Efficient predictor of pressurized water reactor safety parameters by topological information embedded convolutional neural network

Abstract

Accurate forecasts for pressurized water reactor safety parameters are essential to ensure the safe operation of nuclear reactors. Potential of artificial neural networks on this task is limited owing to the lack of extracting the core location information. Sparse connections have unique advantages in discovering correlation between neighboring components and convolution kernels are designed to deal with two-dimensional information. In this paper, topological information embedded convolutional neural network (TCNN) was firstly established and utilized. This model enhanced the ability of fusing location features and component attributes through sparse connections and convolution layers. Datasets of China’s Qinshan Nuclear Power Plant Phase II PWR was used to evaluate the performance of TCNN. Comparative and ablation experiments demonstrated that TCNN has superiority in working as efficient predictor for pressurized water reactor safety parameters, indicating that the proposed model promoted the digitalization of nuclear power plants.