Traffic Speed Prediction Based on Multi Spatial-Temporal Graph Fusion Convolution Network

Traffic speed prediction is an important and challenging task in urban planning. Although there are many existing traffic prediction methods, most of them use two separate modules to capture temporal and spatial correlation respectively, which will limit the effectiveness and efficiency of spatial-temporal feature learning, so they cannot get satisfactory prediction results. In order to capture spatial-temporal correlation at the same time with one module, a new model of Multi Spatial-Temporal Graph Fusion Convolution Networks (MST-FGCN) for traffic speed prediction is proposed. MST-FGCN module first use a data-driven method to generate a temporal graph to capture the temporal correlation and the historical data of the same time period and adjacent time periods are processed by fusing multiple different graph. Then, by integrating the graph convolution module and the gated multi-branch convolution module into a unified layer, the fused spatialtemporal fusion graph is processed. Finally, MST-FGCN can also learn long-term spatial-temporal correlations by stacking deeper layers. Experiments on three real public data sets show that our method is superior to the most advanced baseline.