Spatio-Temporal Attention Hybrid Network for Short-Term Origin-Destination Flow Prediction in Urban Subways

In the context of rapid urbanization and socio-economic development, the subway system has positioned itself as pivotal in meeting the growing demand for transportation due to its efficiency. Given the challenge of managing escalating passenger flow pressure, the ability to accurately predict passenger volumes is crucial. Such predictions serve as a vital strategy to guarantee passenger safety, and also to mitigate congestion during peak periods. Specifically, precise prediction of origin-destination (OD) passenger traffic is elevating for improving the quality and efficiency of subway services, but the challenges posed by the high dimensionality and complex spatio-temporal dependencies of OD matrix sequences. To this end, we proposed an innovative Spatio-Temporal Attention Hybrid Network (STAHN) model aimed at addressing the issue of short-term OD passenger traffic prediction in urban subway systems. The model includes a spatio-temporal (ST) block and a fusion block, where the ST block integrates local and global spatial features and employs attention mechanisms to extract temporal features, effectively capturing the spatio-temporal dependencies of the subway system. The fusion block utilizes the extracted OD features and DO features, enhancing predictive performance through DO feature-assisted OD prediction. Test results on the actual operational dataset of the Shenzhen subway show that the model proposed in this study demonstrates a significant advantage in OD prediction accuracy compared to other advanced prediction methods. Furthermore, the effectiveness of each component of the model was further confirmed through ablation studies.