Abstract: Based on SCMOC refined grid forecast products and observation data of traffic meteorological stations,using the Fully Connected Neural Network model (FCNN),the hourly road surface temperature was forecasted for the next 24 hours in winter along the complex section of the G6 Beijing-Xizang Highway in Inner Mongolia.The results show that according to the experiment,the optimal training period of FCNN model is 20 d.Increasing the time characteristic variable could reduce the forecasting error of the road surface temperature effectively,MAE decreases by 0.3~0.8 ℃,AR increases by 4%~18%.The final FCNN model of road surface temperature forecast is developed with high accuracy,in which ME,MAE and RMSE are -0.2~0.0 ℃,1.1~1.3 ℃,and 1.6~1.9 ℃,respectively,r exceeds 0.94,AR is above 90%,and better performance for temperature below 0 ℃.From the perspective of daily variation of each test index,except for the poor forecast performance from 12:00 to 15:00,the other time periods have good forecast performance,with the mean values of ME,MAE,RMSE,r,and AR are -0.07 ℃,1.20 ℃,1.58 ℃,0.94,and 93%,respectively.Case studies conducted for stable weather,cold wave weather,and snowfall weather during the winter of 2022 confirm that the FCNN model maintains a forecasting MAE within 2.0 ℃ across different weather scenarios,demonstrating its applicability for road surface temperature prediction under diverse meteorological conditions.