IEEE Transactions on Power Systems, Vol. 9, No.

4,November 1994

1927

GENE1'1C ALGORITHMS IN OPTIMAL MULTISTAGE DISTRIBUTION NETWORK PLANNING Vladimiro Muanda J. V. Ranito L. M. Pmenqt INESC Inst. de Engenharia de Sistemas e Computadores and FEUP/DFdiC Faculdade de Engenharia da Univ. do Port0 L. Mompilher - 4000 Porn - PORTUGAL

Fax:351.2.318692 a n : [email protected]

A 4 - This paper describes a genetic algorithm approach to ttrc optimal multistage planning of distribution networks. Tlw authors describe a mathematical and algorithmic model that thcy havc developed and experimented with success. The paper alw prcsents application examples, with real size systems. Thc advantages of adopting this new approach are discussed in Lhc planning context, namely in conjunction with decision making methods. the adoption of rr~ullicriteria

Can genetic algorithms adequately model the DPP problem? Is it possible to develop efficient computing algorithms for real size networks? Has the genetic approach an easy interface with multicriteria decision making procedures ? Do genetic algorithms show any advantages over previous proposed approaches ?

I "
For opt imol distribution network planning, many mathematical wodcls have been proposed in the past - for 2,4,5,71 are static instance, scc rcl'u 11 - 91. Many of them [l, models, giving on optimal solution for a fixed set of data and a single time prr id. Essays in dynamic planning, considering the evolution in power demand through time and consequent topological changcs in the networks (new branches, new substations, rcinforccmcnts, etc.) have never been a definite success, whcn applicd to real sized networks. The mathematical models behind thosc proposals either were heavy or neglected several projcct Ccatures that engineers take as important on disuibution systcm dcsign; for example, economies gained from cables in the same trench, reuse of disclassified lines, reshaping the whole network by introducing a new substation, etc. Furthermorc, the vast majority of the model proposals, on the last 15 years, was aimed at a so called "Optimal solution". Howcver, during the last few years the objective of rcaching this optimal concept has been challenged more and more, namely within the US, with the acceptance of principles of the "least cost planning" approach . This evolution favors the option for multi-criteria methods and for algorithms giving, as an answer, a large set of possibly good solutions, instead of a single optimum. Genetic algorithms share precisely this property, and therefore an invcstigation has been made on their behavior in the distribution network design or distribution planning problem (DPP).This research has been directed into answering the following questions :
A paper recommended and approved 94 WM 229-5 PWRS by the IEEE Power System Engineering Committee of the IEEE Power Engineering Society for presentation at the IEEE/PES 1994 Winter Meeting, New York, New York, January 30 February 3, 1994. Manuscript submitted July 12, 1993; made available for printing December 6, 1993.

As the following sections report, the answer is positive to all these four questions. In fact, differently from many models previously proposed, this new model is flexible enough so that many realistic features and conditions of practical nature may be taken care of. For instance:

the inclusion of multiple feeders in the same trench, with the related savings achieved; the possibility of disclassifying lines at some stage, and of re-using them at a later stage; the possibility of generating solutions with some open loops, which may be valuable from a reliability point of view; considering multi-objectives; dealing naturally with load diversity factors; no need to specify, in advance, at which year should a substation be built or put into service.

C ALGenetic Algorithms (GA) are search and optimization methods based on natural evolution [Golberg89]. They consist on a population of bit strings transformed by three genetic operators: selection, crossover and mutation. Each string (chromosome) represents a possible solution for the problem being optimized and each bit (or group of bits), represents a value for some variable of the problem (gene). These solutions are classified by a evaluarion function, giving better values, or fitness, to better solutions. Although there are many forms [Grefestette91] for Genetic Algorithms, we will only refer to the canonical algorithm. This means that we will be dealing with three genetic operators (selection, crossover and mutation) and linear, binary, fixed-size chromosomes. Canonical GA use a fixed-size, non-overlapping population scheme and each new generation is created by the selection operator and altered by crossover and mutation. The first population is generated at random. Genetic Algorithm comDonentS Each chromosome represents a potential solution for the problem to solve and must be expressed in binary form. For instance, if we want to maximize the function f ( X ) = X Z , in 1 1, we could simply code X in the integer interval I = [8,3

0885-8950j94/$04.000 1994 IEEE

1928

binary base, using 5 bits. Each solution must be evaluated by the fitness function to produce a value. In our example, the chromoscjmc 11011 would receive the fitness value 272=729, while the chromosome 00111 would receive the value 72=49. The pair (chromosome, fitness) represents an individual. The selection operator creates a new population (or generation) by sclccting individuals from the old population, biased towards thc best. This means that there will be more copies of thc k ' s t individuals, although there may be some copies of thc worst. This operator can be implemented in a variety of ways, although we use here a technique known as Stochastic Tournanic'nt [Goldberg91]. This implementation is suited to a luturc distributed implementation and is very simple: evcry tinic we want to select an individual for rcproduction, we chtwse two, at random, and the best wins with some fixed probability, typically 0.8. This scheme can be enhanced by using more individuals on the competition [Goldberg91 c J or cvcn by considering evolving winning probability, cvcnlually leading to Bolttman Tournament [Goldberg91],gi'ncralizing the Simulated Annealing paradigm [Kirkpatrick83I.
Crossowr is the main genetic operator and consists in swapping chroinosome parts between individuals. The simplest crossover opcralor is implemented by selecting a random crossover point in thc chromosome, and swapping the genes that reside bclwecn the crossover point and the end of the chromosornc. For example, if we have two individuals: A=O 10 I 00 B=010 I 11 and choose a crossover point C=3 (indicated by '1') the resulting individuals aftcr crossover would be: A'=010 I 11 B'=010 IO0

Obviously, there is the need for some bookkeeping functions, for statistics and so on, but they are not central to this explanation. This very simple behavior hides a powerful processing, done by the GA. In fact, the combination of selection and crossover leads to a proliferation of individuals that possess small, tightly coupled blocks of bits leading to good performance. These blocks, usually called s c h e m a t a [Holland75],are replicated through selection and combined or separated by crossover. And mutation, what is its job? Mutation works as a kind of "life insurance". Some important bit values (genes) may be lost during selection; mutation can bring them back, if necessary. Nevertheless, t o o much mutation can be harmful: a mutation probability of 0.5 always leads to random search [Goldberg89], independently of crossover probability.

So, GA tends to select individuals with good performance and recombine some of their building blocks, creating more and more copies of good schemata, simply by the use of selection and crossover. This hidden processing is called implicit parallelism because the number of schemata processed in each generation is typically O(N3),being N the population size [Holland75]. This compares very well with the number of fitness function evaluations, N. This characteristic is distinctive of Genetic Algorithms rGrefenstette911.
Genetic algorithms in Power Svstems There have been some attempts to apply Genetic Algorithms to solve problems in the Power System area, but so far, to our knowledge, not in multi-stage distribution( planning. In references [16] to [22] one may find the reports of attempted approaches to problems such as:

Crossovcr is not performed on every pair of individuals, its frequency bcing controlled by a crossover probability. This probability should have a large value, typically Pc=0.8. The last genetic operator is mutation and consists in toggling a random bit in an individual. This operator should be used with some care, with low probability, typically Pm=O.OOl, for normal populations. How does a Gcnclic Algorithm work?
A canonical GA is a very simple process: we first generate a random initial population, evaluate it and start creating new populations by applying genetic operators. This high-level behavior can be depicted on the following piece of pseudo-C:
m i n ()

'

clustering and network reduction [ 161; optimal capacitor placement [17]; voltage optimization [ 181; harmonics [19]; system observability [20]; reactive power control [21]; load flow analysis [22];

I
i n t gen; generate (oldpop) ; f o r (gen = 0; gen < MAXGEN; gen++)

This section presents a model to solve the problems of the optimal sizing, timing and location of distribution substation and feeder expansion, using genetic algorithms. The model allows the inclusion of constraints related to network radiallity, voltage drops and reliability assessment. The objectives for distribution system planning that will be discussed are related to providing the designs and associated implementation plans necessary for an orderly expansion of facilities, minimizing new facility installation costs and operation costs, as well as achieving an acceptable level of reliability, under the following constraints : a) Operation of the networks under radial configuration (although some open loops may exist);

I
e v a l u a t e (oldpop) ; = select (oldpop) ; crossover (newpop); mutation (oldpop) ; oldpop = newpop;
newpop

1 1

1929

b) Voltagc drop constraints; c) Powcr dctnand specifications; (9 Power llow availability, namely constrained by line thermal 1ill1 i ts; e) Possiblc site location for substation and lines prespecilicul.

Facilily installation costs will be divided in three elements : s i h l i h n cost, substation capacity expansion cost and new fecclc-r cost. Power losses in the network will be taken as operation cc)>is. Already existing elements (substations or feeders) arc iiiiludcd at no investment cost in the model; however, thcir Iwwcr losses are taken in account. The following assumptions are made:

m e coding The direct coding of the above variables into a chromosome has been tried and tested. Although the results obtained were satisfactory, the process was not very efficient, because the extremely large number of unfeasible solutions appearing at each generation led to a large computing time before reaching an acceptable stability. Therefore, we devised a new coding process. Its requirements were: it should lead to a minimum amount of unfeasible solutions genera^, and it should provide very fast decoding, as this operation is required at every fitness evaluation. Its characteristics are:
a) at one time stage, some nodes mat be assigned with load

a) A pcak load is considered for each stage expansion planning (load lorwasting is out of the scope of this paper). Some extra inlormation about load curves (such as load factor) is required, in ordc-r to evaluate losses and assess reliability indices such a.\ avcragc load disconnected or average annual energy not supplic'd. b) NCH'inslallation facility candidates are known beforehand, arid lhcir location and installation costs estimated.
An cxpansion slrategy will be driven by load growth. The planning pcriod will be divided into several stages (one year, for institncc). One aims at having, as a result, a list of investments to bc madc at each stage.

values, while other may have a forecasted zero load; b) a node with a positive load must be connected to, at least, one feeder, while nodes with zero load may either have no connections or have one or more connections in order to allow power flow to other nodes; c) each node is represented in the chromosome by a number of bits needed to encode the number of possible connections to it - e.g., if four lines connect to the node, one needs two bits, if it is a positive load node, o r a three bit string, if it is a zero load node (to include the case where no line goes through that node); (0 a substation is assumed as a special type of line, that may be connected to the nodes chosen as possible building sites; This encoding strategy is such that it gives, as a result, only solutions with a number of lines=(nodes-1). a radiallity condition. A few other details were included in the model, which are not relevant to its understanding but helped in gaining computer efficiency. Fitness Function T h e fitness function must reflect both the desired and the unwanted properties of a solution, rewarding the former and strongly penalizing the latter. In the DPP,desired properties are, for inktance, low cost and high reliability, while unwanted features are non-radial configurations (open loops are accepted, but not closed loops), violations of thermal cable limits or of voltage drop constraints. Fitness is evaluated a posteriori; therefore it may be non-linear, non-continuous, non-convex, whatever. This is very advantageous over strict mathematical programming approaches. The general trend is to maximize fitness. Figure 1 presents the general scheme of evaluation of the fitness of a solution, represented by a chromosome, at any generation. The functions g(), h() and v(), referred to in this figure, must be chosen so that, for no matter what solutions x are evaluated, me always obtains

The gcnctic algorithm approach to the DPP is drawn under the following general lines:
1. A sct or variables is chosen to represent a multi-stage network solution; these variables are encoded in a chromosome. 2. A genctic algorithm is applied to a family of solutions, giving birth U, new generations. 3. Each solution in the new generation is evaluated through a fi mess function, that includes investment costs, power loss costs, reliability, voltage drop deviations; non compliance with other constraints is dealt with by the fitness function - an unfeasible solution, with a low fitness value, will hardly survive. 4. At the end of the process, a family of well fitted plans is available.

Variablcs: For a m-stage planning problem, the following (0-1) integer decision variables could be defined: Fis : Fis = 1 : if feeder i is used at stage s Fis=O : otherwise Sis : Sis = 1 : if substation i is used at stage s Sis = 0 : otherwise Eis : Eis = 1 : if subst. expansion i is used at stage s Eis = 0 : otherwise Existing facilities may be considered by fixing their respective valucs LO 1.

In function f, we have included the following:

1930

clzromosomc decoding

chcck coniicclivity and radiality

corn. to
sources)

uc h i i d Flow

h(PNS)

3) else, if at all nodes the voltage drop stays below that lower limit, the voltage quality index receives 0 value. the reliability RB of a solution is evaluated through the approximate calculation of the expected annual energy not supplied - this calculation takes in account the existence of open loops in the network; we have used the following scheme: 1) an upper bound U, in reliability level is calculated assuming that no switching devices are included in the network therefore, disconnections take place d y at the substation; i s calculated assuming that all 2) a lower bound branches are equipped with switching devices, allowing the isolation of failed branches and service restoration (taking in account the line capacities); 3) reliability fitness is given by RI3 = uu, +(1-ct))L, where a E [0,1] is an "improvement coefficient" that aims at simulating the effect of a compromise odution in switching device location policy - as this, in itself, is a very complex problem [23].

The fitness value f of a solution x i s given by where f(X) = M - Cl(IC+PL) - c ~ V Q c ~ R B M - Large (enough) constant value; Ci - constants externally fixed.

FITNESS = f (power loss cost, volt. drop quality,

Other swial features a) the DC load flow calculations are used as a first filter for unfeasibilities, and are performed at the same time as the structure of the network is recognized - this strategy proved very efficient; b) for AC load flow calculations, we used the method described in [24], applied to balanced systems, as it was specially developed for radialnetworks; c) before load flow calculations, loads values t o be used may be affected by coefficients to take in account the diversity in demand load curves - this feature also distinguishes this GA approach f r o m other models.

m
Fig.1 - Fitncss function evaluation scheme

investment costs IC are added for all the planning stages, considering a fixed discount rate; power loss costs PL are treated in the same way; voltage quality VQ may be assessed in many ways we have used the following scheme: 1) if thc volbge drop at any node exceeds some threshold (e.g., 8%), the solution has been taken as unfeasible and its fitness evaluated by function v; 2) else, if the voltage drops at some nodes lie within a givcn inlcrval (e.g., [5%,8%]),the voltage quality index of Lhe solution is assessed by the sum of Lhe squarc dcviations to the lower limit of this interval;

The GA approach was applied to the system shown in Fig. 2, where solid lines represent existing cables in the initial radial system, and dotted lines represent possible sites for the expansion of the system. A complete data listing may be obtained from the authors, by request. The proposed substation sizes and other complementary system data are shown below Number of time stages Discount Rate Nominal voltage Voltage thresholds No. of nodes No. of branches - total No. of potential branches Total load for each time stg. (MVA) Feeder cost (PTE)

3 (+ initial) 10% 15 kV 5%, 8% 3 x 50 3x64 3 x48 45; 63;79.5 4* 106/km

1931

s2

do Km

Fig. 3. Solution for Stage I

Fig. 2 Initial system

s2

Fig. 4 Solution for Stage 2

Fig. 5 Solution for Stage 3

Run - Several runs were made. The results presented were obtained for a population of 40, and for 300 generations (a value for which a robust stability in the "best-so-far" solution has been consistently observed). On average, user run times have been of around 300 sec., in a NeXTStation (33 MHz), but we are now aware that these times may still be largely improved by carefully revising the computer code.

1932

ENS (MWh)

330

CONCLUSIONS

300 -270

i,
\

-240 -210

The research reported in this paper clearly demonstrates that a GA approach b a dynamic multi-stage planning problem is both feasible and advantageous: It provides the planner with a set of time-ordered investment decisions which is not obtained from static suboptimizations, but directly from the considemtion of the time dimension of the problem. In this respect, it is more complete than many publisha apprmhes which claim feasibility under constrained computing environments.
9(X)

-850
950 lo00

Furthermore, it allows the representation of non1050 linearitks which are hard to include in pure mathematical programming methods, in fact, the existence of non-linearities COST ("EXlo6) h e advantages of using GA against pure mathematical enhances t Fig. 6 - PIOI or a I-aIuily of solutions i n a bi-attribute space programming. These non-linqities arise not only from the ' (invcstment vs. reliability). non-linear character of objective functions and constraints but also from the discrete nature of many aspects of the distribution planning problem. These in some cases could lead to a nonconvex domain, perhaps in some cases not even connex - but GA are able to deal with such environments and can detect local Results minima or even "islands" of solutions. In I'jprcs 3-4-5, a substation symbol within a square The results of a GA are a generation of solutions, means that hcrc has been a reinforcement in capacity of that filtered through the struggle for survival. Therefore, many substation; shadowcd lines represent feeders already installed that are not uwd in a given stage, remaining, however, as open interesting and valuable exercises on comparisons and trade offs loops. See, for instance, branches 33-34-35-36, built in stage may be executed, helping the planner to gain insight on the 1; in stage 2,3435-36 iu# unused, but in stage 3 they are put problem he is faced with and Jlowing field for better decisions in active scrvicc whilc 33-34 is flow open. N o t i c e that node 35 to be taken. has no forccnsicd load until stage 3; nevertheless, a solution The fact that many solutions will be available also was built through it in d e r to reach node S3,right in stage 1, which had a positive load value attached. Notice also the enhances the opportunity for multicriteria methods to be explicitly applied, which is, in our point of view, a step rearrangemcnl in the network when a new substation gets in. towards an adequate direction in distribution planning. Figure 6 presents a family of solutions, obtained at the end of the evolution process, in a bi-attribute space REFERF,NCES (investment, in Portuguese Escudos, vs. reliability, in average annual energy not supplied, both to be minimized). Solutions Y. Baklund, J.A. Bubenko, Computer-aided distribution for which thcrc exists an alternative with less investment and, system phning", Electrical Power & Energy System, at least, no incrcase in EiNS, or a more reliable alternative with, vol.1, no.1, Apr 1979 ai least, no incrcase in cost are said t o be dominated.
I'

We have drawn a curve representing a convex approximation to the non-dominated border; the black dot, pointed by the arrow, stands for the solution depicted in figs. 25, and the thrcc shadowed dots represent some special nondominated soluiions: they would not be found by classical mathematical programming methods such as the weighted addition of thc functions representing the objectives nevertheless, thcy may be thought of as good compromises, by h e planner. Onc may clearly see how few the non-dominated solutions (also callcd Pareto optimal) are in this example which, in principle, are the interesting solutions t o look at. This clearly indicates that the combination of genetic o be a algorithms Ibllowcd by multi-criteria screening is likely t powerful tool in dccision aid for network planning. We believe that there is ficld hcre for further research.
I

M.F. Oliveira, V. Miranda, "Etudes d'optimisation dans les rdseaux de distribution comprenant des calculs de fiabilitd", Proceedings of CIRED'79, s.6, Li&ge,Belgium, Apr 1979

R.N. Adams et al., "A methodology for distribution system planning", Proceedings of the 8th PSCC, Helsinki, Finland, Aug. 1984
M.A el-Kady. "Computer aided planning of distribution substation and primary feeders", IEEE Transactions on P A S , vol. 103, Jun 1984

T. Burkhardt et al., "Decision making including forecast uncertainties and optimal routing in distribution networks", Proceedings of CIRED'85, s.6, Brighton, U.K., Apr 1985

1933

[6] T. Gijncn, I .J. Ramirez-Rosado, "Review of distribution planning modcls: a model for optimal multistage planning", TEEE Proceedings, Vol. 133, Pt. C, NO. 7, Nov 1986
[7]

[19] H. Yang, "Worst case analysis of distribution system harmonics using genetic algorithms", Proc. IEEE SOUTHEASTCON '92 New York, NY, USA, IEEE 1992 [20] H. Mori, "A genetic approach to power system topological observability", Proc. IEEE International Symp. on Circuits and Systems, New York, NY,USA, IEEE 1991 [21] K. Iba, "Reactive Power Optimization by Genetic Algorithm", PICA93, Phoenix, AR, USA, May 1993 [22] X. Yin, N. Germany, "Investigations on Solving the Load Flow Problem by Genetic Algorithms", Electric Power Systems research,22 (1991). 1991 Elsevier [23] V. Miranda ,"Using fuzzy reliability indices in a decision aid environment for establishing interconnection and switching location policies", Proc. CIREDP1, s.6, Libge, Belgium, Apr 1991

K.Aoki cl ; i l . , "Ncw approximate optimization method for distrihut ion systcm planning", IEEE Transactions on PWRS, vol 5 , no.1, Feb 1990

181 N. Kagan, K.N. Adams, "Application of Benders decompo4 lion tcchnique to the distribution planning problcni", I'rocccdings of the 10th. PSCC, Graz, Austria, Aug 1990, cd. Butterworths, London
[9] K.Nara ct al., "Multi-year expansion planning for distribulion systcms", IEEE Transactions on PWRS,vol 6, no.3, Aug 1991
[lo] D. E. Goldhcrg, "Genetic algorithms in Search, Optiini;/aiion and Machine Learning", 1989 AddisonWesley

[I 11 J. Grclcnwttc, "Conditions for Implicit Parallelism", Navy Ccnlcr lor Applied Research in Artificial

[%I V. Miranda, F.M. Barbosa, "Three phase load flow for

radial networks', Proc. of MELECON'83, rp. D4.07, Athens, Greece, 1983

Intelligcncc, lnlcrnal Report, 1991 Washington

[12] D. E Goldbcrg, "A Note on Boltzman Tournament

Selection for Gcnctic Algorithms and Population-Oriented Siniulatcd Annealing", 1991, Complex Systems 3
[13] D. E Goldbcrg, K. Deb, "A Comparative Analysis of Selection Schcnics Used in Genetic Algorithms" ,Genetic Algorilhms in Search, Optimization and Machine

Learning Suinincr School, 1991 Stanford

[ 141 S.

Kirkpatrick et al. "Optimization by Simulated Annealing" 1983 Science

1151 J. H. Holland, " Adaption in Natural and Artificial Systcnis", The University of Michigan Press, 1975 Ann Arbor
1161 H. Ding et al, "Optimal clustering of power networks using genclic algorithms", Proc. 3rd Biennial Symp. Indust. Elcct. Applications, Ruston, LA, USA, LA Tech. Univ., 1992 1171 V. Ajjarapu et al., "Application of genetic based algorilhins LO oplimal capacitor placement", Proc. First International Forum on Applications of Neural Networks to Power Systems, New York, NY, USA, IEEE 1991
[ 181

t Vladimiro Miranda was born in Oporto, Portugal, on March 11, 1955. He received his Licenciado, Ph.D. and Agregado degrees from the Faculty of Eng. of the University of Oporto (FEUP)in 1977, 1982 and 1991, in Electrical Engineering. In 1981 he joined FEUP and currently holds the position of Professor Associado. In 1985 he joined also INESC - Research Engineering Institute for Systems and Computers and holds presently the position of Project Manager - Head of Information and Decision in Energy Systems area.

J o b Vasco Ranito was born in Oporto, Portugal, on January

23, 1965. He received his Licenciado and M.Sc. degrees from FEUP in 1988 and 1993, in Electrical Engineering and Computers. In 1988 he joined INESC as a researcher in the Information Systems area. He presently holds there the position of Team Leader in the Software Engineering Development Center and is a Ph.D. student. Luis Mirruel Proenca was born in Lisbon, Portugal, on May 10, 1966. He received his Licenciado degree from FEUP in 1989 in Electrical Engineering and Computers. He is now a M.Sc. student at INESC, in the Information and Decision in Energy Systems group.

T. Haida et al., "Genetic algorithms approach to voltage optimization", Proc. First International Forum on Applications of Neural Networks to Power Systems, New York, NY, USA, IEEE 1991