Dual Flow Fusion Graph Convolutional Network for Traffic Flow Prediction

In recent decades, motor vehicle ownership has increased worldwide year by year, which causes that the accurate prediction of traffic flow on urban road networks becomes more important. However, the dual dependence on the micro layer and the macro layer creates a huge challenge for the prediction task. Previous models lack comprehensive analysis of the macro features at different time granularities. In this paper, we propose a novel Dual Flow Fusion Graph Convolutional Network (DFFGCN) to solve this problem. For capturing more macro features, we build the interactions between the micro layer and the macro layer at more time granularities. Then the spatial-temporal normalization model is introduced to separate the temporal and spatial influences. Therefore, the proposed DFFGCN has a better learning ability compared with other advanced models. Finally, we give experiments to show the effectiveness and superiority of our proposed model. Experimental results on three traffic datasets demonstrate that DFFGCN can achieve state-of-the-art performance consistently. And the ablation studies confirm the importance of each element of DFFGCN.