Based on multiple physical quantity data issued by the T639 numerical forecast product, the localized temperature forecasting equations for counties and cities are established using multiple regression method, which are applied to forecast temperature in Zhenjiang of Jiangsu province.Combined with conventional meteorological observation data and the 6 h reanalysis data of United States NCEP/NCAR1 in 1°×1° grid point, the localized temperature forecast results are comparatively tested and analyzed.The results indicate that the proposed temperature forecast technique based on T639 is more accurate in minimum temperature prediction than in maximum temperature prediction, and more accurate in summer temperature forecast than in other seasons.Yangzhong and Danyang have the most accurate prediction result in maximum temperature and minimum temperature, respectively, compared with other counties or cities in Zhenjiang.Systemic negative deviation existed in objective forecast methods, which can be corrected through the use of seven day moving average of prediction error for corresponding forecast time-length.The correction results show that the prediction performance is improved, especially the maximum temperature prediction.The comparative analysis between subjective and objective prediction methods show that the objective one is weak in maximum temperature forecast, yet is of some value in temperature trend forecasting;the objective one is roughly similar in prediction accuracy for minimum temperature and thus can replace the subjective forecast.The objective prediction method in county level is valuable for forecast of disastrous weathers like high temperature and cold wave, yet is still weak in forecast performance compared with real-time weather data.