Abstract
Predictions of daily maximum and minimum temperatures (Tmax and Tmin) are key components of operational weather forecasting. Here we show how a deep learning scheme can be used to improve their predictions based on the numerical weather prediction (NWP) output from the European Centre for Medium-Range Weather Forecasts − Integrated Forecasting System (ECMWF-IFS). Using an optimal factor set screened by a regression method, an error-correction model for Tmax and Tmin forecasts based on the Spatio-Temporal Stacked Residual Network (STS-ResNet) is established. We find that errors in Tmax and Tmin forecasts for Hunan Province, China, can be reduced by approximately 21% and 33% respectively. However, although the Tmax and Tmin forecasts at almost all terrain elevations have been improved, the improvement decreases with the increasing terrain elevation. To solve this problem, we designed the Residual and Spatial Attention STS-ResNet (SASTS-ResNet) based on spatial attention mechanism. In mountainous regions, the SASTS-ResNet makes up for the deficiency of STS-ResNet in improving the Tmax forecasts of the ECMWF-IFS (with the improvement increasing from 1.45 to 42.53%), which has also largely improved the Tmin forecasts (from 27 to 83%). Moreover, the ECMWF-IFS model, STS-ResNet and SASTS-ResNet all have some uncertainties in Tmax and Tmin in high-elevation areas, where the smallest uncertainty is found in the SASTS-ResNet model.