A theory-guided multi-scale temporal fusion network for road surface temperature nowcasting

The durability of road infrastructures strongly correlates with thermal variations. This study proposed a theory-guided multi-scale temporal fusion network (MSTF-Net), integrating multi-scale temporal feature pyramids, dual-pooling mechanisms, and real-time spatiotemporal attention mechanisms for road surface temperature (RST) nowcasting. Through comprehensive evaluations, MSTF-Net achieved 0.66°C mean absolute error with 0.01 ± 0.01°C uncertainty and 74% and 92% threat score for hot and cold extremes respectively, significantly outperforming Multiple linear regression (Linear), Random Forest (RF), eXtreme Gradient Boosting (XGB) and Long Short-term Memory (LSTM) (p < 0.05) in RST prediction accuracy, spatial generalization, weather adaptability and extreme event detections. Systematic ablation studies revealed a feature extraction hierarchy of maximum pooling in micro branches>average pooling in micro branches>real-time spatiotemporal attention mechanisms>macro branches. MSTF-Net combined lightweight computational characteristics (0.11 MB) with high-resolution modeling capabilities, enabling efficient real-time emergency response optimization (38.46 FPS), potentially providing new capabilities in infrastructure-resilient smart systems for thermal risk identification and road predictive maintenance.