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

Area of Science:

• Civil Engineering
• Artificial Intelligence
• Environmental Science

Background:

• Road infrastructure durability is significantly impacted by thermal variations.
• Accurate road surface temperature (RST) prediction is crucial for infrastructure maintenance and safety.

Purpose of the Study:

• To develop a novel deep learning model for precise real-time road surface temperature (RST) nowcasting.
• To enhance the prediction accuracy and spatial generalization of RST forecasting models.
• To enable efficient real-time emergency response and predictive maintenance for road infrastructure.

Main Methods:

• Proposed a theory-guided Multi-Scale Temporal Fusion Network (MSTF-Net) integrating multi-scale temporal feature pyramids, dual-pooling, and spatiotemporal attention.
• Evaluated MSTF-Net against traditional machine learning (Linear, RF, XGB) and deep learning (LSTM) models.
• Conducted systematic ablation studies to determine the optimal feature extraction hierarchy within MSTF-Net.

Main Results:

• MSTF-Net achieved a mean absolute error of 0.66°C with low uncertainty (0.01 ± 0.01°C).
• Demonstrated superior performance (p < 0.05) over benchmark models in RST prediction accuracy, spatial generalization, and extreme event detection (74% hot, 92% cold threat score).
• Achieved efficient real-time processing (38.46 FPS) with a lightweight model (0.11 MB).

Conclusions:

• MSTF-Net offers a significant advancement in road surface temperature nowcasting, outperforming existing methods.
• The model's efficiency and accuracy support infrastructure-resilient smart systems for thermal risk management.
• This approach provides new capabilities for road predictive maintenance and real-time emergency response optimization.