Prediction of safety parameters of pressurized water reactor based on feature fusion neural network. (February 2022)

• Record Type: Journal Article
• Authors: Chen, Yinghao; Wang, Dongdong; Kai, Cao; Pan, Cuijie; Yu, Yayun; Hou, Muzhou
• Is Part Of: Annals of nuclear energy. Volume 166(2022)
• Publication Date: 2022-02
• Subjects: Pressurized water reactor -- Feature fusion -- Neural network -- Subassembly power peaking factor -- Rod power peaking factor -- Limited effective multiplication factor; 65K05 -- 68T15 -- 68U20 -- 68W25; Nuclear energy -- Periodicals; Nuclear engineering -- Periodicals; 621.4805
• DOI: 10.1016/j.anucene.2021.108803
• Languages: English
• ISSNs: 0306-4549

Abstract

Highlights: We propose a feature fusion neural network model for data regression tasks. We use deep learning to predict the safety parameters of nuclear reactors. Model's presentation ability enhanced by residual connection and feature fusion. Abstract: Rapid and accurate calculation of safety parameters based on the state and layout of the core is essential to ensure the safe operation of nuclear reactors. Based on artificial intelligence technology, we established a neural network model with seven layers, called the feature fusion neural network (FFNN). A residual compensation layer is added in the third hidden layer to avoid gradient dispersion, and a feature fusion layer is added in the last hidden layer to fuse low-dimensional features and high-dimensional features. Numerical experiments simulate the pressurized water reactor of Qinshan Nuclear Power Station, and 10, 000 sets of safety parameters are calculated by CASMO5 under different arrangements as ground truth. Experiments show that the mean relative error of the subassembly power peaking factor and rod power peaking factor estimated by FFNN is about 1%, and the maximum relative error of the limited effective multiplication factor is within 0.5%.