[{"data":1,"prerenderedAt":225},["ShallowReactive",2],{"content-query-N0X8byPbbs":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"category":13,"body":14,"_type":219,"_id":220,"_source":221,"_file":222,"_stem":223,"_extension":224},"/technology-blogs/zh/2025","zh",false,"","MindSpore AI科学计算系列（22）：AI+海洋气象业界进展","在以下几节中，我们回顾了近年来深度学习模型在短临降水、气象预测、极端天气识别等方面的进展。","2022-12-09","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2023/01/03/e4296401fd8644d18d2f686747edc332.png","technology-blogs","大V博文",{"type":15,"children":16,"toc":216},"root",[17,25,31,36,40,51,56,68,73,81,86,96,101,106,116,126,131,136,141,146,151,156,161,171,176,186,191,196,201,206,211],{"type":18,"tag":19,"props":20,"children":22},"element","h1",{"id":21},"mindspore-ai科学计算系列22ai海洋气象业界进展",[23],{"type":24,"value":8},"text",{"type":18,"tag":26,"props":27,"children":28},"p",{},[29],{"type":24,"value":30},"背景：",{"type":18,"tag":26,"props":32,"children":33},{},[34],{"type":24,"value":35},"人类总是致力于预测和理解这个世界，从过去古希腊时代哲学推理到中世纪占卜的伪科学方法，到现代科学论述包括假设检验、理论发展和计算机建模，这些都是以统计和物理关系，即相关定律为基础的。在地球科学中，天气预报是一个成功的案例，主要依赖于理论与观测系统的完善、日益增长的计算能力。即便如此，对于天气的准确预报也只停留在天的尺度，尚且无法达到月尺度。目前地球系统数据激增，譬如广泛用于周期性气候评估的CMIP5数据集，模式输出的数据总量已超过3PB，而下一代CMIP-6数据集估计将达到30PB。与其他领域的大数据相同，地球科学领域海量的数据也具备四大特征：volume, velocity, variety and veracity（体积，速度，多样性和准确性），例如各种遥感、定点观测、模式数据。如何从这些大数据中提取与解读有效的信息，如何利用深度学习提升模型的预测能力，是地球科学系统如今面临的巨大挑战。",{"type":18,"tag":26,"props":37,"children":38},{},[39],{"type":24,"value":9},{"type":18,"tag":26,"props":41,"children":42},{},[43,49],{"type":18,"tag":44,"props":45,"children":46},"strong",{},[47],{"type":24,"value":48},"1.",{"type":24,"value":50}," E级深度学习模型实现气候数据像素级分割",{"type":18,"tag":26,"props":52,"children":53},{},[54],{"type":24,"value":55},"2018年的“戈登·贝尔”奖颁发给了来自美国的劳伦斯伯克利国家实验室通过深度学习识别极端天气，该成果运行在目前世界排名第一的Summit系统上，从而达到了史无前例的计算规模。这项工作的难点在于CAM5气候数据集分辨率高，要素多，且目标事件在样本中出现的概率比较低，正负样本不平衡。针对这些问题，该成果对图像分割模型Deeplabv3进行多项优化，包括系统级优化（数据流水线、数据规约优化等）以及模型级优化（加权损失函数、自适应学习率、梯度延迟更新等等），有效提升模型吞吐量与并行效率。",{"type":18,"tag":26,"props":57,"children":58},{},[59],{"type":18,"tag":60,"props":61,"children":65},"a",{"href":62,"rel":63},"https://link.zhihu.com/?target=https%3A//ieeexplore.ieee.org/abstract/document/8665799",[64],"nofollow",[66],{"type":24,"value":67},"Exascale deep learning for climate analytics",{"type":18,"tag":26,"props":69,"children":70},{},[71],{"type":24,"value":72},"如图1所示，该模型吞吐量最终达到325.8 PF/s，为10的15次方的量级，并行效率达90.7%。面向3.5T的气候数据，该模型可扩展至27360块GPU，吞吐量最高可以达到1.13 EF/s。",{"type":18,"tag":26,"props":74,"children":75},{},[76],{"type":18,"tag":77,"props":78,"children":80},"img",{"alt":7,"src":79},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2023/02/03/d59cc46f2b6b46ccb9d8e8abe0de8abc.png",[],{"type":18,"tag":26,"props":82,"children":83},{},[84],{"type":24,"value":85},"图1 Deeplabv3+模型吞吐量",{"type":18,"tag":26,"props":87,"children":88},{},[89,94],{"type":18,"tag":44,"props":90,"children":91},{},[92],{"type":24,"value":93},"2.",{"type":24,"value":95}," 基于深度学习的厄尔尼诺预测",{"type":18,"tag":26,"props":97,"children":98},{},[99],{"type":24,"value":100},"这是一项2019年发表在Nature上的成果：利用AI进行ENSO的预报。ENSO是发生在赤道东太平洋地区风场和海表面温度震荡的一个异常现象，具有2-7年的准周期，存在中性、暖性（正）、冷性（负）3个相位，数值模拟结果认为，太阳辐射的季节性变化、亚洲季风、全球变暖等因素与ENSO有关。将深度学习应用于气候预报的最大局限性之一是，对全球海洋温度分布的观测始于1871年，观察期太短，由于enso存在2-7年的周期，因此样本数量较少。",{"type":18,"tag":26,"props":102,"children":103},{},[104],{"type":24,"value":105},"这是韩国科学家和南京信息工程大学罗京佳教授利用CNN和迁移学习做的enso预测工作，这项工作的亮点是不仅使用了1871-1973的海表温度再分析数据来做预报，为了增加训练数据的个数，这项成果还利用了CMIP5的气候模拟数据。模型首先在CMIP5上进行训练，然后将训练结果作为初始值，用再分析数据进行训练，最终实现enso超前17个月相关系数超过0.5，超前6个月的相关系数超过0.7。",{"type":18,"tag":26,"props":107,"children":108},{},[109],{"type":18,"tag":60,"props":110,"children":113},{"href":111,"rel":112},"https://link.zhihu.com/?target=https%3A//www.nature.com/articles/s41586-019-1559-7",[64],[114],{"type":24,"value":115},"Deep learning for multi-year ENSO forecasts",{"type":18,"tag":26,"props":117,"children":118},{},[119,124],{"type":18,"tag":44,"props":120,"children":121},{},[122],{"type":24,"value":123},"3.",{"type":24,"value":125}," MindSpore助力降水预测",{"type":18,"tag":26,"props":127,"children":128},{},[129],{"type":24,"value":130},"这是华为在短临降水预测方面做的一项创新成果。针对目前AI外推短临预报技术长程预测衰减快、云团的移动受风的影响等难点，提出了自适应递归神经网络 （AdaRNN）。模型选取历史的雷达回波图像作为训练数据，进行训练，通过模型的调参优化，以及测试集的交叉检验，生成雷达回波最优预报模型，在实时业务中，以前20帧（前2小时）雷达回波作为输入数据，实现 0-2小时雷达反射率因子外推的业务化运行。",{"type":18,"tag":26,"props":132,"children":133},{},[134],{"type":24,"value":135},"针对云团生消以及预测图模糊等问题，模型进行了多项优化：",{"type":18,"tag":26,"props":137,"children":138},{},[139],{"type":24,"value":140},"1. 生消建模：时空模型加入了建模云团生消的模块；",{"type":18,"tag":26,"props":142,"children":143},{},[144],{"type":24,"value":145},"2. 基于特殊设计的对抗损失训练模型，能够生成细节清晰的雷达回波图；",{"type":18,"tag":26,"props":147,"children":148},{},[149],{"type":24,"value":150},"3. 使用神经网络实现纹理迁移，在保持区域特性的情况下，增加细节。",{"type":18,"tag":26,"props":152,"children":153},{},[154],{"type":24,"value":155},"4. 中期气象预报：英伟达FourCastNet & 华为Pangu-Weather",{"type":18,"tag":26,"props":157,"children":158},{},[159],{"type":24,"value":160},"今年春季，英伟达、劳伦斯伯克利国家实验室、密歇根大学安娜堡分校、莱斯大学等机构的研究者发布了一种基于傅里叶的神经网络预测模型 FourCastNet，该模型首次达到了能与NWP进行比较的精度，并将AI中期预报分辨率从之前的2°提升至0.25°，高精度网格输入对于捕捉细微结构至关重要，有助于提升降水预测精度。相较于传统数值模型，AI模型预测时间提升45000+倍。",{"type":18,"tag":26,"props":162,"children":163},{},[164],{"type":18,"tag":60,"props":165,"children":168},{"href":166,"rel":167},"https://link.zhihu.com/?target=https%3A//arxiv.org/abs/2202.11214",[64],[169],{"type":24,"value":170},"Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators",{"type":18,"tag":26,"props":172,"children":173},{},[174],{"type":24,"value":175},"前不久，华为对FourCastNet的工作进行全新升级，推出了Pangu-weather气象大模型。在FourCastNet基础上，盘古研究团队发现：AI气象预报模型的精度不足，主要有两个原因。第一，现有的AI气象预报模型都是基于2D神经网络，无法很好地处理不均匀的3D气象数据。第二，AI方法缺少数学物理机理约束，因此在迭代的过程中会不断积累迭代误差。为此，Pangu-weather提出了3D Earth-Specific Transformer（3DEST）来处理复杂的不均匀3D气象数据，并且使用层次化时域聚合策略来减少预报迭代次数，从而减少迭代误差。这项工作最终使用了192块GPU，模型训练时间长达两周。模型预报精度超越了FourCastNet，并且首次超过传统数值方法，终结了AI模型能否击败数值预报模式的争论。",{"type":18,"tag":26,"props":177,"children":178},{},[179],{"type":18,"tag":60,"props":180,"children":183},{"href":181,"rel":182},"https://link.zhihu.com/?target=https%3A//arxiv.org/abs/2211.02556",[64],[184],{"type":24,"value":185},"Pangu-Weather: A 3D High-Resolution Model for Fast and Accurate Global Weather Forecast",{"type":18,"tag":26,"props":187,"children":188},{},[189],{"type":24,"value":190},"参考文献：",{"type":18,"tag":26,"props":192,"children":193},{},[194],{"type":24,"value":195},"[1] Reichstein M, Camps-Valls G, Stevens B, et al. Deep learning and process understanding for data-driven Earth system science[J]. Nature, 2019, 566(7743): 195-204.",{"type":18,"tag":26,"props":197,"children":198},{},[199],{"type":24,"value":200},"[2] Kurth, Thorsten, et al. \"Exascale deep learning for climate analytics.\" SC18: International Conference for High Performance Computing, Networking, Storage and Analysis. IEEE, 2018.",{"type":18,"tag":26,"props":202,"children":203},{},[204],{"type":24,"value":205},"[3] Ham, Yoo-Geun, Jeong-Hwan Kim, and Jing-Jia Luo. \"Deep learning for multi-year ENSO forecasts.\" Nature 573.7775 (2019): 568-572.",{"type":18,"tag":26,"props":207,"children":208},{},[209],{"type":24,"value":210},"[4] Pathak J, Subramanian S, Harrington P, et al. Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators[J]. arXiv preprint arXiv:2202.11214, 2022.",{"type":18,"tag":26,"props":212,"children":213},{},[214],{"type":24,"value":215},"[5] Bi K, Xie L, Zhang H, et al. Pangu-Weather: A 3D High-Resolution Model for Fast and Accurate Global Weather Forecast[J]. arXiv preprint arXiv:2211.02556, 2022.",{"title":7,"searchDepth":217,"depth":217,"links":218},4,[],"markdown","content:technology-blogs:zh:2025.md","content","technology-blogs/zh/2025.md","technology-blogs/zh/2025","md",1776506118260]