Hybrid Road Surface Temperature Prediction Using Integrated Road Weather Observations

This study presents a hybrid road surface temperature (RST) prediction framework that integrates fixed Road Weather Information System observations, vehicle-based mobile road weather measurements, and GIS-derived terrain and structural information along a 330-km section of the Seohaean Expressway, and systematically evaluates its performance using a leakage-aware validation design. A background model (M1), formulated using harmonic regression to represent the diurnal cycle and solar zenith angle, is linearly combined with a persistence forecast to construct the Base model, which achieves modest yet consistent reductions in RMSE of approximately 0.6-1.0% relative to persistence across lead times of H = 300, 600, and 1,800 s. A residual correction model (M2), implemented using GRU-TCN architectures with mobile RST and temporal and terrain/structural predictors, further reduces mean prediction errors at 300 s but leads to increased residual variance and larger extreme errors at 1,800 s when compared with the Base model. Regime-wise analyses across bridges and tunnels, sky view factor (SVF) classes, and coastal versus inland segments reveal substantially larger errors in tunnel environments and low-SVF sections. The proposed evaluation protocol, which incorporates run-wise splitting, purge windows, and the explicit exclusion of future-dependent features, provides a practical and transferable reference for mitigating information leakage in spatiotemporal RST prediction.