Multiplication Factor Prediction Feasibility Study for the BEAVRS Core Based on the LSTM Algorithm

Abstract

Abstract This paper focuses on exploring the feasibility of the LSTM (Long Short-Term Memory) algorithm in deep learning for effective multiplication factor keff prediction at the core level, modeled by BEAVRS (Benchmark for Evaluation And Validation of Reactor Simulations) core first cycle loading with keff of operating at full power for 0–300 days was used as the study subject. The first 65% of the dataset is the training and validation set, and the last 35% of the dataset is the prediction target. The training and alignment results of the physical parameters of the components were obtained using the DRAGON4.1 and DONJON4.1 codes, and the LSTM algorithm in deep learning was applied. By adjusting the number of LSTM cells, L2 regularization parameters, optimizer type, and other parameter coefficients in the algorithm, the results showed that the absolute error of the predicted core effective multiplication factor keff could be made within 2 pcm by adjusting the appropriate parameters, which validated the successful application of machine learning to transport equations.