Graph Wavelet Neural Controlled Differential Equations Method for Speed Prediction Under Traffic Accidents

Abstract

Accurate speed prediction is a crucial component of intelligent transportation systems, as it enhances traffic management and operational efficiency. While the majority of existing research concentrates on speed prediction under normal traffic conditions, the occurrence of traffic accidents significantly disrupts typical urban traffic patterns, leading to reduced predictive accuracy. Considering that the disruption caused by accidents is localized and severe, and that the dynamic behavior of traffic flow can be effectively modeled through differential equations, we propose a novel traffic speed prediction model, graph wavelet neural controlled differential equations (GW-NCDE). The GW-NCDE model leverages graph wavelet transforms to effectively capture the spatial characteristics of the road network under accident conditions and employs a dual-layer neural controlled differential equation structure for enhanced predictive performance. Experiments conducted on a real-world dataset from Wangjing, Beijing, demonstrate that our model outperforms several existing benchmark methods. Particularly in accident scenarios, compared to the best-performing benchmark, the short-term prediction error of our model is reduced by more than 10%. These results underscore the model's robustness and superior predictive capability in complex and dynamic urban traffic environments.