Spatio-Temporal Graph Attention Convolution Network for Traffic Flow Forecasting

Abstract

Because of the complexity of the traffic network and the non-linearity of traffic data, it is extremely challenging to accurately predict long-term traffic flow. Spatio-temporal graph neural networks are currently the best paradigm for traffic flow forecasting, but most studies are still conducted on predefined graphs or graphs entirely generated by parameter training, which fail to extract the genuine spatio-temporal correlations in road networks. While the attention mechanism is effective in capturing global information, it may overlook local changes. We propose a novel spatio-temporal graph attention convolution network for traffic flow forecasting. In the time dimension, we combine temporal convolutions, which are good at capturing short-term features, with temporal attention, comprehensively considering short-term and long-term temporal correlations. In the spatial dimension, we utilize graph convolutions of fused multiple graphs to thoroughly extract the hidden information and local changes in road networks. By integrating this with spatial attention, we fully consider both local and global spatial correlations. Experimental results on real-world datasets demonstrate the effectiveness of our approach.