By using local surface temperature and high temporalspatial resolution numerical forecasting products of GFS (Global Forecast System) provided by NCEP,the preliminary forecast equations for daily average, highest and lowest temperature within 24 and 48 hours are established with the SPSS (Statistical Product and Service Solution) stepwise regression method, and the daytoday rolling forecasting was made with the Kalman filtering method. The results from June 2005 to May 2007 show that the average temperature forecast performed the best, the minimum the second, and the maximum the worst; the errors of forecasts increase with time; the significance of the factors differs in different seasons, smaller in autumn, higher in winter. The findings can be used as reference in forecasting operation.