A Genetic Algorithm Approach for Optimal Distribution Feeder Restoration During Cold Load Pickup Sudhakar Chavali Anil Pahwa Sanjoy Das Department of Electrical and Computer Engineering Kansas State University Manhattan, KS 66506 Abstract: Restoration of distribution feeders after long interruptions creates cold load pickup conditions due to loss of diversity among the loads. The distribution system load may have to be restored step-by-step using sectionalizing switches under such conditions to prevent overheating of substation transformer. The restoration time is dependent on the order in which sections are restored. A method based on genetic algorithm to find the globally best restoration sequence of sections is presented in this paper. Results obtained using this method for a test case are presented. Keywords: Distribution system restoration, cold load pickup, genetic algorithm, mutation, and genetic edge recombination. enduring load demand may result in excessive transformer heating. Heating is the limiting factor in transformer loading. As per ANSI /IEEE Transformer loading guide C57.91 – 1995 [3], for 65 0C average winding rise transformers limitations are; maximum top-oil temperature of 110 0C, maximum hottest-spot winding temperature of 180 0C, 4% loss of life in any one emergency operation and maximum short time loading of 2 • SR, where SR is the nameplate rating of the transformer. B. Step by step restoration: I. INTRODUCTION A. Cold load pickup: Distribution systems are used to feed power to various loads such as air conditioners, lighting, heating, electronic equipment, etc. These loads can be classified as cyclic loads and non-cyclic loads. Cyclic loads such as air conditioners, refrigerators, space heaters, water heaters, etc., are generally thermostatically controlled. Non-cyclic loads such as lights, washing machines are manually switched as per user requirements. During normal operation diversity among the loads persists and aggregated load on the substation transformers would be less than the connected load. If a distribution system experiences an extended outage, the diversity among these loads would be lost and during restoration undiversified load demand may be much higher than the distribution substation capacity. This condition is known as cold load pickup. The cold load pickup has four stages [1,2]. These four stages are 1) inrush current, 2) motor starting current, 3) motor running current and 4) enduring demand. The first three stages are transient and generally last for about 15 seconds and the load during these stages will be about 3 to 15 times the normal load. The enduring demand on the system remains until the normal diversity among the loads is regained. The enduring load may vary from 2 to 5 times normal load and it may last for several hours depending on rate of load decay. Under certain conditions all of the system load during cold load pickup cannot be restored simultaneously as overloading of transformers due to 0-7803-7282-4/02/$10.00 ©2002 IEEE Sectionalizing switches can be used on distribution system feeders to restore the load in steps in case total load on the system cannot be restored in one step. The shortterm load level and its duration during step-by-step restoration can be determined based on thermal characteristics and loading limitations of the transformer. To model the aggregate load during cold load pickup a delayed exponential model as shown in Fig. 1 is used [4]. Where S(t) is load as a function of time, Tout is the outage duration, SU is the undiversified load, SD is the diversified load, α is the rate of load decay, and ∆ t is the undiversified load duration. S(t) SU α SD Tout ∆t Fig. 1. Cold load pickup model t (time) Fig.2 shows single line diagram of a distribution feeder typical system with automatic sectionalizing switches [4]. Since sections 1, 2, 3; 4, 5; 6, 7 and 8, 9, 10 are on four different feeders, in the order mentioned, the precedence constraints are 1st ! 2nd ! 3rd ; 4th ! 5th ; 6th ! 7th and 8th th th ! 9 ! 10 . In other words section 3 cannot be restored unless the sections 1 and 2 are restored. have 10! (3,628,800) combinations. Although some combinations are not valid due to precedence constraints of sectionalizing switches the total number of valid combinations is still quite high. Hence, it is very time consuming to calculate the restoration times of all valid sequences among all the possible combinations to find the best restoration sequence. C. Optimization of Restoration Sequence: Section 1 Section 2 Section 3 Section 4 Section 5 Section 6 Section 7 Circuit Breaker Section 8 Section 9 Section 10 Sectionalizing Switch Transformer Fig.2. One line diagram of a distribution system The sequence in which sectionalizing switches are closed is important since different sequences would result into different restoration times, as undiversified loads of sections are not the same. Restoration time is the time at which the last sectionalizing switch in the sequence is closed. Some of the restoration sequences are not valid because of precedence constraints. For example, (1 4 6 2 8 7 9 10 3 5) is a valid sequence whereas (4 3 6 8 1 7 5 9 2 10) is not so, since section 3 appears before 1 and 2 in the sequence. The restoration time of different sequences can be calculated from [4], which is given by, k −1 T[k] = ln S MT − SU [ k ] − ∑ S D [ i ] k −1 ∑ (S i =1 U [i ] i =1 − S D [ i ]) e , (1) α ( T [ i ] + ∆ t [ i ]) Where, T[i] α[i] SMT SU[i] SD[i] ∆[i] – Restoration time of ith section, hr – Rate of load decay of ith restored section, 1 / hr – Maximum transformer capacity, VA – Undiversified load of ith section, VA – Diversified load of ith section, VA – Undiversified load duration or time delay of ith restored section, hr We need to optimize the restoration sequence of sections so that restoration time is minimized. Some of the combinations would give the same restoration time since more than one section can be restored in the first step. The main problem is to find optimal solution from a large search space. Even a small system with 10 sections will 0-7803-7282-4/02/$10.00 ©2002 IEEE Ucak and Pahwa [4] used an adjacent pair wise interchange method (APIM) for minimizing the restoration time. The maximum and minimum restoration times of the system are 326.1 minutes and 275.14 minutes. The disadvantage with APIM approach is that the resulting sequence will not guarantee the global minimum. The best sequence obtained from a run of APIM is (1 8 6 9 4 10 7 2 3 5) with restoration time of 282.6 minutes. The global search gave the sequence (1 2 8 9 3 4 10 6 7 5) with restoration time of 275.14 minutes. A genetic algorithm approach for optimizing the order of restoration of sections during cold load pickup is presented in this paper. II. THE GENETIC ALGORITHM Genetic algorithms [5] are stochastic search algorithms based on principles of natural evolution. These algorithms maintain a large population of interacting individuals, where each individual represents a potential solution to the problem. This population is constantly pruned to maintain constant size by applying the Darwinian criterion of survivability. This is possible by assigning a fitness value to each individual. The fitness function should be appropriately chosen to reflect the quality of the individual. The creation of new individuals (offspring) from old ones (parents) takes place through reproduction. Beneficial ‘genes’ are passed on from parent to offspring. When the reproduction is sexual, multiple parents, usually two, are involved to produce offspring. This operation is called crossover. Mutation is another operation that is applied to the individuals. Mutation introduces genetic diversity to a finite population, thereby allowing the search process to explore new regions of the solution space. Genetic algorithms have been successfully applied to a wide range of NP-hard combinatorial optimization problems. The sequence of operation typical in a genetic algorithm is shown below, t = 0; P0 = initial_population(); while (termination_condition = false) do Pintermediate = selection(Pt); Poffspring = crossover(Pintermediate); Poffspring = mutation(Poffspring); t=t+1 Pt = Poffspring; end Our approach was to generate an initial population of valid individuals in a random manner. Each individual was a sequence of sectionalizing switches, i.e. a permutation of the sequence 1, 2,…, n. where n is the total number of switches. We define for each switch i, Pred(i), to be a set of switches whose restoration should precede that of i. For instance, Pred(10) = {8, 9} in Figure 1, since the switches numbered 8 and 9 should be restored before that numbered as 10. A sequence is represented as S = (s1 s2, … sn), where each si is a switch number. A sequence is valid if for each switch si in S, the switches in the set Pred(si) appear before si in S. The algorithm proceeds iteratively to improve the population by means of selection, mutation and crossover that are discussed below. Selection: The better individuals are selected probabilistically from the population to create an intermediate population (Pintermediate), based on the fitness function. If fi denotes the fitness of the ith individual then the probability of selection was given by pi = fi/Σfj. In our approach, the fitness was computed as, every other iteration, if the last switch to be inserted into the offspring O was i, a new switch is picked from either nbd(i, P1) ∩ Ready, or from nbd(i, P2) ∩ Ready at random. As an example consider the two parent sequences P1 = (1 4 8 2 6 9 10 7 3 5) and P2 = (8 1 4 9 6 2 5 7 10 3). The set Ready is initialized to Ready = {1, 4, 6, 8}. If switch 1 is selected in the first iteration, the switch 2 becomes ready for restoration, and Ready = {2, 4, 6, 8}. The offspring O is now (1). The next switch to be inserted will be picked randomly from nbd(1, P1) ∩ Ready = {4} or nbd(1, P2) ∩ Ready = {4, 8}. Occasionally, before the offspring has been fully determined, if the switch i was last inserted into the offspring, the sets nbd(i, P1) ∩ Ready as well as nbd(i, P2) ∩ Ready are found to be empty. In such a case, any other switch in Ready is picked at random for insertion into the offspring. Since this switch is not determined from either parent, the proposed crossover algorithm contains an intrinsic mutation. We found that an external mutation (simply called mutation) had to be added for faster convergence. Mutation: fi = max(T[ k ] ) − T[ i ] k max(T[ k ] ) − min (T[ k ] ) k k where each T[k] is determined by Equation 1. Crossover: The crossover operator was a variant of the well-known edge recombination method [6]. It was modified to ensure that offspring produced by valid parents preserve their validity. The crossover algorithm starts with an empty sequence. In each iteration a new section switch is inserted into the sequence, in the proper order. In order to create a valid sequence, a switch, i, can only be selected only if all switches in Pred(i), have already been restored. The algorithm maintains a set of switches that are ready for restoration, called ‘Ready’. Each time a new switch, i, is inserted into a partially complete sequence, it is removed from this set, and other switches become available for restoration, are inserted into it. The switches j such that each switch in Pred(j) ∈ S are those that are ready for restoration. We define the neighborhood nbd(i,S), to be the set of switches that are adjacent to switch i in any given sequence S. For instance in the sequence S = (1 4 8 2 9 6 10 7 3 5), nbd(8,S) = {2, 4}. Within each iteration, the crossover operation picks two parent sequences at random, P1 and P2, and the offspring O is initialized to an empty sequence ( ). The set ‘Ready’ is then initialized to contain only the first switches in each. Since both P1 and P2 are valid sequences, the first switch appearing in each is in Ready. The first switch in either P1 or P2 is inserted into O as its first switch. Subsequently in 0-7803-7282-4/02/$10.00 ©2002 IEEE The mutation operator is applied to each sequence with a probability µ, called the mutation rate. During mutation, a left switch is picked at random from the sequence. The position of the left switch is then swapped with that of a right switch, which is determined in the following manner, in order to preserve the precedence constraint. If i is the left switch, another switch appearing to its right, j, is determined such that i ∈ pred(j), and every other switch k such that i ∈ pred(k) appears after j. Clearly, i cannot be exchanged with any switch that appears after j since that would violate the precedence constraints. We therefore have to pick the right switch within the subsequence whose limits are defined by i and j. We then choose a switch l in this subsequence that belongs to a feeder other than that of i and which has no predecessor (a switch in pred(l) that appears to its left in the subsequence). The switch l is the right switch. Finally, the positions of i and l are interchanged. As an example, in the sequence P1 = (1 4 8 2 6 9 10 7 3 5) switches 4 and 6 can be swapped to obtain a new offspring (1 6 8 2 4 9 10 7 3 5). But, switches 1 and 9 cannot be swapped as that will result into an invalid sequence of (9 4 8 2 6 1 10 7 3 5), in which switch 2 appears before switch 1 and switch 9 appears before switch 8. The two genetic operators, edge recombination and mutation, discussed above, have the advantage that by applying these operators to the sequences in Pintermediate and Poffspring the resulting offspring will follow the precedence constraints as defined by Pred(si). Also duplication of switches in the offspring is avoided thereby validity of all the offsprings, with regard to precedence constraint and duplication, is ensured. III. RESULTS The algorithm is applied on a test system shown in Fig.1. The diversified and undiversified loads of this system are given in Table 1. TABLE 1 DIVERSIFIED AND UNDIVERSIFIED LOADS OF SECTIONS OF THE TEST SYSTEM Section Number Diversified Load (MVA) 1 2 3 4 5 6 7 8 9 10 Total 3.0 3.0 2.5 1.5 3.0 6.0 2.5 3.5 4.5 3.5 33 Undiversified Load (MVA) 9.0 5.5 6.0 5.0 5.0 9.5 6.0 6.5 11.0 8.0 71.5 For a randomly chosen initial population of 20 sequences the genetic edge crossover operator is applied with probability rate of 100% and mutation is applied with mutation rate of 10%. After each iteration, 10 offsprings are generated and better individuals among the parents and offsprings are chosen for selection process. The algorithm is stopped after predetermined number of iterations. At the end of each iteration the minimum and average restoration times of the corresponding population are noted. An example plot of this data, at the end of 25 iterations, is shown in Fig.3. 0-7803-7282-4/02/$10.00 ©2002 IEEE A sample optimal sequence of switches obtained, by applying genetic algorithms, is (4 1 8 9 2 3 10 6 7 5) with global best restoration time of 275.14 minutes. IV. ACKNOWLEDGEMENTS The authors gratefully acknowledge the National Science Foundation for supporting the research through Grant No. 9721447. V. REFERENCES [1] O. Ramsaur, “A new approach to Cold Load Restoration”, Electric World, pp. 101-103, October 6, 1952. [2] J. E. McDonald, Armin M.Bruning, and W. R. Mahieu, “ Cold Load Pickup ”, IEEE Trans. Power Apparatus and Systems, Vol. PAS-98, pp. 1384-1386, July/August 1979. [3] IEEE Guide for Loading Mineral Oil-Immersed Transformers, ANSI/IEEE C57.91-1995. [4] C. Ucak and A. Pahwa, “An analytical approach for step-by-step restoration of distribution systems following extended outages”, IEEE Transactions on Power Delivery, Vol. 9, No. 3, July 1994. [5] D.E.Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning, Addison-Wesley publishing company, January,1989. [6] D.Whitley, Tim Starkweather and D’Ann Fuquay, “Scheduling problems and travelling salesman: the genetic edge recombination operator”, Proceedings of the Third International Conference on Genetic Algorithms, Palo Alto, Ca.: Morgan Kaufmann. Pp. 133-140, 1989.