Accurate subway passenger flow prediction is crucial to operation management and line scheduling. It can also promote the construction of intelligent transportation systems (ITS). Due to the complex spatial features and time-varying traffic patterns of subway networks, the prediction task is still challenging. Thus, a hybrid neural network model, GCTN (graph convolutional and comprehensive temporal neural network), is proposed. The model combines the Transformer network and long short-term memory (LSTM) network to capture the global and local temporal dependency. Besides, it uses a graph convolutional network (GCN) to capture the spatial features of the subway network. For the sake of the stability and accuracy for long-term passenger flow prediction, we enhance the influence of the station itself and the global station and combine the convolutional neural networks (CNN) and Transformer. The model is verified by the passenger flow data of the Shanghai Subway. Compared with some typical data-driven methods, the results show that the proposed model improves the prediction accuracy in different time intervals and exhibits superiority in prediction stability and robustness. Besides, the model has a better performance in the peak value and the period when passenger flow changes quickly.

Traffic forecasting has always been an important part of intelligent transportation systems. At present, spatiotemporal graph neural networks are widely used to capture spatiotemporal dependencies. However, most spatiotemporal graph neural networks use a single predefined matrix or a single self-generated matrix. It is difficult to obtain deeper spatial information by only relying on a single adjacency matrix. In this paper, we present a progressive multi-graph convolutional network (PMGCN), which includes spatiotemporal attention, multi-graph convolution, and multi-scale convolution modules. Specifically, we use a new spatiotemporal attention multi-graph convolution that can extract extensive and comprehensive dynamic spatial dependence between nodes, in which multiple graph convolutions adopt progressive connections and spatiotemporal attention dynamically adjusts each item of the Chebyshev polynomial in graph convolutions. In addition, multi-scale time convolution was added to obtain an extensive and comprehensive dynamic time dependence from multiple receptive field features. We used real datasets to predict traffic speed and traffic flow, and the results were compared with a variety of typical prediction models. PMGCN has the smallest Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) results under different horizons (H = 15 min, 30 min, 60 min), which shows the superiority of the proposed model.