复杂环境下基于DRL的列车驾驶自动控制方法研究  

DRL-based automatic control method for train driving in complex environments

作　　者：刘继成 苏璐 LIU Jicheng;SU Lu(Chengdu Institute Of Rall Transit,Chengdu 610065,China)

机构地区：[1]成都轨道交通职业学院,成都610065

出　　处：《自动化与仪器仪表》2025年第3期85-89,共5页Automation & Instrumentation

摘　　要：在现代轨道交通系统中,轨道交通的自动化控制面临着复杂环境带来的多重挑战。研究旨在通过改进深度强化学习算法来优化列车驾驶控制。以全连接前馈神经网络作为列车控制器,引入了ε-greedy策略进行指令和激活函数优化,最终提出了一种新型列车驾驶自动控制模型。实验结果表明,该新模型的奖励值均值最高接近6,敏感值波动范围最小为[-1,1.5]。同时,该新模型控制下的列车运行时间最短为161.37 s,晚点时间最小为0.38 s,停车误差最低为0.33 m,牵引能耗最低为17.52 kW/h。由此可知,研究所提新模型显著提高了列车自动控制的稳定性和效率,具有较好的节能效果,能够为列车自动驾驶控制技术的发展提供新的见解。In modern rail transit systems,the automated control of rail transit faces multiple challenges brought about by complex environments.The research aims to optimize train driving control by improving deep reinforcement learning algorithms.A fully-connected feed-forward neural network is used as the train controller,and an ε-greedy strategy is introduced for command and activation function optimization,and a new train driving automatic control model is finally proposed.The experimental results show that the new model has the highest mean value of reward value close to 6 and the smallest range of sensitivity value fluctuation of[-1,1.5].Meanwhile,the shortest train running time under the control of this new model is 161.37 seconds,the minimum late time is 0.38 seconds,the minimum stopping error is 0.33 meters,and the minimum traction energy consumption is 17.52 kW/h.It can be seen that the new model proposed by the research significantly improves the stability and efficiency of automatic train control,has better energy-saving effect,and can provide new insights for the development of automatic train driving control technology.

关 键 词：深度强化学习 列车 驾驶 自动控制 ε-greedy策略 

分 类 号：TP274.2[自动化与计算机技术—检测技术与自动化装置]