Real-time prediction of nuclear power plant parameter trends following operator actions

Abstract

Operators in the main control room of a nuclear power plant (NPP) oversee all plant operations, and thus any human error committed by the operators can be critical. If the operators can be informed about the future trends of the significant plant parameters that will follow their actions, they will be able to detect a human error in a short time or even prevent it. Future parameter trends, in addition, can be used to confirm the appropriate operational plan and support accident diagnosis. To achieve fast and accurate future parameter trend prediction in NPPs, we propose a data-driven prediction model composed of a multi-step prediction strategy and artificial neural networks. To find the optimal model performance, we applied a multilayered perceptron, vanilla recurrent neural network, and long short-term memory (LSTM) network, and trained the various candidate models with emergency operation data generated from an NPP simulator. Application results showed that the prediction model with the multi-input multi-output strategy and LSTM networks was able to successfully address the multivariate problem of future parameter trend estimation considering operator action in multiple emergency situations. It is believed that the proposed model may support NPP operators in coping with human errors and diagnosing accidents.