Enhancing nuclear safety through automated anomaly response in nuclear power plants

Dong, F.*; 出町 和之*; 高屋 茂  ; 吉川 雅紀 

Dong, F.*; Demachi, Kazuyuki*; Takaya, Shigeru; Yoshikawa, Masanori

Ensuring nuclear safety in nuclear power plants requires timely anomaly detection and rapid, effective responses. Conventional methods, relying on operator knowledge and procedural manuals, fail to handle unexpected anomalies and are vulnerable to human error. This study presents an automated response framework using deep reinforcement learning (DRL) to overcome these limitations. A surrogate model, capable of predicting reactor conditions based on input time-series data, supports efficient learning and broadens the exploration space for policy optimization. The proposed framework was validated using the ACCORD code for the high temperature gas cooled reactor, where the DRL agent successfully proposed countermeasures to restore the system to a near-normal stable state. AI-based approaches promise to strengthen anomaly response mechanisms and improve NPP safety.