Practical Intelligent Automatic Train Operations which Optimize Switching Times Autonomously

Abstract

Automatic train operation (ATO) is gradually replacing manual driving because of its reliable operation on both high-speed railways and urban metro systems. However, several existing methods rely on a substantial amount of expert knowledge and real data to simulate train operations, resulting in limited flexibility and scalability. Additionally, given train depreciation or wear, certain algorithms may not always be adaptive, leading to potential strain on the braking system and compromised riding comfort. To this end, this paper proposes a novel Meta-ATO approach based on meta-learning and reinforcement learning, which was successfully applied to the real-world case of Xi’an Metro Line 9. Specifically, the meta-gradient boosting (MGB) method is presented to construct a high-fidelity simulation operating environment with fewer data. Meta-ATO enables trains to quickly perceive actual conflicts among multiple objectives and set control parameters through reinforcement learning. Based on this knowledge, a train can adjust its control strategy autonomously and optimize switching times. The results indicate that Meta-ATO provides good riding comfort, precise parking, and low energy consumption. Moreover, it no longer requires periodic adjustment by human experts, as it can optimize its control strategy and operations autonomously.