Communication-Based Train Control System Performance Optimization Using Deep Reinforcement Learning

Communication-based train control (CBTC) systems are automated train control systems based on continuous and bidirectional train-ground communications. CBTC is the direction of future train control systems. Complex channel condition and frequent handoff can severely affect urban rail CBTC train-ground communication performance, and consequently affect CBTC operation efficiency. Current research regarding CBTC does not consider the impact of train-ground communications on train control performance. This paper studies the joint optimization of communication and control in CBTC systems. With the objective to minimize the optimal operation profile tracking error and energy consumption, linear quadratic cost is defined as the control performance measure. In order to ensure train operation safety, the optimization model puts constraints on train control actions related to safety. Moreover, based on the stochastic channel characteristics and the real-time train position information, handoff decision and train control policy are jointly optimized using deep reinforcement learning. Extensive simulation results based on real-field channel measurements illustrate that the proposed optimization method can significantly improve the train control performance in CBTC systems, and CBTC systems need to sacrifice part of performance to ensure system safety.