Most existing automatic train operation (ATO) models are based on different train control algorithms and aim to closely track the target velocity curve optimized offline. This kind of model easily leads to some problems, such as frequent changes of the control outputs, inflexibility of real-time adjustments, reduced riding comfort and increased energy consumption. A new data-driven train operation (DTO) model is proposed in this paper to conduct the train control by employing expert knowledge learned from experienced drivers, online optimization approach based on gradient descent, and a heuristic parking method. Rather than directly to model the target velocity curve, the DTO model alternatively uses the online and offline operation data to infer the basic control output according to the domain expert knowledge. The online adjustment is performed over the basic output to achieve stability. The proposed train operation model is evaluated in a simulation platform using the field data collected in YiZhuang Line of Beijing Subway. Compared with the curve tracking approaches, the proposed DTO model achieves significant improvements in energy consumption and riding comfort. Furthermore, the DTO model has more advantages including the flexibility of the timetable adjustments and the less operation mode conversions, that are beneficial to the service life of train operation systems. The DTO model also shows velocity trajectories and operation mode conversions similar to the one of experienced drivers, while achieving less energy consumption and smaller parking error. The robustness of the proposed algorithm is verified through numerical simulations with different system parameters, complicated velocity restrictions, diverse running times and steep gradients.