Generally speaking, the majority of existing ATO systems achieve the train operation problem by focusing on the energy-efficient SD trajectory calculation, real-time tracking methods and train station parking algorithms, separately. For example in Beijing subway, the process of train operations by ATO can be described as follows. Before a train departs from stations, the managers of the subway companies first put forward a set of SD trajectories of the train at every segment according to the timetable [23]. Then, when the train is running on a segment, the ATO controller calculates the appropriate traction or braking output in order that the train tracks the corresponding SD trajectory precisely. Finally, when the train approaches the next station, the ATO controller determines the braking rate and parks the train at the destination. However, due to the complexity of train dynamic models, involving time-varying traction and braking prosperities, variable resistances, etc., this multiobjective problem in Eq. (6) is difficult to be solved in most existing ATO systems by the classical optimization approaches. Besides, it usually takes a long time for subway engineers to calculate the optimal trajectory. This makes it infeasible when there are unexpected disturbances, and the timetable needs to be updated in real-time. Thus, the existing ATO systems are also lack of flexibility with disturbances caused by variable and uncertain train dynamic models.