Gaussian Process Based Prediction of Density Distribution in Core of Research Nuclear Reactor

Abstract

Research nuclear reactors operate in a partial refueling mode, which leads to the formation of local areas with high nonuniformity of power density distribution. Such areas impact the economic efficiency of fuel consumption and the reactor core reliability. This necessitates the power density distribution profiling and underscores the importance of identifying the patterns of power distribution formation within the heterogeneous structure of the reactor core. In this study, an analysis of the reactor’s operational experience under various fuel loadings was conducted, and the characteristics of power density distribution in each cell were determined. An approach to applying a machine learning model for predicting power density distribution nonuniformity across the fuel cells of the IRT-T reactor core is presented. It is shown that the application of the supervised learning concept and Gaussian process regression with combined covariance (kernel) function enables the prediction of power distribution parameters in each reactor cell, regardless of the specific loading pattern and fuel burnup depth. The model achieved an overall accuracy of over 99\(\%\), with a mean absolute error not exceeding 0.5\(\%\).