A simple accident management support tool based on source-term category using RNN-LSTM

Abstract

Severe accidents in nuclear power plants can cause significant damage to both human life and property. Due to the inherent complexity and uncertainty of severe accident progression, managing such accidents is challenging for operators. Consequently, computational aids are crucial in supporting their decision-making processes. Among these computational tools, data-driven approaches hold considerable promise by suggesting expected plant states. However, these methods often require large datasets to cover a wide range of scenarios. In this study, a simplified data-driven accident management support tool was proposed using Recurrent Neural Networks with Long Short-Term Memory (RNN-LSTM). The model predicts the consequences of severe accidents in terms of source-term categories based on nuclear power plant monitoring parameters. To assess the effectiveness and robustness of the suggested model, sensitivity analyses were conducted focusing on sensor failure, sampling intervals, duration, and noise levels. Results showed that the model's performance degraded with sensor failures, data scarcity, and increased noise but maintained meaningful performance overall. A notable observation was that denser time intervals generally enhance model performance; however, overly dense intervals can make the system vulnerable to errors. Thus, an optimal sampling interval for monitoring parameters is crucial to achieve the best performance.