Research On Train Automatic Stop Control Algorithm Based On Knowledge And Deep Reinforcement Learning

Over the past ten years,China has achieved rapid development in high-speed railway industry.Efficient operation of high-speed trains has become increasingly prominent in protecting people’s lives and promoting social development,and has been widely concerned.Train automatic stop control,as a vital function of automatic train operation system,has direct impacts on the operational efficiency of train,service quality of rail transit and passenger satisfaction.The current methods of train automatic stop control often face some practical problems,such as rough control rules obtained or very expensive cost for technology updates.Therefore,it is paramount to develop precise train automatic stop control algorithm to achieve ensure efficient train operation.This thesis proposes train automatic stop control algorithms based on knowledge and deep reinforcement learning.These algorithms transform the experience of excellent train drivers and experts into knowledge,and combine it with deep reinforcement learning to make full use of the perception of deep learning and the decision-making of reinforcement learning,and provide precise control rules for complex train automatic stop control system via the end-to-end learning process.The specific work of this thesis is as follows:(1)Modeling of high-speed train stop control and statement of performance indices.First of all,this thesis focuses on high-speed trains,and establishes train stop control simulation model based on dynamic train model and braking model,so that it can be used to simulate actual high-speed trains to execute related commands.Then,performance indices related to stopping error,comfort and number of commands change are described to evaluate the reliability of the proposed algorithms.(2)Train automatic stop control algorithm based on single train single line model.In the case of single train single line,this chapter proposes a train automatic stop control algorithm combining expert knowledge and double deep Q-network.First,the experience of excellent train drivers and experts is transformed into knowledge that is used to estimate the ideal commands executed by the train.Furthermore,double deep Q-network determines the execution time of the command.The proposed algorithm actually allocates the execution time of the execution command reasonably,and is compared with the traditional train automatic stop control algorithms,which verifies that the proposed algorithm can achieve high stopping precision.(3)Train automatic stop control algorithm based on multiple trains multiple lines model.In the multiple trains and multiple lines model,in order to solve the problem of optimal control parameters corresponding to the model and considering that different types of data are suitable for feature extraction with different neural networks,inspired by deep recurrent Q-network algorithm and Wide & Deep model,a multi input single output neural network is designed based on the algorithm in(2),and then a more general and convenient automatic train stop control framework is proposed.Finally,simulation experiments verify that the proposed algorithm can achieve precise stopping results.