Ding, Tianyue (Tsinghua University), Li, Zhongcan (Tsinghua University), Wang, Chong (Tsinghua University), Dong, Wei (Tsinghua University), Yan, Xiang (CRSC Research & Design Institute Group Co.,Ltd.), Li, Wei (CRSC Research & Design Institute Group Co.,Ltd.), Cao, Junwei (Tsinghua University)

An Optimization Method for Marshalling Station Period Scheduling Plan based on Reinforcement Learning with Graph Neural Network

Scheduled for presentation during the Regular Session "S33a-Intelligent Control for Next-Generation Railway Systems" (FR-LM-T33), Friday, November 21, 2025, 11:30−11:50, Southport 3

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Marshalling stations are critical nodes in railway freight transportation, and their transportation efficiency is directly influenced by the period scheduling plan. However, most of the existing research on period scheduling plan optimization relies on traditional mathematical programming methods or heuristic optimization methods, whose computational speed cannot support online decision-making. Although the existing reinforcement learning-based optimization method enable online decision-making, it simplifies some environmental assumptions, making it impractical for real-world deployment. To solve this problem, a simulation model is first constructed to fully capture the core operational process of the real marshalling station. A new entity association graph structure is subsequently proposed that can more comprehensively describe the complex entity attributes than the existing reinforcement learning-based studies. Finally, based on the proposed graph structure, a reinforcement learning algorithm with graph neural network is designed for dynamically optimizing the period scheduling plan. Simulation experiments based on the real data of a typical Chinese marshalling station show that, compared with a genetic-algorithm-based period scheduling plan optimization method, the proposed method improves average number of dispatched railcars in three hours by 16.89%, with an average strategy generation time of 5.67 seconds.