Optimization of train Real-time dynamic scheduling under unexpected events in High-speed railway based on MAPPO

Abstract With the expansion of high-speed railway network scale and the increase of train operation density, the propagation effect of train delays triggered by emergencies such as equipment failures, natural disasters, and transportation safety incidents has become increasingly prominent. Traditional methods like the Mixed Integer Programming (MIP) model and fireworks algorithm optimization, while effective in stable environments, struggle to handle dynamic multi-train interaction scenarios. This paper innovatively proposes a dynamic scheduling framework based on multi-agent proximal policy optimization, aiming to minimize train delays. Through a distributed decision-making architecture, each train operates as an independent agent, leveraging 11-dimensional state features to achieve autonomous decision-making under local observation. A composite reward mechanism is designed, integrating punctuality incentives, safety distance constraints, overspeed rules, and rewards/penalties for minimum dwell time. A centralized training-decentralized execution collaborative paradigm is established to optimize group behavior coordination. Empirical studies demonstrate that this method significantly reduces delay propagation under unexpected events, achieving an average station punctuality rate of 56.3%, an average destination punctuality rate of 91.7%, and a destination punctuality rate of 100% for key trains, verifying its engineering applicability in complex transportation systems.