Vison-based Predictive Energy-efficient Train Control via Deep Reinforcement Learning

The Virtually Coupled Train Set (VCTS) represents a next-generation train operation paradigm that significantly enhances line capacity through cooperative control among multiple trains. However, existing coordination strategies primarily emphasize safety assurance while overlooking systematic energy optimization, and they heavily rely on inter-train communication, making them highly sensitive to latency and packet loss. To address these issues, this paper proposes a predictive energy-efficient control framework based on visual context perception and deep reinforcement learning (DRL), enabling the following train to achieve autonomous energy optimization under weak communication conditions. The proposed framework employs onboard cameras to capture forward track scenes and uses a variational autoencoder (VAE) to extract latent environmental features, constructing a compact state representation. A Soft Actor-Critic (SAC) algorithm is then applied to derive optimal control policies by jointly considering safety, energy consumption, and ride comfort, Simulation results demonstrate that, compared with the conventional model predictive control (MPC) method, the proposed approach effectively reduces energy consumption and improves average speed while maintaining formation stability.

Keywords: Virtually coupled train set, Deep reinforcement learning, Variational autoencoder, Soft Actor-Critic.