Learning spatial-temporal dynamics and interactivity for short-term passenger flow prediction in urban rail transit

Accurate short-term passenger flow prediction in urban rail transit is critical in ensuring the stable operation of urban rail systems. However, accurate passenger flow prediction still faces challenges, including modeling the dynamics of passenger flow data in spatial and temporal dimensions and capturing the interactions between the inflows and outflows. To solve these problems, a novel model called the multi-feature fusion graph convolutional network (MFGCN) is proposed. Firstly, parallel graph branch networks are established to describe inflow and outflow information from geographic and semantic perspectives. Then, in the spatial dimension, the graph convolutional networks with spatial attention are designed to learn the dynamic spatial correlations of nodes in the two graphs. In the temporal dimension, the long short-term memory networks with temporal attention are developed to learn the dynamic temporal dependencies of passenger flow data. Finally, a three-dimensional residual network is established to capture the spatial-temporal interactive dependencies between inflows and outflows. Experiments on Nanning Metro Line 1 passenger flow datasets demonstrated that MFGCN outperformed the existing baseline models, which could provide technical support for URT network operation management.