Development of Genetic Algorithms for Plant Reload Optimization for an Operating Pressurized Water Reactor

This paper summarizes the development, results, and enhancement activity of the Artificial Intelligence (AI) based automated nuclear power plant fuel reload optimization platform under the guidance of the United States Department of Energy, Light Water Reactor Sustainability Program, under the Risk-Informed Systems Analysis Pathway. The research focuses on the optimization of the fuel arrangement to maximize fuel utilization. The AI-based Genetic Algorithms work with both convex and non-convex, constrained or unconstrained problems. This can help explain the relationship between the fuel arrangement and fuel cycle length, in particular, the surrogate models used to reconstruct the Multiphysics problem maps the features/inputs of the problem to the fuel cycle length to provide such an explanation. The Genetic Algorithm is composed of several evolutionary processes: fitness evaluation, parent selection, crossover, mutation, survivor selection, and termination. Crossover and mutation are the main steps responsible for injecting randomness/heuristics to prevent the algorithm from getting stuck in local minima. In this paper, roulette wheel parent selection, one-point crossover, swap mutation, and fitness-based survivor selection are used for demonstration to convert the fuel arrangement problem from the physical world (phenotype space) to the computational word (genotype space) via a user performed encoding/decoding step. Here, the search variables (genes) are the fuel locations in the core, whereas the values each variable takes, represent the fuel identification that will be placed in that specific location. The optimization process was demonstrated with a quarter core initial loading problem. In the core, 56 locations are loaded with five types of fuel assemblies, each type has different amount of enrichment and burnable poisons. As a result, the fuel cycle length increased to over 590 days, which is very close to the expected value. The results and enhancements in the optimization algorithm are also discussed in this paper.