A digital twin-based system for full-lifecycle safety management and dynamic risk assessment in nuclear power plants

Abstract In order to respond to the complex safety problems involved with the lifecycle of Nuclear Power Plants (NPPs) and to eliminate the bottleneck of traditional static risk assessment, this paper proposes a new safety management framework based on Digital Twin (DT) technology and deep reinforcement learning. Legacy paradigms of safety management fail in high-risk, dynamic environments, such as planned refueling outages and unplanned fire events, due to the latency of the data and rigid decision-making. To this end, the current work proposes first the "Nuclear Safety Twin Construction and Integration Framework" (NST-CIF) to build systematically a high-fidelity virtual environment time-synchronized with the real plant and coupled across multiple physics domains. With large-scale experiments performed in the Digital Twin environment on two critical scenarios—dynamic refueling outage scheduling and emergency response to fire accidents—the results indicate that Proactive Safety Policy Optimization (PSPO) completely outperforms both traditional procedure-based policies and naive deep reinforcement learning algorithms both in operational efficiency and safety. This research presents a tested and potentially effective technological pathway for the transformation of NPP safety management from passive, reactive to intelligent, proactive, and predictive.