Improvement of a reinforcement learning model to assist operators in controlling a nuclear power plant under abnormal operating conditions

吉川 雅紀 ; 関 暁之*; 沖田 将一朗  ; 高屋 茂  ; Yan, X. 

Yoshikawa, Masanori; Seki, Akiyuki*; Okita, Shoichiro; Takaya, Shigeru; Yan, X.

Controlling a nuclear power plant (NPP) under abnormal operating conditions requires rapid and effective responses, which is particularly challenging for advanced NPPs owing to limited practical experience. To address this issue, a countermeasure proposal module (CMPM) based on a reinforcement learning model was previously proposed. The CMPM receives measurement data from an NPP under abnormal conditions and proposes countermeasures. In the previous study, the CMPM successfully proposed effective countermeasures. However, two main issues were identified: first, the CMPM occasionally recommended corrective actions even when the NPP was operating under normal operating conditions; second, it sometimes suggested opposing operations for components with the same function. In this study, these issues are addressed by improving the reinforcement learning model through the design of revised reward functions.