An Energy-Efficient Train Control Approach Based on Deep Q-Network Methodology

Application of the automatic train operation can efficiently reduce the energy consumption under the premise of ensuring safety and punctuality. Most of the traditional research mainly focused on solving the driving strategy with the given model or fixed parameters. To flexibly respond to the dynamic changes during the train operation, this paper proposes a data-driven based approach, specifically a reinforcement learning (RL) approach, which can optimize the train control strategy without using the prior knowledge of train dynamics and the predesigned velocity profile. The optimization objective is to minimize the energy consumption under the constraints of punctual arrival and speed restrictions. Numerous driving experiences are used to train the deep Q-network (DQN) until it approximates the optimal action-value function. The trained neural network can output the values of all actions with arbitrary input state such that the energy-efficient control strategy can be obtained according to action values. Based on the real-world data, some cases are conducted to illustrate the effectiveness of the proposed approach.