A multiple spatio-temporal features fusion approach for short-term passenger flow forecasting in urban rail transit

Timely and precise short-term passenger flow forecasting (STPFF) is a key to intelligent urban rail transit (URT) operation. The uncertainty of passenger travel behaviors leads to passenger flow data showing high nonlinearities and complex spatio-temporal correlation. Therefore, the ability to capture such correlations determines the prediction performance of the model. However, current methods either consider only the physical topology of URT network or lack the ability to model global spatial dynamics, making it difficult to obtain satisfactory predictions. Therefore, a Multiple Spatio-Temporal Features Fusion (MSTFF) network is proposed for STPFF in URT networks. Specifically, the connectivity and passenger flow similarity relationships in URT networks are encoded as physical adjacency graph and virtual functional similarity graph, respectively, and then these two complementary graphs are input into two designed Residual Graph Convolutional Gated Recurrent Unit (RGCGRU) modules to extract their spatio-temporal features. Second, an Attention-based Gated Recurrent Unit (AGRU) module is designed to capture global dynamic evolutionary trends in both spatial and temporal dimensions. Finally, a fully connected neural network is used to fuse these features. Experiments on two real passenger flow datasets of URT show that MSTFF performs well on various prediction tasks (15, 30, 45, and 60 min).