Deep Learning Methods for Multi-Element Bias Correction in Numerical Weather Forecasts

Modern weather forecast operations rely on numerical model predictions, and improving the accuracy of model forecasts is crucial for enhancing refined intelligent gridded forecasts. In this study, based on three deep learning models, namely Coupled U-Nets (CU-Net), Attention U-Net (Att-Unet) and Swin-Unet, the integration of meteorological factors and terrain data for modelling is conducted in the Huang-Huai River Basin and its surrounding areas by using the European Centre for Medium-Range Weather Forecasts multi-element forecast data from March 2017 to April 2022. The results from the deep learning-based methods are analyzed and compared with those from the traditional Anomaly Numerical-correction with Observations (ANO) method. The results indicate that the deep learning methods exhibit more remarkable positive correction effects than the ANO method. Specifically, the Swin-Unet model performs the best in correcting 2-m temperature (T2m) forecasts and 2-m specific humidity forecasts, while the Att-Unet model is optimal for correcting 10-m zonal wind, 10-m meridional wind and surface pressure forecasts. The credibility curves of the corrected results indicate that the correction benefits of various deep learning methods tend to be related to the different ranges of corrected elements. For example, the Att-Unet and Swin-Unet models perform the best when T2m < 0°C and T2m > 25°C, respectively. However, in terms of the correction for 10-m wind speed forecasts, the advantages of deep learning methods are outstanding when wind speed is less than 8 m s–1. The CU-Net model excels in high-humidity regions. Moreover, the Swin-Unet model, which is based on a U-shaped encoder-decoder with a pure transformer, demonstrates a stronger ability to portray the details in individual correction cases, such as characterizing high-temperature regions, extreme humidity areas and local wind speed variations in the Yangtze River Basin.