摘要
Digital twins (DTs) enhance the performance and resilience of complex systems such as nuclear power plants (NPPs). By leveraging machine learning and optimization techniques, DTs analyze real-time data to provide actionable insights for operations and maintenance (O&M). These insights facilitate optimal control solutions that reduce failures, increase adaptability, enhance safety, and minimize human error. Despite their potential, challenges remain in implementing DTs, particularly in generating reliable O&M recommendations. This study addresses this issue by presenting a genetic algorithm-based approach to optimize control setpoints in accident mitigation systems during medium break loss of coolant accidents (MBLOCAs) at the Maanshan NPP. Human error events influencing the core damage frequency (CDF) of MBLOCAs are identified from the probabilistic safety assessment report, with a focus on refining cooldown and depressurization (CND) actions. These actions are simulated using RELAP5-3D. Two strategies are proposed: one aimed at reducing CND time in successful cases and another designed to convert failed cases into successful ones. The results indicate that the first strategy shortens depressurization time by approximately 45 %, while the second ensures successful CND in previously failed cases, reducing total CDF by nearly 90 %. These findings demonstrate the effectiveness of the proposed strategies in mitigating MBLOCAs and advancing nuclear safety.