Deep learning based correction of ECMWF forecast products with fusion of DEM and FY-4A data

Accurate numerical weather prediction is an important prerequisite for refined public and commercial meteorological services. ECMWF forecast products are widely used around the world, where as systematic forecast errors always exist. As a correction of numerical prediction products, multi-source data fusion can effectively reduce prediction errors, which is also a typical high-dimensional nonlinear mapping problem. Due to the heterogeneity of geographic data and ground truth data and satellite data, it is necessary to establish a mechanism to fully extract and utilize effective information from these data while avoid noise and redundancy of the information. In recent years, deep learning methods have been extensively applied to data post-processing in meteorological field. Aiming at errors in numerical weather prediction and the nonlinear mapping problem in multi-source data fusion, this study designs a correction deep learning model NFC-Net for ECMWF numerical prediction products, which mainly includes a multi-source data spatial resolution alignment module, a spatiotemporal feature extraction module, and a UNet correction module. NFC-Net optimizes and corrects the forecast results by integrating multi-source data such as FY-4A satellite data, DEM, and ERA5 historical truth data, and utilizes multi-source data spatial resolution alignment module and spatiotemporal feature extraction module to achieve feature extraction and fusion for multi-source heterogeneous data. At the same time, this paper also proposes a spatial resolution alignment algorithm based on convolutional neural networks (UPS-MSR algorithm) and a dual self-attention mechanism (DSA). The UPS-MSR algorithm uses up-sampling and multi-scale residual networks to achieve grid alignment of meteorological and geographic data with different resolutions, which can effectively avoid information loss. The DSAConvlstm network embedded in DSA module can balance the spatiotemporal correlation and element correlation when extracting features from high-dimensional meteorological information. To evaluate the performance of the proposed method NFC-Net, correction experiments on two weather elements, i.e., 2 m temperature and 10 m wind speed in the ECMWF products, are carried out and the results are compared with the ECMWF forecast results, ANO, Convlstm, Fuse-CUnet and ERA5. The experiments show that the root mean square errors (RMSEs) of 2 m temperature and 10 m wind speed corrected by the NFC-Net model decrease by 49.71% and 50.86%, respectively when compared to that in the ECMWF forecast products. The experimental results indicate that the introduction of high-resolution DEM data in the NFC-Net model can obviously optimize land surface process of the model, and the correction effect is more pronounced under complex terrain condition. The use of FY-4A satellite data enables the model to obtain more three-dimensional information during correction. The application of DSA module can make the model pay more attention to variables that have strong correlations with correction elements, and thereby significantly improves the quality of correction. The proposed method can prospectively be applied in the correction of ECMWF forecast results and promote the accuracy of numerical weather prediction.