Traffic flow prediction based on spatial-temporal multi factor fusion graph convolutional networks

Recently, graph convolutional networks (GCNs) have become one of the important models for solving traffic flow prediction, but existing models still have two problems: (1) insufficient information utilization: there is a lack of adequate consideration of the relevant characteristic information between the original data flows. Ignoring external weather information. Only considering the interaction influence of limited surrounding nodes on the target node fails to effectively characterize the joint spatial-temporal correlation; (2) receptive field limitation: the existing graph convolution network model may cause the network node features to be too smooth and lose the original information when analyzing the spatial features extracted by the filter used in the traffic flow data. To address the above issues, we proposed a spatial-temporal multi factor fusion graph convolution network (STFGCN), which is composed of multi factor graph fusion module, the GCN based on the auto-regressive moving average (ARMA) filter and the gated recurrent unit (GRU). Specifically, we consider the correlation between historical data, the joint spatial-temporal correlation between nodes, and external weather factors. The GCN based on the ARMA filter is used to extract the spatial features, and the GRU is utilized to capture temporal features from traffic flow data. Experimental results on four public real-world datasets prove the superiority of our model in terms of prediction performance and capturing the dynamic spatial-temporal correlation.