Topology augmented dynamic spatial-temporal network for passenger flow forecasting in urban rail transit

Millions of residents travel by urban rail transit (URT) every day, and it is meaningful to accurately predict passenger flows for the purpose of intelligent system management. Considering the high traffic volume and complex spatial-temporal traffic patterns, accurate forecasting is quite challenging. Recently, advanced spatial-temporal graph neural networks that combine graph convolution and recurrent units have achieved superior traffic forecasting performance. However, existing methods either simply use a predefined physical network or directly learn latent static graphs by assigning a large number of trainable parameters. Moreover, they only consider capturing local temporal dependencies. However, the passenger flows of URT exhibit both physical adjacent and virtual distant spatial correlations and both local and global temporal fluctuations. In this paper, we propose a Topology Augmented Dynamic Spatial-Temporal Network (TADSTN) to fully exploit the complex spatial-temporal dependencies of passenger flows and make accurate forecasting in URT. Specifically, we first propose a virtual graph generation algorithm with no training parameters to obtain a topology-augmented URT graph, and we also sample multipattern temporal features from raw passenger flow data. Then, we propose a parallel spatial-temporal learning architecture with a time-aware global-level attention mechanism to simultaneously capture both distant and dynamic spatial dependencies and both local and global temporal dependencies of passenger flows. Finally, we evaluate the effectiveness and efficiency of our TADSTN on two real-world datasets.