Noninvasive Data-Driven Prediction of Reactor Power Field: An MLP-Based Approach

The Reactor Operation Digital Twin (RODT) is a numerical representation method designed for the operational phase of nuclear reactors, aiming to solve forward problems (such as state prediction) based on input parameters and inverse problems (such as parameter identification) utilizing additional measurement data, in order to enhance the economic benefits and safety of nuclear power. In this paper, we propose a solution for the forward problem in RODT based on a Multilayer Perceptron (MLP) approach, where the MLP is used to directly learn the mapping relationship from low-dimensional inputs to high-dimensional outputs. Compared to previous methods, the proposed approach shows improved prediction accuracy, especially in situations with limited training data, demonstrating superior generalization capabilities and accuracy. To validate our contributions, we conducted a comparative experiment based on a engineering case, namely the reactor core simulation in the life cycle of a real-world nuclear ractor, which confirmed the effectiveness of the proposed method.