An Informer-based Spatio-Temporal network for traffic flow forecasting

Spatio-temporal data is one type of multivariate time series(MTS), which has strong time dependence and complex spatial correlation. Spatio-temporal data prediction requires mining the temporal-spatial relations implied in the data to obtain trends in the data as they change over time and space. Traffic flow is a typical class of spatio-temporal data, due to the complexity of the traffic road network, the high spatio-temporal correlation of road conditions, and the strong coupling of traffic information in time and space, which makes it much more difficult to mine the correlation of data from traffic flow. However, existing traffic flow forecasting methods often focus only on its temporal attributes leading to the loss of spatial information or can only effectively handle short-time data, To solve this problem, we propose a spatio-temporal data network, which uses Informer to process the masked spatio-temporal data to enhance its data mining capability, and subsequently using spatio-temporal graph neural network (STGNN)to complete the spatial data processing. In this way, we achieve effective processing of long-time data. Finally, after the experimental validation on public dataset, the forecasting error of the proposed model is reduced and the training speed is improved.