High-speed train automatic stopping control method based on deep reinforcement learning

Train Automatic Stop Control (TASC) system based on reinforcement learning(RL) is a crucial technology to improve self-driving performance for automatic train operation(ATO). Although significant progress has been achieved, existing RL-TASC systems still face challenges related to sparse rewards and lower control precision. In this paper, we design a train automatic stop strategy based on the Deep Q-Network (DQN) algorithm with both global rewards and single-step rewards to enhance the precision and training stability of RL-TASC systems. Initially, a single-point motion equation for the train was established and the state-action space was defined. Subsequently, a global reward function was designed based on stopping error. To address the challenges posed by sparse global rewards, which led to poor training effects, and imprecise stopping performance, a method for setting a single-step reward function based on expected velocity was proposed. Finally, through simulation experiments, the results showed that the improved control method could maintain the train stopping error within 0.05 meters. Comparing to the pre-improved model, the average value of stopping error decreased by 76.56% and the standard deviation of stopping error decreased by 67.74%, confirming the effectiveness of the improved method. This paper provides a high-precision and highly robust method for automatic stopping control in high-speed train ATO systems.