Transformer-based Spatial-Temporal Graph Attention Network for Traffic Flow Prediction

Abstract

Traffic flow prediction, which plays an important role in intelligent traffic systems, has become a pressing problem to be addressed with the continuous development of smart cities. Currently, the fundamental obstacle lies in effectively modelling the complex spatial-temporal dependencies present in traffic flow data. Deep learning models such as Graph Neural Network based models and Transformer based models have shown promising results in this field. However, methods founded on a single model or framework have one significant limitation: Such methods cannot adequately represent the spatial and temporal features of traffic flow data, restricting the model's ability to learn the dynamics of urban transportation. In this paper, we propose a transformer-based spatial-temporal graph attention network model called TSTGAT for traffic flow prediction, which integrates Transformer and Graph Attention Network. Experiments on two real-world traffic datasets from the Caltrans Performance Measurement System (PeMS) demonstrate that the proposed TSTGAT model outperforms well-known baselines.