Development of Neural Network for Predicting Local Power Distributions in BWR Fuel Bundles Considering Neutron Absorber

Abstract

A neural network model is under development to predict the local power distribution in a BWR fuel bundle as a high speed simulator of precise nuclear physical analysis model. The relation between ²³⁵U enrichment of fuel rods and local peaking factor (LPF) has been learned using a two-layered neural network model ENET. The training signals used were 33 patterns having considered a line symmetry of a 8x8 assembly lattice including 4 water rods.

The ENET model is used in the first stage and a new model GNET which learns the change of LPFs caused by burnable neutron absorber Gadolinia, is added to the ENET in the second stage. Using this two-staged model EGNET, total number of training signals can be decreased to 99. These training signals are for zero-burnup cases. The effect of Gadolinia on LPF has a large nonliniality and the GNET should have three layers. This combined model of EGNET can predict the training signals within 0.02 of LPF error, and the LPF of a high power rod is predictable within 0.03 error for Gadolinia rod distributions different from the training signals when the number of Gadolinia rods is less than 10. The computing speed of EGNET is more than 100 times faster than that of a precise nuclear analysis model, and EGNET is suitable for scoping survey analysis.

KEYWORDS:

• neural networks
• BWR type reactors
• reactor core
• fuel bundles
• power peaking factor
• uranium enrichment
• uranium 235
• burnable neutron absorbers
• learning
• training signal
• Gadolinia
• two-staged model
• power distribution
• accuracy
• errors