Prediction of Severe Accident Progression Using Machine Learning With Data-Driven Surrogate Modeling as Operator Support Tool

Abstract

Severe accidents in nuclear power plants (NPPs) pose critical challenges due to heightened environmental harshness that can impair instrumentation functionality. This impairment leads to "blind conditions," where operators lack essential thermal-hydraulic data, hindering decision-making during pivotal moments, as exemplified by the Fukushima Daiichi accident. To address this, Operator Support Tools enhancing nuclear safety are essential for substituting failed instruments, requiring reliability, prompt responsiveness, and situational resilience. This study proposes a deep learning-based surrogate methodology to predict severe accident progression in real-time, enhancing Operator Support Tool capabilities. By constructing a comprehensive dataset using the Modular Accident Analysis Program (MAAP) 5.0.3, the surrogate model approximates complex severe accident analysis codes without the computational burden. Advanced deep learning models, including Transformer and Mamba architectures, are employed to handle multivariate time series forecasting of thermal-hydraulic variables and reactor pressure vessel (RPV) status with variable-length inputs. The developed surrogate models enable rapid and accurate prediction of key variables, operating on portable devices and meeting the Operator Support Tool requirements. This approach advances previous work by improving accuracy through state-of-the-art methodologies and enhancing flexibility in input handling. Performance evaluations demonstrate the models' effectiveness in supporting operators during severe accidents, mitigating blind conditions, and contributing to the safety and resilience of operations.