Identification of a nuclear reactor core (VVER) using recurrent neural networks

Abstract

Recurrent neural networks (RNNs) in identification of complex nonlinear plants like nuclear reactor core, have difficulty in learning long-term dynamics. Therefore, in most papers in this area, the reactor core is used to identify just the short-term dynamics. In this paper we used a multi-NARX (nonlinear autoregressive with exogenous inputs) structure, including neural networks with different time steps and a heuristic compound learning method, consisting of off-line and on-line batch learnings. This multi-NARX was trained by an accurate 3-dimensional core calculation code. Network responses show that this procedure solves the difficulty in identification of complex nonlinear dynamic MIMO (multi-input multi-output) plants like nuclear reactor core, and can be used in fast prediction of nuclear reactor core dynamics behavior.