国内AI中期天气预报大模型进展及面临的挑战

Progress and Challenges of Large-Scale AI Mid-Term Weather Forecasting Models in China

摘要: 随着人工智能技术的快速发展，AI在天气预报领域的应用已成为提升预报准确率的重要手段。本文深入介绍了国内AI中期预报大模型的最新研究成果与创新价值，具体聚焦于风乌气象大模型、伏羲气象大模型和华为盘古气象大模型。风乌模型通过多模态多任务学习方法、不确定性损失函数及重放缓冲机制，显著提升了预报的准确性和时效性，特别是在全球500 hPa位势高度和台风路径预报方面展现出卓越性能。伏羲模型则采用级联机器学习架构和纬度加权损失函数，优化了长期天气预报，尤其在对小到中等降水事件的预报上表现出色。华为盘古气象大模型凭借三维地球特定变换器和分层时间聚合算法，实现了高空气象变量和地表气象变量的高精度预测，同时在极端天气事件的预报上展现出强大能力。这些模型不仅提高了预报精度和效率，还减少了对高性能计算资源的依赖。本文还探讨了AI气象大模型与数值预报的互补关系，并对未来AI在气象预报领域的发展趋势进行了展望，强调了实时数据同化和极端天气模型构建的重要性。尽管AI气象大模型在常规气象预报方面取得了显著成果，但在极端天气事件预报、数据同化、高分辨率数据处理等方面仍面临挑战。这些研究成果和创新价值为气象预报领域的研究和实践提供了有价值的参考和启示。

Abstract: With the rapid development of artificial intelligence technology, AI has become an important means to improve the accuracy of weather forecasting. This article provides an in-depth introduction to the latest research achievements and innovative value of domestic AI mid-term forecasting models, focusing on the Wind Wu Meteorological Model, the Fuxi Meteorological Model, and Huawei’s Pangu Meteorological Model. The Wind Wu model, through multi-modal multi-task learning methods, uncertainty loss functions, and replay buffer mechanisms, has significantly improved the accuracy and timeliness of forecasts, especially in global 500 hPa geopotential height and typhoon path forecasting. The Fuxi model adopts a cascaded machine learning architecture and latitude-weighted loss functions to optimize long-term weather forecasts, particularly excelling in the prediction of small to moderate precipitation events. Huawei’s Pangu Meteorological Model, with its 3D Earth-Specific Transformer and layered temporal aggregation algorithm, has achieved high-precision forecasting of both upper-air meteorological variables and surface meteorological variables, while also demonstrating strong capabilities in extreme weather event forecasting. These models have not only improved forecasting accuracy and efficiency but also reduced the dependence on high-performance computing resources. The article also discusses the complementary relationship between AI meteorological models and numerical forecasting and looks forward to the future development trends of AI in the field of meteorological forecasting, emphasizing the importance of real-time data assimilation and the construction of extreme weather models. Although AI meteorological models have achieved significant results in routine meteorological forecasting, they still face challenges in extreme weather event forecasting, data assimilation, and high-resolution data processing. These research achievements and innovative values provide valuable references and insights for research and practice in the field of meteorological forecasting.