Abstract:Based on forecast products of the European Centre for Medium-Range Weather Forecasts-Integrated Forecasting System (ECMWF-IFS) and hourly temperature observation data from the CMA Land Data Assimilation System (CLDAS), an enhanced model named ED-LSTM-FCNN is constructed, with an embedding layer module incorporated to handle high-dimensional spatial and temporal features. A fully connected neural network is utilized to integrate various feature types, achieve regression prediction of temperature, and generate gridded hourly temperature forecast products with a resolution of 0.05°×0.05°. Verification for the forecast products in Hunan Province in 2022 shows that this model exhibits a notable capacity to mitigate forecast errors inherent in the numerical model, and can enhance the overall forecast stability. The root mean square errors (RMSEs) of forecasts with lead time ranging from 1 to 24 hours exhibit a reduction of 25.4%-37.7% when compared to ECMWF-IFS and a decrease of 15.8%-40.0% relative to the National Meteorological Centre forecast (SCMOC). The model can significantly improve the forecast performance of ECMWF-IFS forecast, in spatial scale, particularly in regions characterized by intricate terrain. The RMSEs across most areas vary within the range of 1.2-1.6℃. The forecast accuracy of the model, with an error margin of ±2℃, surpasses 83.0% across various seasons, demonstrating a significant improvement compared to both ECMWF-IFS and SCMOC. The forecasting performance is notably superior, particularly in stable extreme high-temperature weather conditions compared to alternative products. In conclusion, this model has proved to be effective in the high-resolution temperature forecasting operations.