SSTCU:A Spatial-Temporal Correlation Unit based Traffic Flow Prediction Approach

Abstract

Traffic flow prediction is a crucial application in traffic guidance and control. Existing approaches rarely consider the dynamically correlated spatial-temporal features between multiple road segments. To effectively capture the spatial-temporal features between multiple road segments, we propose a novel approach, Spatial-Temporal Correlation Unit (STCU). STCU utilizes a fast Fourier transform-based autocorrelation mechanism to extract the correlations between temporal sequences, a graph attention mechanism to extract the correlations between spatial traffic monitors, and a feedforward neural network to fuse the interacting spatial-temporal correlations. We construct a traffic flow prediction model with a stacked STCU module called Sequential STCU (SSTCU). We conduct a lot of experiments and compared them with the several baselines to verify the effectiveness of SSTCU. The results show that the proposed method outperforms the baselines and achieves state-of-the-art performance. We also conduct ablation experiments to verify the effectiveness of the STCU module. Moreover, we change the layer depth of the model to find the most efficient setting for a computation efficiency consideration.