Application of Deep Neural Network to an Accelerated Prediction of a Severe Accident in Nuclear Power Plants

Abstract

Recent nuclear severe accidents have spurred interest in the development of advanced accident management support tools (AMSTs) to enhance decision-making during crises. This study examines the efficacy of deep neural networks (DNNs) in accelerating severe accident predictions within nuclear power plants (NPPs), focusing on a loss-of-component-cooling-water (LOCCW) accident scenario. Through analysis of 10,780 simulated LOCCW accident scenarios across varied component failures and mitigation strategy implementations, time series datasets were synthesized at 15, 30, and 60-min intervals. The evaluation demonstrated that convolutional neural network (CNN)-integrated models outperformed standalone architectures in prediction accuracy across all temporal resolutions. Notably, higher temporal resolutions in training datasets significantly improved mean absolute error (MAE) and root mean squared error (RMSE), thereby enhancing prediction precision for immediate subsequent time steps. However, the augmentation of temporal resolution did not uniformly improve overall scenario prediction performance, as assessed by dynamic time warping (DTW) distance, due to cumulative prediction error in higher resolution models. These findings elucidate the nuanced relationship between temporal resolution and predictive accuracy, offering valuable insights for the development of sophisticated AMSTs aimed at bolstering nuclear safety and accident management strategies.