CCDSReFormer: Traffic flow prediction with a criss-crossed dual-stream enhanced rectified transformer model
Zhiqi Shao, Michael G H Bell, Ze Wang, Glenn Geers, Xusheng Yao, Junbin Gao
Abstract
Accurate, efficient, and rapid traffic forecasting is essential for intelligent transportation systems and plays a pivotal role in urban traffic planning, management, and control. While existing spatiotemporal transformer models have demonstrated effectiveness in traffic flow prediction, they face notable challenges in achieving a balance between computational efficiency and accuracy. Additionally, they often prioritize global trends over local time series information and treat spatial and temporal data separately, limiting their ability to capture complex spatiotemporal interactions. To overcome these limitations, we propose the criss-crossed dual-stream enhanced rectified transformer (CCDSReFormer). This model introduces a novel rectified linear self-attention (ReLSA) mechanism combined with enhanced convolution (EnCov) to reduce computational overhead and sharpen the local feature focus. Furthermore, our cross-learning strategy seamlessly integrates spatial and temporal data, improving the model's ability to capture intricate traffic dynamics. Extensive experiments on six real-world datasets show that CCDSReFormer outperforms existing models in both accuracy and efficiency. An ablation study further validates the contributions of each component, confirming the model's superior ability to forecast traffic flow accurately and efficiently.