Dynamic Decoupling Control for Virtually Coupled Train Set: A Reinforcement Learning-Based Model Predictive Control Approach

This paper presents a novel dynamic decoupling control method designed to manage the decoupling operations of virtually coupled train sets (VCTS) within the throat area of a railway line. The primary objective is to dynamically adjust the spacing between trains within the constraints of safety, thereby enhancing the efficiency and smoothness of decoupling maneuvers. Traditional controllers with fixed parameters often fail to effectively manage the complex constraint challenges inherent in such scenarios. To address the multi-objective and nonlinear nature of this constrained control problem, we develop a reinforcement learning (RL) based model predictive control (MPC) approach. The MPC framework is proficient at handling stringent constraints imposed by speed limits, relative braking distances, and signaling protection requirements. The RL-based strategy is capable of learning and fine-tuning MPC parameters in response to varying line conditions and train dynamics parameters by calculating the gradient from the MPC cost. Numerical experiments demonstrate that this approach significantly improves safety while reducing the decoupling time of VCTS from 59.6 seconds to 47.4 seconds after convergence.