Study On Mechanism And Application Of Stochastic Optimization Method Based On PWR Loading Pattern Optimization Benchmark

Because of extraordinarily large search space, it is still a challenging job to find a safe and economical loading pattern (LP) quickly so far. As stochastic optimization algorithms, their application to LP optimization is widely studied. For lack of complete information about solution space of the benchmark problem, there is not quantitative evaluation for the validity of their search and optimization capability of LP. In this paper, by using a PWR LP searching benchmark problem with all LPs enumerated and pre-evaluated, and by exploiting search efficiency of the defined feasible solution and quality of the searched optimal solution as evaluation indices, study on the mechanism of Simulated Annealing Algorithm (SA) and Genetic Algorithm (GA) for LP search and optimization is carried out, calculation of these indices is also performed. At last, a new LP optimization method is discussed.Firstly, by using a PWR LP searching benchmark problem with all LPs enumerated and pre-evaluated, the mechanism of SA for LP optimization is revealed and for the first time the searching efficiency of SA method is quantitatively evaluated. Investigations find that generally speaking, when objective function is defined, no matter what the initial LP or the cooling schedule is, search efficiency and quality of the searched optimal solution of SA will not vary dramatically. But for effect of random factor in the search process of SA, some exceptional phenomena will occur, the optimization results under these conditions will deviate from the average value rather far. Investigations also find that the constraint on the distribution of average k-inf value for a fuel assembly crisscross is quite efficient to screen the non-feasible patterns, very high searching efficiency can be achieved by a SA method adopting this technique.Secondly, by using a smaller benchmark problem with whole solution space information, study on the mechanism of GA for LP search and optimization is conducted. Optimization results of classical GA are calculated in the first instance. For short of guidance of effective physical mechanism in the generation of new LPs of classical GA, the algorithm tends to converge early, and its search efficiency is relatively low. In this study, the age technique, the concepts of relativeness degree and worth function are exploited to improve the performance of GA for PWR LP search. Among them, the age technique endows the algorithm be capable of learning from previous search"experience"and guides it to do a better search in the vicinity of a local optimal; the introduction of the relativeness degree checks the relativeness of two LPs before performing crossover between them, which can significantly reduce the possibility of prematurity of the algorithm; while the application of the worth function makes the algorithm be capable of generating new LPs based on the statistics of common features of evaluated good LPs. Numerical verification demonstrates that the adoption of these techniques is able to significantly enhance the efficiency of the GA while improves the quality of the final solution as well. In addition, the quantitative comparison of search efficiency of SA and GA is made based on the same smaller benchmark problem.Finally, a new LP optimization method is put forward. Because all the other processes are definite except that the initial generation is obtained randomly, this results in too small search space and premature convergence, so quality of the final optimal solution is not desirable. This new method still has much to be improved.