Bayesian Optimization for PWR-Core Loading Pattern Optimization

The optimization of loading patterns in pressurized water reactor (PWR) cores is essential for maintaining the safety, operational efficiency, and cost-effectiveness of nuclear power plants. This task poses a challenging combinatorial optimization problem characterized by nonlinear, non-convex, and integer constraints, which complicate the search for global optimal solutions. Traditional optimization techniques often struggle with low computational efficiency and a tendency to become trapped in local optima, limiting their effectiveness for complex core configurations. This study introduces an innovative loading pattern optimization approach that combines variational autoencoders, deep metric learning, and Bayesian optimization to overcome these limitations. Variational autoencoders convert discrete core layouts into a continuous latent space, facilitating more efficient exploration of potential solutions. Deep metric learning further structures this latent space, grouping configurations with similar physical characteristics closer together to improve the interpretability of the latent space and enhance search performance. Subsequently, a multi-objective Bayesian optimization process is used to effectively identify optimal core configurations within this structured latent space. The identified latent variables are then decoded back into discrete core layouts. Experimental validation using initial loading data from the first cycle of an M310 core demonstrates that this integrated approach significantly improves both the efficiency of loading pattern optimization and the quality of the resulting configurations, outperforming traditional methods and providing a more robust solution framework.