Surrogate Model of Predicting Eigenvalue and Power Distribution by Convolutional Neural Network

Abstract

During loading pattern (LP) optimization and reactor design, a lot of time consumption spent on evaluation is one of the key issues. In order to solve this issue, the surrogate models are investigated in this paper. The convolutional neural network (CNN) and fully convolutional network (FCN) are adopted to predict the eigenvalue and the assembly-wise power distribution (PD) for a simplified pressurized water reactor (PWR) during depletion, respectively. For the eigenvalue prediction during depletion, the error in the begin of cycle (BOC) and middle of cycle (MOC) is higher than that in the end of cycle (EOC). For the BOC and MOC, the samples with discrepancy over 500 pcm are less than 1%, except four burnup points. For the EOC, the fraction of samples with error over 500 pcm is less than 1%. As for the error of assembly power, the average absolute error is on the same level for all test cases. The average absolute relative error in the center region and the peripheral region is higher than that in the inter-ring region. The prediction results indicate the capability of neural network to predict core parameters.