Abstract:Digital intelligent weather forecasting has grown rapidly in global meteorological operations in recent years, yet conventional methods face inherent limitations in nonlinear error correction, spatial and temporal downscaling, and multi-source data integration. Deep learning, with its exceptional nonlinear approximation and pattern recognition capabilities, has emerged as a transformative tool in digital intelligent forecasting workflows. This paper presents representative advances and achievements of deep learning in five key areas: numerical weather prediction (NWP) bias correction, high-resolution downscaling, heterogeneous multi-source data fusion, hazardous weather prediction (exemplified by typhoons), and datadriven weather forecasting. The results show that deep learning models can substantially enhance forecast accuracy by learning complex forecast-observation relationships end-to-end to correct systematic biases, employing generative adversarial networks to refine precipitation structures and improve heavy rainfall prediction skill, integrating radar, satellite, and model data to extend severe convective weather nowcasting lead times, and developing data-driven models that generate forecasts in seconds with accuracy comparable to numerical models. Notably, the Fenglei model demonstrates the superiority of generative approaches in convective nowcasting (achieving about 30% improvement in hit rate for radar reflectivity ≥50 dBz), while the Fengqing model has been operationally implemented for 15 d global forecasting. Deep learning integration has been advancing intelligent forecast products in spatial and temporal resolution and uncertainty quantification. However, its current applications have persistent challenges including limited training samples, model interpretability, extreme event prediction, cross-scale consistency, and computational efficiency, all of which require further investigation in future’s researches. The research priorities in the future should be put on the following aspects: expand large-sample and reforecast datasets; incorporate physical constraints to enhance model interpretability and robustness; implement tail-weighted loss functions to improve the reliability of extreme weather forecasts; design cross-scale coherent frameworks to ensure consistency across scales; and optimize training and inference efficiency for operational requirements. The synergistic integration of deep learning and numerical modeling, with their complementary strengths, represents a pivotal pathway for advancing intelligent numerical weather prediction.