In railway systems, the train dynamics are usually affected by the external environment (e.g., snow and wind) and wear-out of on-board equipment, leading to the performance degradation of automatic train control algorithms. In most existing studies, the train control models were derived from the mechanical analyzation of train motors and wheel-track frictions, which may require many times of field trials and high costs to validate the model parameters. To overcome this issue, we record the explicit train operation data in Beijing Metro within three years and develop three data-driven approaches, involving a linear regression-based model (LAM), a nonlinear regression-based model (NRM), and furthermore a deep neural network based (DNN) model, where the LAM and NRM can act as benchmarks for evaluating DNN. To improve the training efficiency of DNN model, we especially customize the input and output layers of DNN, batch normalization based layers and network parameter initialization techniques according to the unique characteristics of railway train models. From the model training and testing results with field data, we observe that DNN significantly enhances the predicting accuracy for the train control model by using our customized network structure compared with LAM and NRM models. These data-driven approaches are successfully applied to Beijing Metro for designing efficient train control algorithms.