Study On Automatic Train Operation Algorithm Towards Uncertain Systems

This dissertation studies the automatic train operation algorithms with consideration of the uncertainties of train dynamics.The study includes two aspects,firstly,for the speed control of high-speed train,based on the longitudinal multi-particle model,where the uncertain parameters,time delays,nonlinearities and randomness are explicitly included in the model,robust H_∞control criteria are proposed for the system.Secondly,for the optimal operation of heavy-haul freight train,by using the reinforcement learning algorithms,the agent can obtain the optimal policy for operating multiple electric locomotives only based on the sampled data,without any prior knowledge about train dynamics.In the environment,the errors of sensors and actuators are simulated.The main work and contributions of this dissertation are listed as following.1.High-speed train is modeled as a multi-input multi-output(MIMO)system with uncertain parameters,where the train mass,elastic coefficients of couplers,the coefficients of mechanical resistance and aerodynamic drag are all taken as uncertain parameters.The system is linearized at the equilibrium point.Based on Lyapunov stability theorem,the sufficient condition that ensures the stability and H_∞performance of the system is obtained,namely Criterion A.Based on the linearized model,the high-speed train is modeled as a MIMO system with uncertain parameters and single time-varying delay.Similarly,a H_∞control criterion is obtained by using a Lyapunov function,namely Criterion B.The simulation indicates that compared with Criterion A,Criterion B shows stronger stability when time delay exists.2.Considering the difference between the delays of vehicle units,and nonlinearities of the resistance during high-speed train operation,the high-speed train is modeled as a MIMO system with uncertain parameters,multiple time-varying delays,nonlinearities and randomness.Under the assumption that the upper bounds of the time delays and their derivatives are known,a Lyapunov-Krasovskii function is used to obtain H_∞control criterion,namely Criterion C.The simulation indicates that compared with an existing robust adaptive controller for high-speed train,it takes more time for the H_∞controller to achieve stability,but its requirements for traction/braking force and power are low,and it won’t cause high-frequency vibration.3.Under the assumption that the upper bounds of time delays are unknown,and those of the derivatives of time delays are known,another H_∞control criterion is proposed for the MIMO system with uncertain parameters,multiple time-varying delays,nonlinearities and randomness,namely Criterion D.The simulation indicates that compared with Criterion C,Criterion D has higher tracking accuracy,but the coupler force and the vibration of couplers are stronger,and the requirements for traction power and energy consumption are higher.4.The optimal policy for operating multiple electric locomotives in a heavy-haul train is obtained based on Q-learning and on-policy Monte Carlo control.Without the prior knowledge about train dynamics,these two algorithms can optimize multiple objectives,such as speed,coupler force and energy consumption,and satisfy several constraints on the control inputs,such as the traction/braking characteristics of electric locomotives,notch change rate.The control-response process of the train is simulated based on a parallel traction calculation framework,and the sampled data are used to estimate the action-value function.The optimal policies for heavy-haul train running on wave-like ramps are obtained.5.A novel action-value function approximator is proposed for the optimal control of multiple electric locomotives in a heavy-haul train,it is named Double-Switch Q-network(DSQ-network).The simulation indicates that compared with Q-learning and Monte Carlo control,DSQ-network can significantly improve the solving efficiency,enlarge the scale of the problem and denoise the action-value function.Based on 28 cases,the optimal policies for the heavy-haul trains operating on wave-like ramps are obtained,and the factors that influence the convergence rate of DSQ-network are discussed.