A hybrid framework for spatio-temporal traffic flow prediction with multi-scale feature extraction

Efficient and accurate traffic flow prediction has become increasingly crucial with the advancement of intelligent transportation systems. This paper proposes a hybrid framework that combines depthwise separable convolutions and Transformer modules to learn spatio-temporal dependencies in traffic flow data. First, multi-scale features are extracted by depthwise separable convolutions, which decompose the convolution operation into independent spatial and temporal dimensions. This approach aims to reduce computational costs and effectively capture complex local spatiotemporal flow patterns in road networks. By adopting hierarchical processing, the model can learn dynamics across various scenarios and adapt to diverse traffic flow conditions. Then, we integrate a Transformer module into the model, leveraging its self-attention mechanism to capture the global patterns within traffic data. The integrated Transformer learns long-range dependencies across different road sections, which is particularly beneficial in road networks with complex interaction effects. Experiments on multiple real-world traffic datasets demonstrate that the proposed model outperforms traditional methods in both prediction accuracy and computational efficiency. The integration of depthwise separable convolutions and Transformer-based modeling exhibits superior performance in traffic flow prediction, providing a sufficient tool for urban traffic management.