【作者】 李亚；

【导师】 付文秀；

【作者基本信息】 北京交通大学 ， 控制工程（专业学位）， 2021， 硕士

【摘要】 城市轨道交通(城轨)作为我国大中型城市公共交通的核心,是市内通勤人员出行的首选方式。近年来,随着城轨的快速发展,城轨客运量呈逐年增长的趋势。在城轨运营网络已成型的条件下,利用先进技术进一步缩短列车追踪间隔、提高行车密度、充分释放线路运能是满足城轨客流需求的重要手段。现有的城轨采用基于通信的列控系统实现的是“撞硬墙”的移动闭塞技术。该方式由于只考虑前车当前的静态位置信息,在一定程度上限制了线路运能的释放。基于车车通信的新型列控系统实现的是“撞软墙”的理想移动闭塞技术,在充分考虑前车位置、速度、加速度等更多动态信息,能够进一步缩小列车运行间隔、提升系统运力。论文针对新型列控系统的列车控制方式,基于强化学习、模糊控制等算法,对列车区间追踪间隔控制问题进行研究,论文的主要工作如下:(1)研究基于车车通信的新型列控系统的系统结构和主要功能,分析新型列控系统与传统列控系统在系统结构、通信方式及列车控制等方面的差异。根据城轨列车特点建立列车动力学模型,分析列车ATO控制性能指标为列车间隔控制指标的制定提供参考。(2)基于新型列控系统的“撞软墙”技术构建列车间隔控制系统模型,采用蒙特卡洛树搜索算法(Monte Carlo tree search,MCTS)对列车区间运行间隔控制问题进行研究。MCTS算法用于生成列车动作序列,实现对列车未来运行轨迹的预测;采用动态规划算法(Dynamic Programming,DP)对动作序列处理,以获得更具远见性的列车控制策略;在判定列车运行安全的前提下,执行列车控制策略。建立列车间隔控制指标,对基于MCTS的列车间隔控制算法进行评价。(3)针对在线的MCTS算法在列车当前状态距离控制目标较远时所存在的运算效率问题,引入模糊控制对算法进行改进。采用二维Mamdani型模糊控制器对模糊控制算法进行设计,并基于MCTS算法的运算数据生成模糊规则表。形成了以强化学习和模糊控制为核心的集成智能列车间隔控制方法。(4)仿真对比三种不同的控制策略生成方法,确定采用DP算法生成MCTS列车控制策略。对MCTS算法的不同搜索树深度进行仿真,确定树深度为7时算法的效率和控制精度最佳。为验证本文提出的控制算法实际性能,首先对列车区间追踪的多个运行场景进行仿真。然后,基于某一列车真实运行数据进行追踪间隔仿真。最后,针对同一列车追踪场景分别采用本文控制算法和模糊控制算法,进行列车追踪间隔仿真对比。论文研究结果表明,基于强化学习的新型列控系统区间行车间隔控制方法,在解决部分可观测、高维度、复杂非线性等控制问题方面,相比传统的控制算法具有响应速度快、调节时间短、控制精度高等特点,适合于新型列控系统区间行车间隔控制。图42幅,表8个,参考文献80篇。更多还原

【Abstract】 As the core of public transportation in large and medium-sized cities in China,urban rail is the first traveling choice for urban commuters.In recent years,with the rapid development of urban rail,the passenger volume of urban rail has increased sharply.Under the condition that the urban rail operation network has been formed,using advanced technology to further shorten the train tracking interval,improving the traffic density,and fully releasing the line capacity are important means to meet passenger demand.The existing urban rail uses Communication Based Train Control System(CBTC)to realize the moving block technology of "concrete hard wall".This method only considers the static position of the train ahead,which limits the release of line capacity.The VehicleBased Train Control System(VBTC)realizes the ideal moving block technology of "concrete soft wall",which considering more dynamic information such as the position,speed,and acceleration of the train ahead,and the train operation interval can be further reduced and the system capacity can be improved.The thesis focuses on the train interval control method,based on algorithms such as Reinforcement Learning and Fuzzy Control.The main work of this paper is as follow:(1)Research on the system structures and the main function of VBTC and analyze the differences between VBTC and CBTC in system structures,communication mode and train control.Establish the train dynamics model according to the characteristics of urban rail train,and analyze the ATO control performance to provide reference for the formulation of train interval control index.(2)Building the model of the train interval control system based on the VBTC "concrete soft wall" technology,and use Monte Carlo Tree Search algorithm(MCTS)to research the control problem of train interval.Using MCTS algorithm to generate train action sequences,and to realize the prediction of the future track of trains;using Dynamic Programming algorithm(DP)to process the action sequences to obtain a far-sighted train control strategy;Judging the safety of the train operation and adopt the train control strategy.Establish the train interval control index,and evaluat the train interval control algorithm.(3)Introducing Fuzzy Control to improve the problem that MCTS algorithm has low computing efficiency when the current state of the train is far from the control target.Designing Fuzzy Control algorithm by using a 2-D Mamdani fuzzy controller and generating fuzzy rule table based on the calculation data of MCTS algorithm.Forming an integrated intelligent train interval control method with reinforcement learning and fuzzy control as the core.(4)The simulation compared three different control strategy generation methods,and decided to use DP algorithm to generate MCTS train control strategy.The simulation of different tree depths of MCTS algorithm shows that the algorithm has the best efficiency and control precision when the tree depth is 7.Firstly,in order to verify the actual performance of the control algorithm,multiple running scenarios of train interval tracking are simulated.Then,based on the actual running data of a train,the tracking interval is simulated by using the control algorithm.Finally,the proposed control algorithm and the Fuzzy Control algorithm are used to compare the simulation of the train tracking interval in the same tracking scene.The thesis research results show that the new interval control method based on Reinforcement Learning in the VBTC is suitable for train interval control.Compared with the traditional control algorithm,it has the characteristics of fast response speed,short adjustment time,and high control precision in solving the control problems such as partially observable,high dimensional,and complex nonlinear control problems.Figures 42,Tables 8,References 80更多还原