AbstractThis paper presents a comparative study for four gorithms [1], [3], [5], expert systems [2], data base [4], and
modern heuristic algorithms (MHAs) to service restoration in fuzzy reasoning [7]. However, these methods produce solutions,
distribution systems: reactive tabu search, tabu search, parallel which may not even be sub-optimal. In [6], the Hopfield type
simulated annealing, and genetic algorithm. Since service restora-
tion is an emergency control in distribution control centers to neural network was applied. However, they devoted their efforts
restore out-of-service areas as soon as possible, it requires fast to analyzing the neural network itself and the method itself is not
computation and high quality solutions for customers satisfaction. practical.
The problem can be formulated as a combinatorial optimization Recently, modern heuristic algorithms (MHAs) [8] such as
problem to divide the out-of-service area to each power source. genetic algorithm [9], parallel simulated annealing [10], tabu
The effectiveness of the MHAs is compared against each other on
typical service restoration problems. search [11], [12], and reactive tabu search [13], [14] have been
used for various combinatorial optimization problems including
Index TermsCombinatorial optimization, distribution sys-
tems, genetic algorithm, modern heuristic algorithms, parallel
service restoration in power systems. The authors have devel-
simulated annealing, reactive tabu search, service restoration, oped a parallel genetic algorithm for service restoration and
tabu search. showed promising results on several distribution networks [15].
However, it requires parallel processors and, unfortunately, con-
ventional engineering workstation (EWS) or personal computer
I. INTRODUCTION
(PC) is still utilized for the main computer in practical distribu-
Step 2) Generation of an initial set of states (string Step 2) Generation and evaluation of neighboring states:
population): Generate several perturbation states ( of
Strings are generally generated randomly. the searching point) of the current state ( ) and eval-
However, it is sometimes effective to produce initial uate the energy function at each neigh-
strings using the problem-dependent methods. These boring states. Calculate the energy using the fol-
methods produce sub-optimal initial searching lowing equation:
points in the solution space. Using the sub-optimal
initial searching points sometimes realizes fast (1)
convergence to the optimal solution.
Step 3) Evaluation and selection of each string: Step 3) Generation of the next state:
Strings are evaluated using the fitness function, If the following condition is satisfied:
which represents the tendency of fitness of each
string to the target problem. A good candidate for (2)
the fitness function is the objective function of the or
problem. Basically, the strings that have higher fit-
ness values are selected for the next generation with (3)
high probability. In other words, better searching
points are selected according to their objective func- then the current state is changed to
tion values. The original GA utilizes roulette wheel
selection [9]. However, improved methods such as (4)
stochastic sampling with and without replacement
[9] and elite strategy have been proposed. Update the temperature and repeat steps 2 and 3 until
Step 4) String operations (generation of neighboring the convergence criterion is satisfied.
states): Tabu Search [11], [12]: TS is based on the use of prohibi-
GA performs string operations such as crossover tion-based techniques and basic heuristic algorithms like local
and mutation. The operations produce new searching search. Therefore, the main advantage of TS with respect to con-
points using the current searching points. Crossover ventional GA and SA lies in the intelligent use of the past his-
generates two new searching points from the two tory of the search to affect its future search procedures. Since
current searching points. Mutation generates one the method utilizes a tabu list for storing the past history of the
new searching point from one current searching search, the efficient structure of the tabu list is important for fast
point. Natural selection and string operations are re- computation. The procedure of TS can be expressed as follows:
peated until the strings are converged to the optimal Step 1) Initialization:
or suboptimal solution. Give the initial state, searching point and put
The details on GA could be found in [9]. the current state into the tabu list.
Parallel Simulated Annealing [10]: Parallel simulated an- Step 2) Generation and evaluation of neighboring states:
nealing (PSA) is an extension of simulated annealing (SA). Al- Generate all of possible neighboring states and
though SA is an attractive optimization technique, small state check whether the states are tabu or not.
transition often makes SA gets stuck in local minima in cases Step 3) Generation of the next state:
where the solution space of the target problem is large or con- Move the current state to the next state that is not
straints of the target problem are strict. In order to overcome tabu and have the lowest objective function value.
the problems, PSA parallelizes the routines of state transition repeat step 2 and 3 until the convergence criterion is
in the original SA to obtain better searching points efficiently. satisfied. The details of the method can be found in
PSA basically utilizes one searching point like SA. However, it [11], [12].
generates multiple neighboring states rather than a single state Reactive Tabu Search [13], [14]: The conventional modern
like SA. It allows to find out a solution near a global minimum heuristic methods like GA, SA, and TS require adjusting
over a wide range due to handling a set of solution candidates. search parameters for efficient search. However, in general,
The characteristics of PSA are summarized as follows. the appropriate parameter values depend on each problem.
Therefore, a parameter tuning problem is known as one of
a) PSA has a distinct possibility of reaching a optimal solu-
the disadvantages of the MHAs. The Reactive Search (RS)
tion due to multiple searching points.
framework proposes the introduction of feedback (reactive)
b) PSA has better convergence characteristics because of se-
schemes in heuristics for discrete optimization problems [13].
lecting the best state among solution candidates.
RTS is one of the RS methods and it has feedback-based
The PSA algorithm can be written as follows: tuning mechanism of tabu length (TL) and automated balance
Step 1) Initialization: mechanism of diversification and intensification. RTS stores
Give the initial state, searching point and tem- all searched states. After one move is performed, the algorithm
perature and evaluate the energy function , checks whether the current searching point has already been
which is generated by the objective function of the found. TL increases if a searching point is repeated, while TL
original problem. decreases if no repetitions occur during a sufficient long period.
TOUNE et al.: COMPARATIVE STUDY OF MODERN HEURISTIC ALGORITHMS 175
(6)
where
Fig. 1. A distribution system model. voltage drop at node n;
equivalent impedance at node .
Step 3) Calculate voltage (forward calculation).
This adjusting mechanism allows us to escape from the local
Voltage at each node can be calculated from the
valley. Moreover, since the basic TS mechanism cannot avoid
power source using the following equation:
long search cycles, RTS introduces the escape procedure. It
consists of a number of random steps starting from the current (7)
searching point.
Effective search in the solution space requires balance of di- where
versification and intensification. GA realizes the balance by the voltage at node n;
string operations and the selection mechanism, and sometimes node that is next to node at the upstream side.
requires more effective local search procedure. PSA realizes The voltage and current calculation method is appropriate
parallel search by several conventional SA search procedures. when the target distribution system is always operated in radial
However, it requires parallel processors for practical speedup. network structures, like in Japan. However, if it is required to
TS realizes the balance between diversification and intensifica- consider a network with loops, we have to utilize backward for-
tion using a tabu list. RTS, moreover, strengthens the mecha- ward power flow calculation methods which can handle loops
nism using reaction and escape mechanism. [16].
Objective Function and Constraints: The most important
III. PROBLEM FORMULATION OF SERVICE RESTORATION objective of the service restoration is to restore as many cus-
Distribution System Model: The following assumptions for tomers as possible. In many fault cases, it is enough to energize
distribution systems are usually made for practical application the out-of-service area from the only one neighboring power
of service restoration. source. However, it is necessary to utilize several power sources
a) Power source can be formulated as current injection next to the large out-of-service area. In such a case, service
source. restoration can be formulated as one of the graph partitioning
b) Voltage at the power source is known. problems to divide the out-of-service area to corresponding
c) An area surrounded by switches is called a section. Each power sources. The objective function is to counterbalance
section has a concentrated load and each load can be for- spare capacity of each power source considering next faults
mulated as constant contracted current. and maximize the minimum voltage of the network considering
d) Each section line impedance can be calculated as an customers satisfaction in the large out-of-service area. The
equivalent scalar impedance using load power factor and function can be expressed as follows:
line constants.
According to the above assumption, distribution system can be (8)
expressed as shown in Fig. 1.
Circuit Calculation Method: In order to calculate voltages
where
and currents in the target distribution system correctly, load flow
number of power source;
calculation is necessary. However, the following backward and
spare capacity of source ;
forward sweep circuit calculation is usually performed for the
average of spare capacity of all sources;
sake of fast service restoration [15].
minimum voltage of the target network;
Step 1) Sum up total load currents (backward calculation). coefficients for each term.
Sum up total load currents from the end of the The following are constraints, which should be considered for
branches to the power source according to the fol- practical service restoration.
lowing equation. a) Radial Network Constraint: Distribution network
(5) should be composed of radial structure from an operational
point of view. Therefore, each section has only one up-stream
where section.
total load currents at section ; b) Power Source Limit Constraint: The total load ca-
contracted load currents at section ; pacity of each partial network cannot exceed the capacity limit
node that is next to section at the downstream of the corresponding power source.
side.
If the node is the end of feeder, is (9)
0 (A).
176 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 17, NO. 1, JANUARY 2002
where
number of load for power source ;
capacity of power source ;
capacity of load energized by power source .
c) Voltage Constraint: Voltage magnitude at each section
must lie with their permissible ranges.
(10)
where
allowable minimum section voltage;
allowable maximum section voltage;
voltage at load section .
d) Current Constraint: Current magnitude of each branch Fig. 2. An example of radial distribution system and its representation using
the proposed state representation method.
(switch and line) must lie with their permissible ranges.
(11)
is load no. 1. Therefore, the first two contents of the array is
where 9 and 1.
allowable maximum load section current; Generation of Initial Searching Point: Initial states, namely
current at load section . initial network configurations in the service restoration
Constraints (a) can be checked using a search method. The ob- problem, can be generated by various methods such as random
jective function value and constraints (b)(d) can be checked and problem-dependent methods. This paper proposes the
using the above-mentioned circuit calculation method. following problem-dependent procedure considering efficiency
of fast convergence to the global optimal state.
IV. PROBLEM FORMULATION USING MODERN Step 1) Select a certain load , statistically. The load
HEURISTIC ALGORITHMS should be next to the load sets energized from the
Problem formulation using MHAs includes the following current power source and it has not determined its
items: power source yet.
Step 2) Determine statistically whether the power source
1) representation of state variable;
supplies power to the load or not using the fol-
2) generation of initial states;
lowing probability, . The larger the spare ca-
3) generation of neighboring states.
pacity (SC) of source is, the larger can be.
The above formulation should be made clear explicitly because
Here, minimum value of is set to and
the efficiency of each method depends on the formulation. In
maximum value of is set to .
other words, the comparison of each method is only based on
the formulation as mentioned below. (12)
Representation of the State Variable: Each MHS requires
to store state variable (searched points) for various purposes. where
Therefore, the representation method for the state variable is spare capacity of source ;
one of the key issues for applying MHS to a certain problem. capacity of source ;
The above-mentioned methods generate neighboring states of maximum probability of ;
the current searching point by various methods. Therefore, it minimum probability of .
is necessary to consider whether the neighboring states can be Step 3) If every load has its power source, exit. Otherwise,
generated easily and effectively or not, using the representation go to Step 1.
of the state variable. Conventionally, a network has been repre- Step 4) Convert the obtained network configuration to an
sented by the switch states [17]. However, this paper proposes array. The above method can generate sub-optimal
the following method to represent the state variable considering solution of the problem and it can be an efficient
the whole searching procedures of each method. initial point in the solution space.
(Representation Method): In other words, each power source determines energizing loads
The length of an array equals to the number of loads in by tracing connection of the target network and using .
the out-of-service area. is gradually decreasing when the number of loads con-
Numbering all of nodes including power sources and nected to the target power source is increasing.
loads. Modification of the Current Searching Point (Network Con-
Each array position represents the upstream load or figuration): Modification of the current searching point means
power source section number of each position. modification of the network configuration in the service restora-
Fig. 2 shows an example of radial distribution system and its tion problem. The modification method of the network config-
expression using the above method. For example, load nos. 1 uration is different among the MHAs. Each method generates
and 2 are energized from power source no. 9 in the figure. The the candidates for the next states, evaluates the candidates, and
upstream of load no. 1 is source no. 9 and that of load no. 2 selects the next state among the candidates.
TOUNE et al.: COMPARATIVE STUDY OF MODERN HEURISTIC ALGORITHMS 177
TABLE I
COMPARISON OF OBJECTIVE FUNCTION VALUES BY RTS, TS, GA, AND PSA
REFERENCES
Fig. 8. Comparison of the average objective function values to various initial
tabu lengths by TS and RTS. [1] K. Aoki et al., Outage state optimal load allocation by automatic
sectionalizing switches operation in distribution systems, IEEE Trans.
Power Delivery, vol. PWRD-2, pp. 11771185, Oct. 1987.
fault. It can restore 97% of the de-energized area even in the [2] C. C. Liu et al., An expert system operational aid for restoration and
transformer fault which can be considered as one of the severe loss reduction of distribution systems, IEEE Trans. Power Syst., vol. 3,
pp. 619626, May 1988.
faults. The results indicate the efficiency of the RTS-based [3] K. Aoki et al., Voltage drop constrained restoration of supply by switch
method in practical operation. operation in distribution systems, IEEE Trans. Power Delivery, vol. 3,
pp. 12671274, July 1988.
[4] C. E. Lin et al., A distribution system outage dispatch by data base
VII. CONCLUSION method with real-time revision, IEEE Trans. Power Delivery, vol. 4,
Jan. 1989.
This paper has compared modern heuristic algorithms: ge- [5] K. Aoki et al., A new algorithm for service restoration in distribution
netic algorithm, parallel simulated annealing, tabu search, and systems, IEEE Trans. Power Delivery, vol. 4, July 1989.
reactive tabu search, for service restoration in distribution sys- [6] C. Fukui et al., Switch pattern planning in electric power distribution
systems by Hopfield-type neural network, in Proc. IJCNN, vol. II-591,
tems. The results can be summarized as follows: Washington, DC, 1990.
a) RTS can generate only local neighboring states at moving [7] Imamura et al., An application of fuzzy reasoning for service restora-
tion (in Japanese), Trans. IEE Jpn., vol. 113C, no. 5, May 1993.
to the next states. However, the simulation results for [8] C. R. Reeves, Modern Heuristic Techniques for Combinatorial Prob-
typical distribution systems and a practical distribution lems. Oxford, U.K.: Blackwell Scientific, 1993.
system indicate that RTS can generate the most highly [9] D. E. Goldberg, Genetic Algorithms in Search, Optimization, and Ma-
chine Learning. Reading, MA: Addison-Wesley, 1989.
qualified results. [10] H. Mori et al., Parallel simulated annealing for power system decom-
b) The computation time by RTS depends on the number of position, IEEE Trans. Power Syst., vol. 9, May 1994.
evaluated neighboring states at each iteration. According [11] F. Glover, Tabu search: Part I, ORSA J. Computing, vol. 1, no. 3,
Summer 1989.
to the results, RTS can realize the fastest computation es- [12] , Tabu search: Part II, ORSA J. Computing, vol. 2, no. 1, Winter
pecially for practical service restoration. Therefore, RTS 1990.
is the best method for service restoration. [13] R. Battiti, The reactive tabu search, ORSA J. Computing, vol. 6, no. 2,
pp. 126140, 1994.
If the out-of-service area can not be restored only using the [14] , Reactive search: Toward self-tuning heuristics, in Keynote Talk
power sources next to the area, multistage switching is required at Applied Decision Technologies, Brunel, UK, Apr. 1995.
to increase spare capacity of the neighboring power sources [15] Y. Fukuyama and H. D. Chiang, A parallel genetic algorithm for ser-
vice restoration in electric power distribution systems, in Proc. IEEE
[18]. Multistage switching is a large combinatorial problem and FUZZ/IFES Conference, Yokohama, Japan, Mar. 1995.
we have much knowledge on the problem. Therefore, expert [16] D. Shiomohammadi et al., A compensation-based power flow method
system (ES) is suitable for the problem. On the contrary, decom- for weakly meshed distribution and transmission networks, IEEE
Trans. Power Syst., vol. 3, pp. 753762, Apr. 1988.
position of out-of-service area can be formulated as a combina- [17] K. Nara et al., Implementation of genetic algorithm for distribution
torial optimization problem and it is suitable for RTS as shown systems loss minimum re-configuration, IEEE Trans. Power Syst., vol.
in this paper. We plan to develop a practical service restoration 7, pp. 10441051, Aug. 1992.
[18] Y. Fukuyama et al., A hybrid system for service restoration using expert
using ES and RTS as one of the future works. In such a case, the system and genetic algorithm, in Proc. ISAP 96, Jan. 1996.
role of ES is to move the initial configuration inside the feasible [19] Y. Fukuyama et al., Application of meta-heuristic techniques to service
region in the solution space [18]. Application of the proposed restoration in power distribution systems (in Japanese), in Proc. 6th
Intelligent System Symp., Oct. 1996.
method with the function of multistage switching to larger ser- [20] Fudo et al., A genetic algorithm for network reconfiguration using
vice restoration problems is one of the future works. three phase unbalanced load flow, in Proc. ISAP97, July 1997.
It is important to consider load priority and reliability and
switching complexity in the service restoration. For example,
one of the ways to handle the load priority and reliability is to Sakae Toune received the B.S. degrees in electrical engineering, in 1978, from
utilize weighted penalty terms in the objective function. It is also Shinshu University, Nagano, Japan.
important to consider switching complexity. One simple way He joined Kansai Electric Power Company, Hyogo, Japan, in 1978 and is cur-
rently working as a senior engineer at Technical Research Center. His research
is to minimize the difference of switching states between the interests include distribution system operation and equipment.
post-fault network and the target restored network. Another way Mr. Toune is a member of IEE of Japan.
TOUNE et al.: COMPARATIVE STUDY OF MODERN HEURISTIC ALGORITHMS 181
Hiroyuki Fudo graduated Sumoto-jitugyou High School and joined Kansai Yoshikazu Fukuyama (M90) received the B.S., M.S., and Ph.D. degrees in
Electric Power Company, Hyogo, Japan, in 1966. electrical engineering in 1985, 1987, and 1997, respectively, from Waseda Uni-
He was a lecture audience of Kyoto University from 1971 to 1973. His re- versity, Tokyo, Japan.
search interests include distribution system operation. He has been working at Fuji Electric Company R&D Japan from 1987. He
Mr. Fudo is a member of IEE of Japan. was a visiting scientist at Cornell University from 1993 to 1994. His research
interests include application of intelligent systems such as expert system, neural
network, and evolutionary computation techniques to power systems and power
system analysis including voltage stability and load flow. Dr. Fukuyama is a
member of IEE of Japan.