Autonomous Train Dispatch Method for Regional Rail Transit Based on Deep Reinforcement Learning

Train dispatching of rail transit is an important issue for implementing the transportation plan and ensuring the orderly operation. Recognizing that existing centralized dispatch methods lack efficiency and flexibility, a novel concept of autonomous train dispatch of rail transit is proposed. Additionally, the Deep Reinforcement Learning (DRL) is incorporated into the autonomous train dispatch (ATD) model, and the resultant DRL-ATD model considers both dispatch and control objectives. The Double DQN (DDQN) algorithm is used for the model training, and a prioritized experience replay mechanism is further introduced to improve the stability and convergence. Using a case study involving the Chongqing Rail Transit in China, it is found that the DRL-ATD model can ensure the safety and punctuality of rail transit trains; it also reduces energy consumption by 1.62% during low passenger volume periods and passenger waiting time by 6.83% during high passenger volume periods, compared with the existing centralized method. Sensitivity analysis of the dispatch policy shows that the DRL-ATD model can effectively balance energy consumption of train operations and the total waiting time of passengers by adjusting the train operation schedules automatically, thereby providing the advantages of flexibility and dynamic adaptability for train operation.