Traffic Flow Prediction via Spatial-Temporal Geometric Graph Convolutional Network

Traffic flow prediction is a critical task of building Intelligent Transportation System, especially the long-range traffic flow prediction plays an increasingly important role. Most of the current methods perform well in short-term traffic flow prediction, but they do not perform well in long-term prediction tasks. They have three main problems: 1) when extracting spatial features, the methods based on graph neural network lose the structural information of neighboring nodes and cannot establish long-term dependence among nodes; 2) when dealing with long-term complex forecasting tasks, they show high time complexity, which lead to the bottleneck of time feature utilization; 3) not paying too much attention to the complicated time patterns in time series. To solve the above problems, we propose a new spatio-temporal geometric graph convolutional neural network (STGGCN) for traffic prediction model, which uses geometric graph convolutional neural network and Autoformer to capture the spatio-temporal correlation of traffic data, establish long-term dependencies in both temporal and spatial features, and decomposes complex time patterns in sequence, which reduces the time complexity of the model while focusing on more time patterns. Experiments demonstrate that the model proposed in this paper achieves advantages over other six advanced methods on four datasets, especially in long-term forecasting tasks.