Abstract: To enhance the ice-period water conveyance capacity of the main canal in the Middle Route of South-to-North Water Diversion Project and maximize water supply under safe ice-period operation, accurate ice prediction is crucial. Utilizing prototype ice observation data since the project′s operation, a nonlinear regression model was developed using the Radial Basis Function Neural Network (RBFNN) to predict the next day′s average water temperature based on the current day′s average water temperature, the next day′s highest and lowest air temperatures, weather conditions, and flow velocity. Meanwhile, a two-category classification model for ice condition states was established using the Random Forest (RF) model based on water temperature, air temperature, and flow velocity factors. Integrating both, a daily ice prediction model based on RBFNN-RF was formed. Applied to the typical North Juma River section, the model was tested for lead times of 1, 3, 5, 7 days. Results showed that the root mean square errors of predicted water temperatures for these lead times were 0.17 ℃, 0.36 ℃, 0.52 ℃, 0.64 ℃, respectively, with corresponding ice condition prediction accuracy rate of 95.28%, 92.68%, 89.08%, 85.22%. The established model demonstrates high accuracy and can provide technical support for precise dynamic regulation and benefit maximization of the project during ice periods, offering a valuable reference for ice-period operation of similar long-distance open-channel water diversion projects.