Application of Reinforcement Learning to Deduce Nuclear Power Plant Severe Accident Scenario

Abstract

Probabilistic Safety Analysis (PSA) is crucial for determining severe accident scenarios in nuclear power plants. While it allows the identification of component failure sequences, evaluating the impact of component failure time on accident severity remains challenging. This study introduces novel approach that employs machine learning, specifically reinforcement learning (RL), to complement traditional PSA. Validating this approach involved comparing severe accident scenarios obtained through accident simulations with those reproduced by RL. The comparison demonstrates the feasibility of exploring critical accident scenarios using RL. To implement RL based on the existing system code, supervised learning model predicting the remaining time of reactor vessel failure was developed. Using this prediction model and data from catastrophic accident simulations, RL was implemented and results validated against severe accident code simulations. In summary, this study presents a new methodology for applying machine learning to nuclear accident analysis, discussing the feasibility and potential of the proposed approach.

Keywords: severe accident, severe accident simulation, severe accident scenario, machine learning, supervised learning, reinforcement learning