RM-DSTGN: Dynamic Graph Topology Evolution via Rainfall-Modulated Attention for Spatiotemporal Traffic State Estimation

Abstract

Accurate traffic speed prediction in Intelligent Transportation Systems (ITS) is challenging under adverse weather due to highly non-linear traffic dynamics. Traditional deep learning relies on static graphs and treats weather as a peripheral feature, failing to capture rainfall-induced structural shifts in traffic propagation. To address this, we propose the Rainfall-Modulated Dynamic Spatiotemporal Graph Network (RM-DSTGN). Unlike conventional models, RM-DSTGN internalizes rainfall intensity as a global structural controller that dictates spatial connectivity and temporal reliance. The framework integrates a rainfall-modulated graph generator to dynamically reconfigure spatial dependencies, a dual-branch stability module decoupling macroscopic trends from microscopic speed variances, and a gating mechanism calibrating temporal reliance based on rain severity. Evaluated on real-world data from a critical Hong Kong arterial intersection, RM-DSTGN significantly outperforms state-of-the-art benchmarks, achieving a mean absolute percentage error (MAPE) of 3.62%. Comprehensive ablation and error analyses confirm the model's robustness across varying rainfall intensities, prediction horizons, and look-back windows. Ultimately, RM-DSTGN maintains consistent service levels during extreme weather, validating its potential to enhance urban mobility and safety.