LSTM-Based Road Surface Temperature Forecasting for Winter Road Icing Risk Management

Road surface icing is a major cause of skid-related winter accidents, making preemptive maintenance essential. Current decisions rely mainly on regional weather information, which cannot capture how road surface temperature (RST) varies between sections according to road structure and topography. This study had two objectives: to empirically verify that RST, rather than air temperature, dominantly governs icing, and to develop a deep learning model forecasting future RST. A three-stage approach was adopted. First, a laboratory experiment observed icing conditions on asphalt and concrete specimens. Second, RST and meteorological data were collected over one winter at four sites in Seoul, South Korea—two earthwork sections, a bridge pavement, and a mountainous road. Third, a long short-term memory (LSTM) model was trained for each site to forecast RST at 10, 30, 60, and 120 min ahead, with a case study comparing input variables and window sizes. In the laboratory, icing began only when RST fell to between 0 °C and −1 °C, although air temperature had already reached −16.5 °C. In the field, RST was consistently lower at the bridge and mountainous sites than at the earthwork sections. The best input combination, RST and air temperature, achieved an average RMSE of 0.47 °C and an R2 of 0.95; excluding RST reduced R2 to approximately 0.30, whereas window size had little effect. RMSE increased by approximately 266% from 10 to 120 min ahead, and the model is considered suitable for short-term forecasting within one hour.