With the acceleration of urbanization, the dynamic passenger flow has an ever-growing impact on the actual train operation. In this paper, we propose a learning based intelligent train regulation method with dynamic passenger flow prediction. To capture the characteristics of the dynamic metro passenger flow, a convolutional neural network is established to predict the real-time passenger flow from two dimensions including space and time. As the prediction accuracy is restricted by the insufficiency of the practical passenger flow data, a deep convolutional generative adversarial network is constructed to generate data that have the same distribution as the original passenger flow dataset. Then, by considering the effects of the dynamic passenger flow on the train operation and the train capacity constraints, the dynamic train regulation is formulated as a multi-stage optimal control problem with the objective function of minimizing the train traction energy consumption and the total traveling time of passengers. To efficiently obtain the optimal regulation strategy at each decision step, a deep Q-network algorithm is proposed to solve the formulated problem such the dimensionality curse caused by the excessive state space is avoided. The numerical experiments demonstrate the high efficiency and effectiveness of our proposed algorithm and model.