By using ECMWF-IFS model data and observation data, this study established a fully connected neural network model (DL), tried to make more accurate temperature forecast based on deep learning method for the future 84 hours. The comparative evaluation of ECMWF-IFS and DL was made in 2020. The results show that DL correction can improve the ECMWF-IFS forecasts. The root mean square error (RMSE) of DL in most areas of Hunan is 1.5-2.0℃. There is a significant correction effect for ECMWF-IFS, and the improvement rate increases with the altitude of stations. The diurnal variation is also obvious for the RMSE of ECMWF-IFS/DL model. The RMSE is higher in the afternoon (with higher improvement rate), but lower before sunrise. The improvement rate has different diurnal variation characteristics at different altitudes. DL model shows higher accuracy in the whole year, of which the higher improvement rate appears in October and November (lower in December). In addition, the forecast results in a cold wave process was evaluated. In the accuracy for daily maximum/minimum temperature and the RMSE for 3 h temperature, DL model shows obvious correction ability for the systematic deviation from ECMWF-IFS. The DL temperature curve for single station is much closer to observation than the ECMWF-IFS forecast. Thus, the model can significantly reduce numerical weather prediction error， and its products could basically meet the demands of daily forecast service.