Based on the model prediction outputs from the European Centre for MediumRange Weather Forecasts (ECMWF), Japan Meteorological Agency (JMA) and National Centers for Environmental Prediction (NCEP), the multimodel ensemble forecasts of surface temperature are carried out by methods of the Multimodel Ensemble Mean (EMN), the Running Training Period Biasremoved Ensemble Mean (RBREM) and the Running Training Period Superensemble Forecast (RSUP), for Taizhou, Jiangsu Province. Their forecasts are compared with single model forecasts by means of Root Mean Square Error (RMSE) and Anomaly Correlation Coefficient (ACC). Thus, the multimodel ensemble forecast system is established. For the surface temperature forecast at 08:00 and 20:00, it is found from the comparative analysis of forecast results that the forecast skill of RBREM is apparently superior to those of individual models, EMN, and RSUP. Its average RMSE is reduced by about 0.5 ℃ with respect to the optimal single mode, and ACC increased by about 0.16. Additionally, the RBREM is applied into the daily operational forecast of Taizhou for the temperature. The forecast accuracy rate is improved efficiently due to the RBREM multimodel ensemble forecast system.