高速列车自主追踪的自适应强化学习虚拟编组控制方法

Adaptive Reinforcement Learning Control of Virtually Coupled High⁃speed Trains for Autonomous Tracking

虚拟编组技术是轨道交通领域近年来的研究热点,是进一步提升线路运行效率的有效手段。虚拟编组控制系统通过突破传统闭塞间隔的方式设计极限边界的行车间隔运行,借助AI、车车通信、智能计算等先进方法与技术,实现基于前车信息的自主安全间隔保持与追踪控制。列车运行动力学通常具有时变、随机、未知等特点,同时,虚拟编组的高速列车耦合动态更为复杂,给车载控制方法的智能性与收敛性带来了更大挑战。本文给出一种适用于高速列车自主追踪的虚拟编组控制方法,基于反步设计和Actor⁃Critic强化学习框架,设计哈密顿⁃雅可比⁃贝尔曼方程解析解的神经网络在线逼近方法,针对列车动力学模型的未知参数问题,设计估计误差最速下降的梯度算法,并开展基于京沪线高速列车运行数据的虚拟编组控制仿真,为后续列控技术的智能化水平提升提供了理论基础。

Virtual coupling(VC)has been recognized as research hot⁃spot in railway areas and an effective way to further improve the efficiency of rail lines.In VC⁃empowered control systems,the tracking headway among trains is further reduced by breaking the designed limit boundary of traditional blocking interval,meanwhile,e⁃mergent advanced methods,such as AI,train⁃to⁃train(T2T)communication,intelligent calculation,make it possible that high⁃speed train runs autonomously with safe interval maintenance and tracking control on basis of available information of preceding one.It is known that train dynamics is generally time⁃varying,stochastic and unknown,such status becomes much more complicated for VC high⁃speed trains,which brings challenge to guar⁃antee the intelligence and convergence for on⁃board control equipment.In this paper,a VC control method is de⁃signed for high⁃speed trains autonomous tracking,neural networks⁃based approximation method is proposed to on⁃line solve the Hamilton⁃Jacobi⁃Bellman equations using backstepping and actor⁃critic reinforcement learning,steepest descent gradient algorithm is designed regarding to the unknown parameters of train dynamics.Simula⁃tion results are given using collected data Beijing⁃Shanghai high⁃speed railway,which provides theoretical foun⁃dation for the intelligent improvement of train control technology in the future.