Dynamic adjustment control of train tracking based on reinforcement learning with event-driven and attention mechanisms

A dynamic adjustment control method of train tracking based on reinforcement learning is proposed to address the issues of low tracking efficiency and poor anti-interference ability in the fixed speed mode of urban rail transit. The comprehensive optimization goals include train running safety, tracking efficiency, energy consumption, and comfort. The proposed method utilizes Deep-Q-Network (DQN) algorithm and introduces an attention mechanism to extract important features of the state of the front and rear trains in different running stages. Event triggering conditions are designed based on the joint state of the front and rear trains to drive operation condition transformation, thereby achieving dynamic optimization of train tracking interval. Simulation results demonstrate that compared with fixed speed mode, the proposed algorithm significantly improves departure interval by 24.39% and maximum tracking distance by 49.81%. Additionally, it shows excellent performance in energy saving and comfort. When compared with Model Predictive Control (MPC) algorithm, the proposed algorithm optimizes both comfort and energy consumption while ensuring driving safety.