Based on the 3hour hourly data of air temperature from 99 stations in Shaanxi Province in spring and summer of 2016, the 3hour temperature forecast performance of the ECMWF highresolution model (0 to 72 hours) and the 6hour temperature forecast performance (78 to 240 hours) are evaluated. The results show that the accuracies of air temperature with the forecast error less than 1 ℃ and 2 ℃ are 30% to 50% and 55% to 85% before 72 hours, 10% to 30% and 25% to 55% after 96 hours at most observation stations of Shaanxi Province. The accuracy of air temperature forecasted starting at 00:00 GMT is higher than that forecasted starting at 12:00 GMT, and the accuracy in summer is higher than that in spring. With the increase of forecast lead time, the capability and stability of numerical forecasting of air temperature decrease wavily, and their diurnal variation is obvious. The capability and the stability from 15:00 GMT to 03:00 GMT next day are better than those from 06:00 GMT to 12:00 GMT. There was a high correlation between numerical forecast and observation, but elevation difference between model surface and observation site influences the quality of numerical temperature forecast greatly. To some degree, a method of terrain height revision can improve the numerical forecast accuracy of air temperature, which put the variation of air temperature in the vertical direction and the elevation difference between model surface and observation site into consideration. The method for correcting air temperature forecast has positive correction effectiveness for most areas in Shaanxi Province, but it is problematic and needs for further research.