GSTGAT: Gated spatiotemporal graph attention network for traffic demand forecasting

Abstract Urban traffic demand forecasting is an important component in the implementation of intelligent transport systems (ITS). Urban traffic demand data is a spatiotemporal data, describing the amount of traffic demand generated by different areas or stations within a city along the time dimension. Although there has been considerable research work, researchers still face several challenges in predicting accurately, including the capture of hidden features in the temporal dimension of such spatiotemporal data, and the capture of dynamic dependent changes in the spatial dimension. These are even more difficult for long‐time series prediction tasks. This paper designs a multivariate temporal forecasting model specifically adapted to traffic demand to address these challenges, called the Gated Spatiotemporal Graph Attention Network (GSTGAT). GSTGAT is based on the Transformer framework and the whole model is used in an end‐to‐end manner. First of all, it uses the gated self‐attention to extract temporal features in the sequence. Secondly, graph attention is used to capture the spatial dependencies among different variables in the unstructured space. Finally, the use of gated recurrent units in combination with hidden spatial states is proposed to capture multiple levels of spatial dependencies. Experimental results on the taxi dataset in New York and the bicycle dataset in San Francisco Bay Area show that the authors’ proposed model outperforms other state‐of‐the‐art models and improves the prediction accuracy.