Application of Neural Network-Genetic Composite Algorithm in Core Refueling Design for PWR

The neural network model was trained by large-scale data, to accurately predict effective multiplication factor (k_eff), component power crest factor (Rad) and rod power crest factor (FΔH) of the nuclear reactor core, which were used for core refueling optimization. The improved genetic algorithm can obtain the best solution quickly, and solve time-consuming and cost-effectiveness problem. In modeling of core loading mode, the binary vector was designed as the input parameter, which effectively reduced the neural network complexity and improved the prediction accuracy. In the search of optimal scheme, the genetic algorithm with unique crossover operator and selection operator ensured that the search results were in the feasible region, and improved the search efficiency. The theoretical analysis and numerical experiment results show that, one-hidden-layer adaptive BP network predicts k_eff data well, while three-hidden-layer adaptive BP network is more suitable for Rad and FΔH data. Then the ideal core refueling schemes are obtained by the genetic algorithm. These practices are expected to promote a further application of artificial intelligence algorithms in the nuclear industry.