关键词: 深度强化学习, 核应急撤离, 动态撤离决策, 马尔可夫决策过程, MD-DQN

Abstract: Timely and effective evacuation of people during nuclear accident scenarios is critical to minimize radiation exposure and ensure public safety. Although traditional path planning algorithms can quickly compute static shortest paths, they are difficult to adapt to the challenges posed by dynamic changes in radiation fields. In this paper, a dynamic optimization decision model (MD-DQN algorithm model) for nuclear emergency evacuation based on deep reinforced learning is proposed. By establishing a Markov decision process (MDP) model, and taking the dynamic radiation field information, road network information, and real-time location as the state space, a multifactorial reward function that comprehensively considers the path length, radiation exposure and directional guidance is designed. The inteligent agent is driven to learn the optimal dynamic evacuation decision-making strategy autonomously. Meanwhile, the convergence and generalization performance of the algorithm are improved by optimizing the network structure design and instant reward mechanism. Simulation experiments show that compared with the traditional Dijkstra’s algorithm and A^* algorithm, the MD-DQN algorithm is able to effectively avoid high-risk areas in time, significantly reduce the radiation dose exposure of personnel in the evacuation process, and has better real-time path adjustment ability and environmental adaptability. The research results can provide an efficient, intelligent and decision support tool for practical nuclear emergency evacuation decision-making, and provide new research ideas for the future in the field of intelligent decision-making driven by multi-source radiation, multi-intelligent agent and real-time data.

Key words: deep reinforced learning, nuclear emergency evacuation, dynamic evacuation decision, Markov decision process, MD-DQN