Comparing the average runtime of each heuristic to standard
GRASP (SGr-BI), the sample search GRASP (SGr-SBI) is faster by a
factor of 0.75, the sample construction GRASP (Spl–BI) is slower by a
factor of 2.4, GRASP with both sample algorithms (Spl–SBI) is slower
by a factor of 1.8, and GRASP-PR (SGR-BI-PR) is slower by a factor of
1.4. The other GRASP-PR variants that we tested took longer than
their GRASP counterparts to obtain the same percentage of target

Table 3
Comparison of construction and local search algorithms.

Algorithm Instance B4 Instance R6

Constr. L.S. c tc l lBK tl c tc l lBK tl

Gr BI 3.95 0.826 3.95 (3.95) 2.51 7.37 6.450 0.06 (0.06) 113.54

SGr FI 49.53 0.444 2.18 17.5 29.56 164.41 4.699 2.99 10.0 808.58
BI 2.24 21.3 18.19 3.51 5.0 309.60
SLS 6.18 0.5 1.12 11.34 (1.06) 15.33
SBI 2.64 15.1 9.01 2.78 1.0 152.72

Rnd FI 69.20 0.001 2.40 8.6 44.42 229.47 0.003 4.56 1.0 1182.59
BI 2.36 14.5 22.11 3.79 4.0 328.93
SLS 6.22 0.2 0.74 12.76 (2.14) 11.04
SBI 2.76 13.5 8.53 3.66 4.0 137.41

Spl FI 25.04 0.056 3.37 5.3 18.98 43.01 0.584 2.76 15.0 351.71
BI 3.33 7.8 13.12 2.57 7.0 223.36
SLS 6.87 (0.41) 0.59 11.08 (0.91) 8.65
SBI 3.45 5.8 8.38 2.57 2.0 133.42

c, l: average percentage over best known solution of construction and local search, respectively.
lBK: percentage of tests that reach the best known solution (or relative deviation in parentheses if not reached).
tc, tl: average runtime in seconds of construction and local search, respectively.
30 A.J. Benavides et al. / Computers & Operations Research 40 (2013) 24–32

0.9

0.8

0.7

0.6
Probability

0.5

0.4

0.3

ISCPR (Spl-PR)
0.2 ICPR (SGr-PR)
GRASP (SGr-SBI)
GRASP (SGr-BI)
0.1 GRASP-PR (SGr-BI-PR)
GRASP (Spl-SBI)
GRASP (Spl-BI)
0
0 100 200 300 400 500 600
Time (in seconds) to target (11.262 MWh/year)

Fig. 6. Metaheuristics comparison.

Table 4
Comparison of ISCPR with GRASP.

Inst BK tmax ISCPR GRASP

Itmax S SBK t It Itmax S SBK t It

BR 758.7 40 s 303 0.044 87 7 56 30 0.003 99 11 9

BU 782.3 40 s 320 0.000 100 2 15 32 0.000 100 7 5
R3 3277.8 40 s 162 0.000 100 5 20 15 0.000 100 2 1
B4 11 262.1 5m 452 0.000 100 25 39 15 0.000 100 77 4
R4 3055.9 5m 314 0.152 28 81 86 7 0.235 1 109 2
AR 21 186.6 30 m 1875 0.233 84 303 315 34 2.494 5 936 18
AU 21 487.8 30 m 1827 1.006 43 337 340 36 3.635 3 879 17
R5 7552.6 30 m 977 0.129 57 484 269 10 0.620 8 903 5
R6 1793.0 30 m 772 0.054 64 292 126 6 0.197 28 890 3
R7 351 958.7 6h 683 1.344 20 7782 244 2 6.606 (3.4) 13 701 1

BK: Best known solution (in kW/year).

tmax, Itmax: Time limit, average iterations executed within time limit.
S: average percentage over best known solution.
SBK: percentage of tests that reach the best known solution (or relative deviation if not reached).
t, It: average runtime (in seconds) and iterations to reach best solution.

solutions; they are not shown in Fig. 6 for this reason. This indicates
that the effectiveness of path relinking is reduced when applied to
local minima after a local search for this problem.
ICPR (SGr-PR) performs slightly faster than the best GRASP
combination. This results from the faster semi-greedy reconstruc-
tion that rebuilds only half of the switches, and the reduced
neighbourhood exploration of path relinking. ISCPR always
reaches the best known solution in less than 2 min, about 25 s
in average. ISCPR is more than two times faster than all other
tested combinations. This is a consequence of the fast execution
of the sample reconstruction and of the guided search of path
relinking over the reduced neighbourhood.
Sample local search reduces the search time of GRASP. Sample
construction also reduces the runtime, but it generates biased
initial solutions and GRASP needs more iterations to compensate
this. On the other hand, sample construction combined with path
relinking (ISCPR) reduces the runtime without loosing quality,
and performs faster than semi-greedy construction combined
with path relinking and than GRASP variants.
Table 4 compares ISCPR and standard GRASP (SGr-BI) for other
instances. It shows the best known solution value (BK), the time
limit (tmax), the number of iterations (Itmax) reached within the
time limit, the average relative deviation from the best known
solution (S), the percentage of tests that reach the best known
solution (SBK), or the relative deviation of the best found solution
in parentheses when the best known solution is not found, the
average runtime (t), and the average iterations (It) needed to
reach a best solution.
ISCPR presents better average results than GRASP and reaches
the best known solutions more often. Within the time limit, ISCPR
iterates at least 10 times more than GRASP. Consequently, ISCPR
finds its best solution faster than GRASP. The better performance
of ISCPR is more evident in large instances such as R7. It finds the
best solution in 130 min and 244 iterations in average for R7,
while GRASP completes only one iteration in an average of
172 min. With the exception of instances BR and R3, ISCPR finds
better solutions faster and with higher probability. In particular,
for large instances with more than 80 lines, it finds the best
 A.J. Benavides et al. / Computers & Operations Research 40 (2013) 24–32 31

known value in about half the execution time with a 13 times The use of path relinking as guided local search can improve the
higher average probability. solutions efficiently.
These components together form an efficient heuristic, ISCPR,
that is capable of reducing the number of reliability estimations,
5.5. Quality of the solutions and is able to perform better than GRASP for this problem.

In this section, we compare the solutions of GRASP and ISCPR

with optimal solutions to validate their quality. The optimal Acknowledgements
solutions were obtained with a branch-and-bound procedure.
Our implementation of branch-and-bound starts from a network This work has been supported by the project ‘‘Metodologia
with no switches installed. The lines are processed in Multiobjetivo para Análise e Melhoria da Confiabilidade de
non-increasing order of the EENS reduction that produces the Sistemas de Distribuic- a~ o de Energia Elérica’’ (Auxı́lio Integrado
installation of a unique switch in that line of the network. For CNPq, CTEnerg processo 554900/2006-8). We also thank Prof. Dr.
each partial solution, it generates two new solutions by allowing Alysson M. Costa (ICMC-USP), Prof. Dr. Vinı́cius J. Garcia (UNI-
or forbidding the allocation of a switch in the processed line. To PAMPA), and two anonymous reviewers for valuable comments.
lower-bound the EENS value of a partial solution, the network is
evaluated supposing that every unprocessed line holds a switch.
Observe that the solutions of the branch-and-bound algorithm are References
only optimal relative to the heuristic service restoration algo-
rithm, since the branch-and-bound searches for the best switch [1] Teng J-H, Liu Y-H. A novel ACS-based optimum switch relocation method.
IEEE Transactions on Power Systems 2003;18(1):113–20.
allocation. [2] Brown RE, Hanson AP, Frank HLW, Luedtke A, Born MF. Assessing the
Due to the complexity of the problem, the branch-and-bound reliability of distribution systems. IEEE Computer Applications in Power
algorithm is able to solve only small instances. Our experiments 2001;14(1):44–9.
[3] Feo T, Resende M, Smith S. A greedy randomized adaptive search procedure
with branch-and-bound install five or 10 switches. Table 5 shows for maximum independent set. Operations Research 1994;42:860–78.
the optimum value and the runtime for instances that were [4] Civanlar S, Grainger JJ, Yin H, Lee SSH. Distribution feeder reconfiguration for
solved by branch-and-bound within a time limit of 16 h. We loss reduction. IEEE Transactions on Power Delivery 1988;3(3):1217–23.
[5] Falaghi H, Haghifam M-R, Singh C. Ant colony optimization-based method for
tested GRASP and ISCPR 100 times for each of those instances placement of sectionalizing switches in distribution networks using a fuzzy
with a time limit of 30 s. GRASP and ISCPR always reach the multiobjective approach. IEEE Transactions on Power Delivery 2009;24(1):
optimal solutions within the time limit. Table 5 shows the 268–76.
[6] Baran ME, Wu FF. Network reconfiguration in distribution systems for loss
average time, the maximum time and the average number of
reduction and load balancing. IEEE Transactions on Power Delivery 1989;4(2):
iterations needed to reach the optimal solution. Runtimes of 1401–7.
GRASP and ISCPR are considerably smaller than branch-and- [7] Augugliaro A, Dusonchet L, Sanseverino ER. Multiobjective service restoration
bound runtime. Since we can only compare with branch-and- in distribution networks using an evolutionary approach and fuzzy sets.
International Journal of Electrical Power and Energy Systems 2000;22:
bound in small instances, the difference between ISCPR and 103–10.
GRASP is less than in Table 4, but ISCPR still finds the optimal [8] Miu KN, Chiang H-D, Yuan B, Darling G. Fast service restoration for large-
solution faster in seven of the 12 instances, confirming the overall scale distribution systems with priority customers and constraints. IEEE
Transactions on Power Systems 1998:789–95.
observations of Section 5.4. [9] Benavides AJ. Service restoration and switch allocation in power distribution
networks: bounds and algorithms. Master’s thesis, Prog. Pós-graduac- a~ o em
Ciência da Computac- a~ o, Federal University of Rio Grande do Sul. URL /http://
hdl.handle.net/10183/31136S; 2010.
6. Concluding remarks [10] Ćurčić S, Özveren CS, Crowe L, Lo PKL. Electric power distribution network
restoration: a survey of papers and a review of the restoration problem.
Electric Power Systems Research 1996;35:73–86.
Our results show that sample construction is efficient and can [11] Sudhakar TD, Srinivas KN. Restoration of power network—a bibliographic
create high quality solutions for the switch allocation problem. survey. European Transactions on Electrical Power 2011;21:635–55.
[12] Zambon E, Bossois DZ, Garcia BB, Azeredo EF. A novel nonlinear programming
A restriction of the neighbourhood by random samples also can
model for distribution protection optimization. IEEE Transactions on Power
speed up the local search, leading to good solutions in short times. Delivery 2009;24(4):1951–8.
[13] Bupasiri R, Wattanapongsakorn N, Hokierti J, Coit DW. Optimal electric
Table 5 power distribution system reliability indices using binary programming. In:
Proceedings of the IEEE annual reliability & maintainability symposium,
Comparison of branch and bound with ISCPR and GRASP.
2003. p. 556–61.
[14] Soudi F, Tomsovic K. Optimized distribution protection using binary pro-
Sw. Inst. B&B ISCPR GRASP
gramming. IEEE Transactions on Power Delivery 1998;13(1):218–24.
[15] Soudi F, Tomsovic K. Towards optimized distribution protection design. In:
Opt t t tmax It t tmax It Proceedings of the 3rd international conference on power system planning
and operations, 1997. p. 354–8.
5 BR 1957.3 1.5 0.02 0.05 8.2 0.04 0.11 1.2 [16] Sohn J-M, Nam S-R, Park J-K. Value-based radial distribution system relia-
BU 2220.0 1.8 0.02 0.08 8.8 0.06 0.22 1.7 bility optimization. IEEE Transactions on Power Systems 2006;21(2):941–7.
R3 6469.3 7.7 0.17 0.89 70.3 0.15 0.79 3.1 [17] Billinton R, Jonnavithula S. Optimal switching device placement in radial
B4 17 057.9 39.7 0.26 1.32 17.7 1.17 6.36 2.2 distribution systems. IEEE Transactions on Power Delivery 1996;11(3):1646–51.
R4 4353.9 4553.0 0.28 0.70 7.0 1.43 2.78 1.1 [18] Chen C-S, Lin C-H, Chuang H-J, Li C-S, Huang M-Y, Huang C-W. Optimal
AR 37 268.5 7484.0 1.68 6.90 60.5 1.46 3.44 1.1 placement of line switches for distribution automation systems using immune
AU 38 288.3 7968.0 1.35 6.22 49.7 5.64 17.05 4.7 algorithm. IEEE Transactions on Power Systems 2006;21(3):1209–17.
R5 11 388.6 55 625.0 1.72 7.30 27.6 3.18 4.61 1.0 [19] Lin C-H, Chen C-S, Chuang H-J, Li C-S, Huang M-Y, Huang C-W. Optimal
switching placement for customer interruption cost minimization. In: Power
engineering society general meeting, 2006. IEEE; 2006.
10 BR 1174.1 342.3 3.83 12.48 131.4 1.00 3.88 3.9
[20] Levitin G, Mazal-Tov S, Elmakis D. Genetic algorithm for optimal sectionaliz-
BU 1285.1 273.3 3.65 13.27 129.5 0.91 4.09 3.9
ing in radial distribution systems with alternate supply. Electric Power
R3 4026.4 12 089.0 2.02 5.68 32.5 0.70 1.96 1.3 Systems Research 1995;35:149–55.
B4 14 235.0 50 362.0 3.52 15.11 22.3 4.93 21.58 1.7 [21] da Silva LGW, Pereira RAF, Mantovani JRS. Allocation of protective devices in
distribution circuits using nonlinear programming models and genetic
Opt: Optimal solution value (in kW/year). algorithms. Electric Power Systems Research 2004;69(1):77–84.
t, tmax, It: Average runtime, largest runtime (in seconds) and average iterations to [22] da Silva LGW, Pereira RAF, Mantovani JRS. Optimized allocation of sectiona-
reach the optimal solution value. lizing switches and control and protection devices for reliability indices
32 A.J. Benavides et al. / Computers & Operations Research 40 (2013) 24–32
improvement in distribution systems. In: Transmission and distribution
conference and exposition: Latin America, 2004. IEEE/PES; 2004. p. 51–6.
[23] Carvalho PMS, Ferreira LAFM, da Silva AJC. A decomposition approach to
optimal remote controlled switch allocation in distribution systems. IEEE
Transactions on Power Delivery 2005;20(2):1031–6.
[24] da Silva LGW, Pereira RAF, Abbad JR, Mantovani JRS. Optimised placement of
control and protective devices in electric distribution systems through reactive
tabu search algorithm. Electric Power Systems Research 2008;78(3):372–81.
[25] Villasanti A, Baran B, Gardel P. Multiobjective allocation of remotely con-
trolled switches in an electric distribution power system. In: Transmission
and distribution conference and exposition: Latin America, 2008. IEEE/PES;
2008. p. 1–7.
[26] Moradi A, Fotuhi-Firuzabad M. Optimal switch placement in distribution
systems using trinary particle swarm optimization algorithm. IEEE Transac-
tions on Power Delivery 2008;23(1):271–9.
[27] Stützle T. Applying iterated local search to the permutation flow shop
problem. Technical Report AIDA-98-04. Darmstadt, Germany: University of
Technology, Department of Computer Science, Intellectics Group; 1998.
[28] Glover F. Genetic algorithms and scatter search: unsuspected potentials.
Statistics and Computing 1994;4(2):131–40.
[29] Kavasseri R, Ababei C. REDS: REpository of Distribution Systems. URL
/http://venus.ece.ndsu.nodak.edu/kavasseri/reds.htmlS; 2009.
[30] Benavides AJ, Machado M, Ritt M, Buriol LS. Problem instances for the switch
allocation problem. URL /inf.ufrgs.br/algopt/repos/power-distribution-in
stances.tgzS; 2011.
[31] Resende M, Ribeiro C. GRASP with path-relinking: recent advances and
applications. In: Ibaraki T, Nonobe K, Yagiura M., editors. Metaheuristics:
progress as real problem solvers. Operations research/computer science inter-
faces series, vol. 32. Springer US; 2005. p. 29–63, ISBN 978-0-387-25382-4.
[32] Aiex R, Resende M, Ribeiro C. TTT plots: a perl program to create time-to-
target plots. Optics Letters 2007;1:355–66.
[33] Hoos H, Stützle T. On the empirical evaluation of Las Vegas algorithms—position
paper. Technical Report, Computer Science Department, University of British
Columbia; 1998.