基于深度强化学习的大型高铁站到发线运用方案快速编制方法OA

Large high-speed railway stations serve as critical hub nodes in high-speed rail networks,where the rational arrangement of station track utilization plans constitutes the core of station transportation organization.To enable rapid generation of track utilization schemes for large high-speed stations,this study proposed applying deep reinforcement learning from artificial intelligence to address track utilization problems.Firstly,based on the concept of train passing routes,this paper achieved collaborative optimization of track allocation and route arrangement.The track utilization problem was formulated as a maximum independent set problem with the objective of accommodating all trains,leading to an integer programming model.Subsequently,a conflict-graph-based reinforcement learning environment model was constructed to transform the mathematical model of track utilization into a Markov Decision Process.This paper defined the state representation,action space,state transition,and reward function to establish an interactive data collection framework for agent training.Finally,a graph convolutional network(GCN)-based intelligent agent architecture was designed,combined with proximal policy optimization(PPO)algorithm for efficient pre-training.During the decision phase,a 2Imp-based decision enhancement algorithm was proposed to improve solution quality.The case studies of Beijing South High-Speed Field and Guangzhou South Station were conducted to compare the solving time and final objective function values of three different methods.The results demonstrate that after 4.2 hours of pre-training,the agent achieves convergence and rapidly obtains optimal solutions during the decision phase.Compared to classical operations research algorithms,the proposed intelligent algorithm can achieve equivalent solution quality with 5 to 16 times faster,completing the track utilization planning for 1 133 trains at Guangzhou South Station within 1.2 seconds.Moreover,the proposed method can exhibit strong scalability and transferability,enabling the agent to be directly deployed to stations of different scales and configurations for decision-making and problem-solving after a single training phase.These findings can provide technical support for enhancing the overall efficiency of train timetable preparation.The results can offer methodological references for the intelligent development of railway transportation organization.