The passenger flow of subway station during the period of large-scale activities is a kind of predictable unconventional passenger flow. This paper proposed a short-term passenger flow prediction model suitable for metro stations during the period of planned special events (PSEs) based on the deep learning framework, combining the real-time AFC passenger flow data, PSE data and historical passenger flow rules. Firstly, the historical passenger flow data is split and noise reduction is processed, and the characteristics of active passenger flow are analyzed. Then, based on the deep learning framework, a multi-layer convolution neural network (CNN) is constructed to fit the mapping relationship between the characteristics of active passenger flow and the spatial-temporal distribution of passenger flow, and the adaptive moment estimation algorithm (Adam) is selected to optimize the training process, so as to apply to the situation that the passenger flow is concentrated in the terminal when the special event is over. Finally, taking Beijing Olympic Park subway station as an example, the accuracy of the model is verified by the measured data. The prediction results show that the Adam-CNN model established in this paper has higher accuracy than the general time series method ARMA and the traditional neural network SGD-CNN model, and can provide more favorable support for the organization of planned special events.