Depth modeling and congestion prediction algorithm of large-scale spatio-temporal data of urban traffic flow

Abstract

Aiming at the challenges of limited static graph structure, insufficient multi-modal information fusion and poor robustness of long-term prediction in large-scale spatio-temporal data prediction of urban traffic flow, this paper proposes a congestion prediction framework based on Dynamic Spatio-Temporal Graph Network (DSTGN). The core contributions include: (1) designing an adaptive adjacency matrix learning module to dynamically capture the real-time correlation strength between nodes through attention mechanism, breaking through the traditional static graph assumption; (2) A cross-modal attention fusion mechanism is proposed, which uses BERT to encode the semantics of the event text, and realizes the deep alignment between the text and the traffic state through multi-head attention, thus enhancing the ability of emergency perception; (3) Construct a backbone network with the collaboration of space-time graph convolution and Transformer encoder, model local space-time dependence and global long-range correlation respectively, and adopt course learning strategy to realize multi-scale prediction from 15 minutes to 1 hour. Experiments on two real-world city datasets demonstrate that DSTGN achieves Mean Absolute Error (MAE) values of 4.78/5.64/6.85 for 15/30/60-minute forecasting tasks, respectively. This represents a 5%-12% improvement over advanced baselines such as Diffusion Convolutional Recurrent Networks (DCRNN), while exhibiting the smallest performance degradation over extended forecasting horizons. Ablation experiments verify the effectiveness of dynamic graph structure, cross-modal fusion and global coding module, and visual analysis shows that the dynamic association learned by the model conforms to the traffic flow propagation law. This study provides an effective scheme for high-precision and interpretable congestion prediction by fusing multi-source heterogeneous data.