Você está na página 1de 9

Electrical Power and Energy Systems 25 (2003) 515523 www.elsevier.

com/locate/ijepes

Review

Power distribution system optimization by an algorithm for capacitated Steiner tree problems with complex-ows and arbitrary cost functions
Gang Duan, Yixin Yu*
School of Electrical Automation & Energy Engineering, Tianjin University, Tianjin 300072, Peoples Republic of China Accepted 26 September 2002

Abstract An algorithm called genetic shortest-path algorithm is presented to solve capacitated minimal Steiner tree problems in graphs with complex ows and arbitrary arc cost functions, but without negative cycles. Voltage constraint can also been taken into consideration by the algorithm. Hence, it can solve various power distribution system optimization problems with detailed mathematical models. In the proposed algorithm, a local optimization method based on shortest-path algorithm and heuristics is used to nd the local optimums, in which the minimum cost objective and all constraints are considered and the specialties of the problems are made good use of. Genetic operations are only used to search the global optimum from the local optimums. Therefore, this algorithm overcomes the disadvantage of general genetic algorithm in local searching. An example for distribution system planning problem with large scale is given to demonstrate the power of the algorithm. q 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Power distribution system planning; Power distribution system reconguration; Steiner tree problem; Genetic algorithm

1. Introduction Minimal Steiner tree problem [1] is to nd a tree to feed all demand nodes by passing through some of the Steiner nodes (i.e. transshipment nodes) with the least cost. In this paper, power distribution system optimization problems, for example, reconguration and planning, are modeled as capacitated minimal Steiner tree problems in graphs with complex ows and arbitrary arc cost functions of the corresponding arc ows, but without negative cycles, that is, no cycles with negative total costs. Voltage constraint can also be considered in this model. Up to now, no algorithm for this detailed model has been presented. Typical minimal Steiner tree problem [1] only considers constant arc costs. Although it is too simplied to model power distribution system optimization problems, it is already an NP hard problem [2], which cannot be solved in polynomial time by Turing computers. Some more general Steiner tree problems [3] and mathematical models proposed in literature of power systems [4 8] take linear costs, quadratic costs, capacity
* Corresponding author. Tel.: 86-22-2740-5490; fax: 86-22-27403544. E-mail address: yixinyu@tju.edu.cn (Y. Yu).

constraint and even voltage constraint (for power system problems) into consideration, but all of them are not precise enough to model power distribution system optimization problems. For examples, voltage constraint is only veried after a candidate solution is found instead of in the searching process; arc costs must be of some simple and predened formation; and reactive power is not considered in planning problems. The main obstacle to model power distribution system optimization problems precisely is that they are NP hard problems, for which the computation burden will increase exponentially with the scales of the problems. This is especially true for using mathematical programming methods, such as branch and bound methods, Lagrange relaxation methods, and dynamic programming. However, the proposed detailed Steiner tree model for power distribution system optimization problems in this paper is a much more complicated NP hard problem. From our viewpoint [9], only randomization techniques based methods, such as genetic algorithm, simulation annealing [7] or Tabu searching, can avoid the curse of dimensions, while getting a believable global optimum or near optimum. Moreover, we think that genetic algorithm is the most effective randomization technique [9]. Some genetic algorithms [5,8,10] for the similar problems have

0142-0615/03/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved. PII: S 0 1 4 2 - 0 6 1 5 ( 0 2 ) 0 0 1 2 8 - X

516

G. Duan, Y. Yu / Electrical Power and Energy Systems 25 (2003) 515523

already been proposed. However, in these algorithms, besides that more simplied models are used, nearly all objectives and constraints are taken into account only by genetic operations that realize the optimization by survival of the ttest and are not good at ne local searching [11]. Of course, the specialties of the particular problems are not made good use of either. Hence, these genetic algorithms are inefcient. In the following sections, a problem-oriented genetic algorithm, called genetic shortest-path algorithm, will be presented. Firstly, a local optimization method based on shortest-path algorithm and heuristics is proposed to nd local optimums, in which the objective and all constraints are considered and the specialties of problems are made good use of. Then genetic operations are used to search global optimum from local optimums. The algorithm can solve power distribution system planning or reconguration problems with detailed mathematical model efciently.

Usually, loads are given in complex power. They need to be converted into complex current demands. Since the angle differences and amplitude differences of node voltages are quite small in distribution systems, It is assumed that the node voltages of loads are all the rated voltage Ve with zero angles. Thus, complex electric current demands can be obtained by the following formula id < pd jqd =Ve 2 where id ; pd and qd are, respectively, the complex current, active power and reactive power of load d. Note that the above assumption is only used to obtain demands in electric currents approximately, voltage drops and power loss will still be considered in the optimization process. Thus, a connected directed graph G N ; A is derived, where N Ns < Nd < Nt ; and Nd is the set of demand nodes, Nt is the set of Steiner (or transshipment) nodes. A is the set of all directed arcs. Capacitated minimal Steiner tree problems with arbitrary arc cost functions can be formulated as follows X min fij iij 3
aij [A

2. Problem formulation Each undirected element in a distribution system except loads is treated as two mirror directed arcs aij and aji : The arbitrary cost functions of arc aij is fij iij ; where iij is the complex ow in arc aij ; and in this context, it is the complex electric current. Arbitrary cost function means that fij iij can be any types of function, such as nonlinear, piecewise, discrete or a curve with any shape, as long as there are no negative cycles in the graph derived from the distribution system. For examples, in a planning problem, the cost function per unit length of potential feeder with three candidate conductor types (LGJ-70, LGJ-120, LGJ-240) can be a piecewise function as follows
fUNIT_CDI_TYPE i HTYPE alil2 rTYPE t 8 > 1:53 0:4 1024 6000 0:460 1023 lil2 0 , lil , 87:6 > > < 1:88 0:4 1024 6000 0:270 1023 lil2 87:6 # lil , 119:1 > > > : 2:35 0:4 1024 6000 0:132 1023 lil2 119:1 # lil , 305

s.t. X
aik [A

iik 2

X
akj [A

ikj

8 X it ; 2 > > > t[Nd < ik ; > > > : 0;

k Ns k [ Nd k [ Nt 5 6 7 8 9 10 4

U $ liij l; Iij

aij [ A aij [ A aij [ A; aji [ A k [ Nt k [ Nd k Ns

Realiij $ 0; iij iji 0; dInk [ {0; 1}; dInk 1; dInk 0;

where HTYPE is the annualized xed cost per unit length; rTYPE, the resistance per unit length; a, the electricity price; t is the annual operating time. The upgrade of elements (feeders or transformers) or the selection of the number of paralleled feeders can also be modeled by piecewise functions. A virtual supply or source node is needed and denoted as Ns. All real power sources, including power plants, substations and equivalent external power systems, are modeled by virtual arcs from the virtual supply node to the real source nodes. The costs of the real power sources, e.g. operation costs and xed costs, are the costs of the virtual source arcs. All load nodes are treated as demand nodes, while the other nodes only for transshipment are called Steiner nodes.

where iij [ C; ik [ C ; C is the complex number set; ik is U the demand of node k; Iij is the capacity bound of arc aij ; dInk is the indgree of node k. When the direction of the real part of complex ow is dened as the direction of the complex ow, Eq. (6) indicates that the ow passing through a directed arc must have the same direction as that of the arc. From Eq. (2), it can be seen that the direction of complex electric current is also the direction of active power. Eq. (7) says that only one of the two arcs aij and aji ; which represent an undirected element and have opposite directions, can be chosen. Eqs. (8) (10) form the radialization constraint. For power distribution system optimization problems, the voltage drop constraint should also be considered, that is e 2 # Vk 2 Ve # e ; k[N 11

G. Duan, Y. Yu / Electrical Power and Energy Systems 25 (2003) 515523

517

where Vk is the voltage amplitude of node k, e and e2 are the upper bound and lower bound of the permitted voltage offsets from the rated voltage Ve :

the other genetic algorithms. This not only makes it possible to use the proposed efcient local optimization algorithm, but also can reduce the number of genes because demands are fewer than arcs. As a result, the computing time will be reduced greatly.

3. Prole of the proposed algorithm Firstly, an efcient local optimum searching method based on shortest-path algorithm and heuristics is developed to obtain local optimums. In the local searching method, all demands are ordered in a sequence, the feeding paths for them are searched one by one and the Steiner tree will be formed step by step. In each step, shortest-path algorithm is called to search the minimum cost feeding path for a demand in a subgraph of G. Whenever the feeding path is found, a new subgraph is derived from G, so that the next feeding path for the next demand node will not form a cycle with the already selected arcs having non-zero ows, besides that capacity constraint cannot be violated. In the process of searching for feeding paths, capacity margin is also considered, so that arcs with wider capacity margin have greater possibilities to be selected than arcs with lower costs and narrower capacity margin even though they may be have higher costs. For a power distribution system optimization problem, if a new found feeding path results in a violation of voltage constraint, then this path is abandoned and the costs of arcs on the path are penalized so that these arcs are avoided to be selected as most as possible in the next feeding path searching for the same demand. The above process is repeated until all demands have their own feeding paths, that is a local minimal Steiner tree is found, or until one of the demands cannot nd its feeding path, that means no solution for the loads sequence. To reduce computation burden, knowledge about a practical distribution system should also been used. For example, if some feeders (i.e. arcs) are too far to feed a demand, then they can be removed from the system temporarily when the feeding path for this demand is being searched. It can be seen that this method can obtain a radial network very easily; and besides that, voltage constraint and capacity constraint are considered in the searching progress instead of just being veried after candidate solutions have been obtained. The above-mentioned local optimum searching method will be called by a genetic algorithm that will search global optimum from local optimums. Loads sequences will be chromosomes. Because a local optimum corresponds to a group of loads sequences having similar sequence features, hence, if the genetic algorithm nds any one of the optimal loads sequences, then it obtains the optimal Steiner tree. That is the genetic algorithm does not need to perform ne local searching, which it is not good at. In fact, it only needs to nd what loads sequence feature the optimal solution should have. Note that the proposed genetic algorithm nds an optimal Steiner tree by ordering demands in sequences instead of combining arcs, which has been used by all 4. Local optimization algorithm 4.1. Feeding paths optimization for all demands in a sequence Assuming that the loads, that is the demands, will be added into graph G in a sequence as follows: SQ {D1 ; D2 ; ; Dm21 ; Dm ; ; DNL } 12

where NL is the total number of loads, and Dm indicates the mth load node, its value is im obtained by Eq. (2). Firstly, the minimum cost feeding path for D1 is searched by calling subroutine OneLoadRoute detailed in Section 4.2, then for D2,, until DNL. Whenever a feeding path is found, the related arc ows and indegrees of nodes in G will be revised. For example, if path Rm21 for Dm21 is found, where Rm21 is the set of arcs on the feeding path for Dm21 ; and aij [ Rm21 ; then iij_m iij_m21 im21 dInj_m 1 13 14

where the subscript _m (or _m 2 1) indicates the values of the corresponding variables are the values before searching the feeding path of Dm (or Dm21 ). Before R1 is found, all the indegrees of nodes and the ows in all arcs should be zero. To accelerate feeding path searching for a demand, say Dm ; we search the path in directed graph Gm ; called searching path graph of Dm instead of G that already contains the determined feeding paths for the rst m 2 1 demands, which should be kept unchanged. Gm is a subgraph derived from G by the following rules: 1. If the indgree dInk_m of node k is already 1, then exclude all the arcs directing to node k and with zero ows from Gm : Apparently, these arcs have not been selected before. 2. If arc aij has been selected, that is, iij_m is not equal to zero, then discard the corresponding mirror arc aji : 3. If
U liij_m im l . Iij

15

then discard aij : 4. Check up the amplitude Vi_m of each node voltage, if it is nearly to violate the voltage constraint, for example 0 # Vi_m 2 1 2 1Ve , err 16

where 1 is the permitted percent of voltage drop, e.g. 5%, and err is an error indicating how close Vi_m to the lower

518

G. Duan, Y. Yu / Electrical Power and Energy Systems 25 (2003) 515523

bound of voltage, then discard all the arcs connected to this node. 5. Dene the distance of an arc from a load as follows: q DIij_m kij_m min{ xB_ij 2 xm 2 yB_ij 2 ym 2 ; q xE_ij 2 xm 2 yE_ij 2 ym 2 }

17

where xB_ij ; yB_ij and xE_ij ; yE_ij are the coordinates of the start node and the end node of aij correspondingly. The coordinate of load m is xm ; ym : kij_m is a coefcient to take landform factors into account. If DIij_m is greater than a threshold distance, say DImax, then arc aij is thought being too far from load Dm to feed it. Then, aij is discarded from Gm : 6. Discard the arcs that cannot feed Dm for some other reasons. For example, some loads need to be fed only by feeders with high reliabilities. For planning problems, arc cost should be the sum of xed cost and variable cost of an arc; and for reconguration problems, arc cost may be the active power loss in an arc. However, from the following simple network ows example shown in Fig. 1, it can be seen that capacity constraint should also be considered in arc costs during path searching process. Up to now, we have not given more chances of being selected to arcs with wider capacity margin. In Fig. 1, if the path for D1 is searched rst, then path A1 A3 is found. Take radialization constraint into consideration, D2 cannot nd its feeding path. If the searching sequence is D2, D1, there is still no solution. However, it is apparent that a solution exists: A2 A3 for D1 and A2 A4 for D2. To solve the problem, capacity constraint must be considered in arc costs during the path searching process. The costs will be called capacity costs of arcs. Capacity constraint not only means that the ows in arcs cannot exceed the limits, but also means that the arcs with wider capacity margin should have greater possibilities to be selected even though they maybe have higher real costs. In the proposed algorithm, an index E is used to convert real arc costs to capacity costs. The greater E, the more consideration of capacity margin is taken into arc costs. When E is zero, capacity costs of arcs are just the same as

real arc costs. The interval of E is 0; Emax : E has the following properties: (a) a greater E tends to have greater possibilities to obtain a solution with higher real network cost and wider capacity margin; (b) capacity consideration is an ancillary means for optimization, hence E is not a key parameter; (c) it will be seen from Section 4.2 that the interval of E is a very narrow range. Theoretically, an optimal E should be found to obtain the global optimum. However, based on (b) and (c), we just select the best E as the optimal one from the following discrete set of E {ElE kc h; 0 # kc # kc_max Emax =h} 18

where h is a step and kc is an integer. The G corresponding kc is noted as Gkc ; which will be revised when the feeding path for a load in sequence SQ is found. When feeding paths for all loads are found, Gopt_kc is got. The best Gopt_kc is noted as Gopt, which has minimum real network cost, and the corresponding E is called the best one under this loads sequence. If no solution can be found for all E, then the loads sequence should be discarded. How h and Emax are determined will be given in Section 4.2. 4.2. Feeding path optimization for one demand In this section, subroutine OneLoadRoute(Gm ; SQ, kc ; m ) is presented to search minimum cost feeding path for load Dm in sequence SQ. The found feeding path Rm should not form a loop with the arcs of the determined feeding paths for the rst m 2 1 loads while keeping them unchanged. Because Rm is searched in Gm ; the above two constraints will be satised automatically. Hence, the key for this subroutine is how to determine the virtual arc costs during the searching process to take capacity constraint and voltage constraint into account. The detail of the subroutine is as follows: Step 1. Calculate the (incremental) arc costs without considering capacity constraint and voltage constraint. Assuming that all the ows in arcs of Gm are increased by im ; the incremental arc costs are taken as the real arc costs wP_ij_m Dfij_m fij iij_m im 2 fij iij_m 19

where fij can be any function as stated in Section 2. Step 2. Calculate the capacity costs of arcs. Let cscale_ij_m be a variable that indicates how many demands at most can be further fed through arc aij after the feeding paths for the rst m 2 1 loads have been determined. From Eq. (15), it can be seen that in graph Gm the remainder capacity of arcs can feed Dm at least, hence cscale_ij_m $ 1: If
U Iij

Fig. 1. The necessity to consider capacity constraint in arc costs.

NL X $ iij_m il
lm

20

G. Duan, Y. Yu / Electrical Power and Energy Systems 25 (2003) 515523

519

then let cscale_ij_m NL 2 m 2 1 NL 2 m 1 nr_m 21

where nr_m is the total number of the remainder demands waiting to be fed. If n nij_m 1 ij_m X X U il # Iij , iij_m il ; nij_m , NL 22 iij_m
lm lm

then cscale_ij_m nij_m 2 m 2 1 nij_m 2 m 1 Capacity costs are dened as follows


m wC_ij_m wP_ij_m =ce scale_ij_m

23

24

where em indicates the strength that the remainder capacity of arc affects the arc cost. When em 0; capacity constraint is not considered. The following rules are used to determine em 1. When there is no more load to be fed after the feeding path for Dm is found, that is m NL; then eNL should be zero. 2. For the rst load D1, there are the most number of subsequent loads to be fed, so the strongest capacity conversion is needed. Hence, e1 should be the greatest one of all em ; say an E given by Eq. (18). 3. With the decrease of remained loads nr2m to be fed, em should decrease until to zero. To satisfy these rules, a linear equation (25) or an elliptical equation (26) can be used to obtain em em Enr_m 2 1=NL 2 1 s   NL 2 nr_m 2 em E 1 2 NL 2 1 25 26

Whether the temporal ows should be accepted depends on if voltage constraint is violated. Step 4. Verify the voltage constraint. In graph Gm ; the arcs with non-zero ows and the rst m loads form a radial distribution system. Because the virtual source is an ideal voltage source, only the voltages of nodes in sub-tree TRm that the feeding path Rm belongs to change after Dm is fed. Furthermore, the lowest voltage of TRm must be one of the voltages of load nodes. Therefore, to check up the voltage constraint, only the voltages of load nodes in TRm need to be known. An exact power ow algorithm can be used to get the voltages. However, to save computing time, as a compromise between speed and precision, the following approximate algorithm is used. Given that the voltage of virtual source node is V0 with zero angle, then the amplitudes of voltages of load nodes are X l [ LRm 28 Vl_m V0 2 it_ij_m1 rij jxij ; aij [Rl where LRm is the set of load nodes in TRm ; Rl is the feeding path for load Dl ; rij and xij are the resistance and reactance of aij ; respectively. If all these voltages satisfy lVe 2 Vl_m =Ve l # 1 29

Because em is a power index and cscale_ij_m $ 1; it is necessary for E not to be too large, otherwise capacity costs will be too small and even nearly zero. E is in 0; Emax by Eq. (18), hence Emax should not be too large either. Usually, let Emax 2 and h 0:1: Step 3. Search the minimum cost feeding path for Dm : Call shortest-path algorithm [1] to nd the feeding path Rm with minimum cost from Ns to load Dm in Gm : Ford algorithm is used, which can nd the shortest path in a graph containing negative cost arcs but without negative cycles. If Rm is not found, then Dm and Ns cannot be in the same connected subgraph of Gm ; and this loads sequence SQ with this kc should be discarded, that is the subroutine stop with failure. If Rm is found, revise the related arc ows of Gm and get the following temporal ows ( it_ij_m1 iij_m im iij_m aij [ Rm aij Rm 27

where 1 is the permitted voltage bias, then subroutine OneLoadRoute(Gm ; SQ, kc ; m ) has a valid solution and the temporary ows it_ij_m1 will be converted to determined ows iij_m1 by Eqs. (13) and (14). Otherwise, the found feeding path for load Dm should be abandoned, and the temporary ows it_ij_m1 of arcs should be restored to the former values it_ij_m : Then the costs of all arcs on the path are penalized according to Eq. (30), so that in the next path searching process for the same load these arcs have fewer possibilities to be selected. Next, return to step 3 to search the minimum cost feeding path for Dm again. Because only the costs of some arcs are changed while the remaining capacity of arcs and the structure of Gm keep unchanged, it can be sure that the feeding path can still be found. Then check up the voltage constraint again. This process including steps 3 and 4 will be iterated until a feeding path not violating voltage constraint is found or the maximum failure times NV_fail_max ; say 10, is reached. The arc cost penalization formula is as follows: 8 wV_ij_m_k bwV_ij_m_k21 ; aij [ Rm_k21 > > < wV_ij_m_0 wC_ij_m 30 > > : Rm_0 Rm where k is the times of violating voltage constraint, b is a penalization factor greater than 1, e.g. 1.2, and wV is the virtual arc cost with both capacity constraint and voltage constraint being considered. It should be noted that if a feeding path for Dm is found, then the feeding path searching for the next load Dm1 can start; otherwise, no feeding tree for this loads

520

G. Duan, Y. Yu / Electrical Power and Energy Systems 25 (2003) 515523

Fig. 2. The owchart for the overall local optimization algorithm.

sequence with this kc ; and begin to try the next kc in Eq. (18). The owchart for the overall local optimization algorithm is shown in Fig. 2.

5. Global optimization algorithm The above algorithm can nd an optimal or nearly optimal feeding tree for loads in a sequence. If we change the sequence, difference solution may be obtained. Hence, it is only a local optimum solution for the original Steiner tree problem. The local algorithm is also called subroutine AllLoadsRoute(Graph: G, Sequence of Loads: SQ). How to nd the global optimum from these local optimums is in fact how to nd an optimal loads sequence. The following genetic algorithm, called genetic shortest-path algorithm, is presented to nd the global optimal loads sequence. Step 1: Initialization. Order the loads in descending sequence, and label them with their indexes. Thus, an index will represent a load. Randomly arrange these indexes, different chromosomes will be obtained. The indexes of loads are taken as genes, and the number of loads is the length of chromosome. Given that Nchrom is the size of population, the initial population is made of Nchrom such chromosomes. Each chromosome will be taken as a loads sequence. The position of a gene in a chromosome indicates the order of the corresponding load to be considered in AllLoadsRoute. The descending sequence usually should be included in the rst generation, because, from the viewpoint of heuristics, it has greater possibilities to contain more gene segments that have better gene combinations.

Step 2: Evaluation. Corresponding to each chromosome, call subroutine AllLoadsRoute to obtain the local optimal Steiner tree and its cost that is raw tness. The objective is to minimize the cost, but the greater the tness used by genetic algorithm, the better. Therefore, an appropriate conversion is needed to obtain the genetic algorithm tness from the raw tness. For the local optimal Steiner trees with the same costs, the greater kc ; the greater genetic algorithm tness, because a greater kc usually means the Steiner tree has greater possibilities of having wider capacity margin. Step 3: Stopping criterion. To avoid premature convergence, when the chromosomes of the best tness are more than 50% of the population, genetic operators with strong destructive are used until the population is diversied again. This process can be called drastic evolution [9] relative to the ordinary evolution. Thus, when either the permitted maximum times of drastic evolutions or a specied maximum generation is reached, the global searching is stopped. Step 4: Selection and reproduction. Roulette wheel method [11] is used to realize selection and reproduction. To ensure that at least one of the best chromosomes is reproduced to the next generation, elist method [11] is also used at the same time. Step 5: Crossover. In the process of ordinary evolution, the population is various, and partially mapped crossover (PMX) is used. In the process of drastic evolution, partially reverse matched crossover (R-PMX) is proposed to diversify the population. For both the crossovers, the crossover points are selected randomly. It should be noted that for some special crossover points, PMX given in Ref. [10] will generate unfeasible offspring. For example, given the parent chromosomes and crossover points as follows Parents : A 671l243l58; B 863l451l27: By PMX in Ref. [10], the section between the crossover points denes an interchange mapping: 2$4 4$5 3$1

It will generate the following unfeasible offspring: Unfeasible offspring : A0 673l451l48 B0 861l243l47

That is because gene 4 is mapped twice. To avoid unfeasible offspring, transferred mapping 2 $ 5 is used instead of 2 $ 4 4 $ 5. Hence, the offspring of modied PMX should be PMX offspring : A0 673l541l28 B0 861l423l57

For R-PMX, rstly the genes between the crossover points of A are rearranged in a reverse sequence: R-PMX Reversed A : Ar 671l342l58 Then perform modied PMX on Ar and B, and the interchange mapping will be 2$1 3 $ 5 coming from 3 $ 4 4 $ 5

G. Duan, Y. Yu / Electrical Power and Energy Systems 25 (2003) 515523

521

As a result, the offspring of R-PMX will be R-PMX offspring : B 865l432l17 By R-PMX, even two identical parents can generate new offspring with new genes combinations. It is very useful for exploring new searching space when the evolution converges to a local optimum, that is, most of the chromosomes have similar or the same genes combinations. Although RPMX can diversify the population greatly, it has also great possibilities of destroying the good combinations of genes. Therefore, a lower crossover probability of R-PMX is used. Step 6: Mutation. In the process of ordinary evolution, a low mutation probability Pm of chromosomes is used to avoid destroying good combinations of genes largely, typically Pm 0:1: However, in the process of drastic evolution, a high mutation probability is used, typically Pm 0:5: In Ref. [10], exchange mutation is given, which exchanges two randomly selected genes in a chromosome. However, when there are too many loads, one time exchange usually cannot obtain enough new genes combinations. Hence, multiple-exchange mutation is proposed, in which the exchange is performed several times Nx that is an integer selected randomly from [1, NL/2]. Return to step 2.
00

6. Numerical example The planning of a large-scale power distribution system is presented as an example. The original system to be planned is shown in Fig. 3. The dashed lines represent candidate feeders while the real lines represent existing feeders. The candidate conductor types are LGJ-240, LGJ120, and LGJ-70. If necessary, the existing conductors with smaller sections can be replaced by conductors with larger sections. The candidate substations are at C, D and E. Substation A and substation B are existing ones with capacity 1 20 MVA while substation A is expandable if necessary. The candidate substation types are 2 16 MVA, 2 20 MVA, 2 31.5 MVA, the upper bounds are 60% of the rated capacity with N 2 1 security being considered. The original system to be planned totally has 378 nodes, 271 loads, 485 feeders or transformers, and 85 meshes (note that for some branches combination-based algorithms, e.g. branches exchange algorithm [7], more meshes means much more computation burden; however, for this loads combination-based algorithm, no apparently increase of computation burden will be resulted). The parameters labeled on the loads are the rated apparent power of load transformers in kVA. The total active power of loads is 45 805.3 kW, and the total reactive power of loads is 28 064.9 kVar. The rated voltage of the system is 10 kV.

A00 672l541l38

Fig. 3. The original 10 kV power distribution system to be planned.

522

G. Duan, Y. Yu / Electrical Power and Energy Systems 25 (2003) 515523

Fig. 4. The planning result for the 10 kV power distribution system.

The voltages of the secondary sides of transformer substations are 10.5 kV. The resulted system is shown in Fig. 4, substation A is expanded to 2 20 MVA, and a new substation of 2 31.5 MVA should be set up at C. The active power losses and the costs of the resulted system are shown in Table 1. The lowest voltage in the resulted system is 10.32 kV, and the corresponding node is labeled by y in Fig. 4. The program was written in C language. For this example, the test environment is: Pentium 90, 16 Mb RAM, Windows 98. The average computing time is about 2.8 h. The best solution is given above. It should be noted that although genetic algorithm cannot ensure that the global optimum is found, it provides an efcient

method to perform global searching while overcoming the curse of dimensions.

7. Conclusions The proposed genetic shortest-path algorithm can solve capacitated minimal Steiner tree problems in graphs with complex ows and arbitrary arc cost functions of the corresponding arc ows, but without negative cycles. Voltage constraint can also be taken into consideration by the algorithm. Because of the exibility in modeling, the algorithm can be used to solve various power distribution system optimization problems with detailed models. Moreover, the computational efciency of the algorithm is very high in solving large scale problems, because genetic algorithm is used to perform global searching, which avoids the curse of dimensions, while a high efcient local optimization method based on shortest-path algorithm and heuristics is used to nd local optimums, which overcomes the disadvantage of genetic algorithm in ne local searching. In the local searching method, the objective and all constraints are considered and the specialties of problems are made good use of. Genetic operations are only used to search the global optimum from the local ones.

Table 1 The planning result for the 10 kV power distribution system Items Active power loss (kW) 432.7 153.0 585.7 Variable cost (104 Yuan/year) Fixed cost (104 Yuan/ year) 64.1 15.7 79.8 Total cost (104 Yuan/ year) 167.9 52.4 220.3

Feeders Substations System

103.8 36.7 140.5

G. Duan, Y. Yu / Electrical Power and Energy Systems 25 (2003) 515523

523

To further accelerate the algorithm, parallel computing technique can be used because genetic algorithm is an inherent parallel algorithm. References
[1] Jungnickel D. Graphs, networks and algorithms. Berlin: Springer; 1999. [2] Garey MR, Johnson DS. Computers and intractability: a guide to the theory of NP-completeness. New York: W.H. Freeman; 1979. [3] Cruz FRB, MacGregor SJ, Mateus GR. Solving to optimality the uncapacitated xed-charge network ow problem. Comput Oper Res 1998;25(1):6781. [4] Khator Suresh K, Leung Lawrence C. Power distribution planning: a review of models and issues. IEEE Trans Power Syst 1997;12(3): 11519. [5] Ramirez-Rosado IJ, Bernal-Agusti JL. Genetic algorithms applied to the design of large power distribution systems. IEEE Trans Power Syst 1998;13(2):696703. [6] Sar RJ, Salama MMA, Chikhani AY. System optimization of electrical distribution networks: a new nonlinear formulation. Int J Power Energy Syst 1997;17(3):16974. [7] Chiang HD, Rene JJ. Optimal network recongurations in distribution systems. Part 2. Solution algorithms and numerical results. IEEE Trans Power Deliv 1990;5(3):1568 74.

[8] Levitin G, Mazal-Tov S, Elmakis D. Genetic algorithm for optimal sectionalizing in radial distribution systems with alternative supply. Electr Power Syst Res 1995;35(3):14955. [9] Duan G, Yu Y. Problem-specic genetic algorithm for power transmission system planning. Electr Power Syst Res 2002;21(1): 4150. [10] Reeves CR. Modern heuristic techniques for combinatorial problems. New York: Wiley; 1993. [11] Michalewicz Z. Genetic algorithms data structures evolution programs. Berlin: Springer; 1994.

Gang Duan achieved his PhD degree from Tianjin University, China. Currently, he is working as a visiting scholar at the University of Queensland, Australia, with on leave from Tsinghua University, China. E-mail: gang@duan.com.cn

Yixin Yu is the Professor of Tianjin University. Correspondence address: School of Electrical Automation and Energy Engineering, Tianjin University, Tianjin 300072, China. E-mail: yixinyu@tju.edu.cn

Você também pode gostar