[{"data":1,"prerenderedAt":204},["ShallowReactive",2],{"content-query-z4APQoyC6t":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":198,"_id":199,"_source":200,"_file":201,"_stem":202,"_extension":203},"/technology-blogs/zh/2969","zh",false,"","MindSpore AI科学计算系列 | MetNet3融合稀疏站点数据，实现提前24h天气预报","作者：于璠 来源：知乎","2024-01-30","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/02/02/7c32e72474624361af96357632e6301c.png","technology-blogs",{"type":14,"children":15,"toc":195},"root",[16,24,43,51,56,61,71,76,84,89,94,99,106,111,116,124,129,134,144,153,164,174,185],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"mindspore-ai科学计算系列-metnet3融合稀疏站点数据实现提前24h天气预报",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28,30,36,38],{"type":23,"value":29},"**作者：**",{"type":17,"tag":31,"props":32,"children":33},"strong",{},[34],{"type":23,"value":35},"于璠",{"type":23,"value":37}," ",{"type":17,"tag":31,"props":39,"children":40},{},[41],{"type":23,"value":42},"来源：知乎",{"type":17,"tag":25,"props":44,"children":45},{},[46],{"type":17,"tag":31,"props":47,"children":48},{},[49],{"type":23,"value":50},"背景",{"type":17,"tag":25,"props":52,"children":53},{},[54],{"type":23,"value":55},"准确预测未来几分钟到几周的天气是一项基本的科学挑战。目前，绝大部分气象机构采用传统数值模式进行天气预报，这些数值模式是基于大气的物理模型进行构建的，但同时这些模式本身也会受到计算成本的限制。深度神经网络为天气变化建模提供了另一种范式，一旦模型训练完成，神经网络可以在不到一秒的时间内做出预报，同时兼备非常高的时间分辨率和空间分辨率。与此同时，AI模型还可以直接从大气观测数据中进行训练，以此获得更贴近观测的预测效果。",{"type":17,"tag":25,"props":57,"children":58},{},[59],{"type":23,"value":60},"2023年6月，谷歌研究院与DeepMind合作发布了最新的天气模型MetNet-3，与之前的MetNet和MetNet-2相比，MetNet-3的预报时效从8小时、12小时提升到了24小时，并且增加了预报变量，同时扩大了预报范围。",{"type":17,"tag":25,"props":62,"children":63},{},[64,66],{"type":23,"value":65},"**1、**",{"type":17,"tag":31,"props":67,"children":68},{},[69],{"type":23,"value":70},"方法与实验",{"type":17,"tag":25,"props":72,"children":73},{},[74],{"type":23,"value":75},"在近期大量涌现的AI气象预报模型中，如Graphcast、盘古气象大模型，均采用再分析数据作为训练样本，即通过背景场与观测同化后的数据。相较于这些中期预报模型，MetNet-3选择引入一部分大气观测数据来进行模型的训练和评估。然而观测数据与再分析数据不同的是，观测数据来源多样且异构，数据存在不同程度的稀疏性。为了融合大气观测数据（稀疏点状数据），MetNet-3采取了一种Densification方式，这种方法将数据同化与模拟预测直接在AI模型中进行融合，Densification的过程如图1所示。",{"type":17,"tag":25,"props":77,"children":78},{},[79],{"type":17,"tag":80,"props":81,"children":83},"img",{"alt":7,"src":82},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/02/02/84e19af92a0941e790adec0f56a2704e.png",[],{"type":17,"tag":25,"props":85,"children":86},{},[87],{"type":23,"value":88},"图1 Densification训练方式",{"type":17,"tag":25,"props":90,"children":91},{},[92],{"type":23,"value":93},"Densification的目的是利用站点观测数据，预测所有位置上的基础气象要素。首先将所有站点取80%作为训练站点，20%作为验证站点，这20%的站点是不参与训练的，只参与精度验证；在实际训练中，随机mask25%的站点，然后输出所有训练站点的气象要素，并计算loss；然后在模型验证阶段，会把所有的训练站点都输进去，在所有验证站点上验证精度。最后在模型推理阶段，模型将所有的站点都作为输入，输入全部位置的预测值。",{"type":17,"tag":25,"props":95,"children":96},{},[97],{"type":23,"value":98},"MetNet-3整体网络架构如图2所示，MetNet-3选择Unet作为backbone，由于数据集部分包含了两种分辨率，因此在Unet输入中，首先输入4km分辨率的数据，并在一次DownSample后加入8km分辨率数据，实现两种分辨率数据的融合。其次，MetNet-3在Unet中加入了MaxVit模块，主要考虑利用grid attention来捕捉一些全局信息。模型最终有三个输出，在4km的分辨率下会输出600多个同化的基础要素和稀疏的站点数据。在模型的最终输出层，数据会UpSample到1km的降水数据上。值得注意的是，模型虽然会输出同化结果作为loss用来约束整个模型，但是模型的评估里并不会考虑这些要素的预测质量。",{"type":17,"tag":25,"props":100,"children":101},{},[102],{"type":17,"tag":80,"props":103,"children":105},{"alt":7,"src":104},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/02/02/edf2a0a15895417f9da97a26aba75a5c.png",[],{"type":17,"tag":25,"props":107,"children":108},{},[109],{"type":23,"value":110},"图2 MetNet-3模型架构",{"type":17,"tag":25,"props":112,"children":113},{},[114],{"type":23,"value":115},"另外，实验也发现如果同时训练这些变量，降水的预报质量和其他地面要素的预报质量会存在trade off，因此MetNet-3采取的训练策略是先减少OMO的损失权重训练降水要素，然后在把权重扩大100倍来进行模型微调。",{"type":17,"tag":25,"props":117,"children":118},{},[119],{"type":17,"tag":31,"props":120,"children":121},{},[122],{"type":23,"value":123},"2、结论与感谢",{"type":17,"tag":25,"props":125,"children":126},{},[127],{"type":23,"value":128},"MetNet-3是一种新型的AI天气预报模型，可以对核心天气变量进行提前24小时的预报，其性能优于最先进的数值模式。目前MetNet-3已投入运行，可生成实时12小时降水预报，这些预报现已在与天气相关的Google产品中提供服务。",{"type":17,"tag":25,"props":130,"children":131},{},[132],{"type":23,"value":133},"往期回顾",{"type":17,"tag":18,"props":135,"children":136},{"id":7},[137],{"type":17,"tag":138,"props":139,"children":143},"a",{"href":140,"rel":141},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247613185&idx=1&sn=e3a032720426aab204582e244d318953&chksm=c11e2b4ef669a2585c1ed10ef619133359b4008b3a227b5c089e58131b810bdc9dd1bafd814b&scene=21#wechat_redirect",[142],"nofollow",[],{"type":17,"tag":18,"props":145,"children":147},{"id":146},"_1",[148],{"type":17,"tag":138,"props":149,"children":152},{"href":150,"rel":151},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247614352&idx=1&sn=6eddd5488123a7bfbb85681317d6b950&chksm=c11e2fdff669a6c95c412f2e5711d34310f56cd9d468bd7a5c0edecc6cfd6cd55ecefee097a0&scene=21#wechat_redirect",[142],[],{"type":17,"tag":18,"props":154,"children":156},{"id":155},"mindspore-ai科学计算系列-化学深度学习模型chemgpt的性能评估公式拟合",[157],{"type":17,"tag":138,"props":158,"children":161},{"href":159,"rel":160},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247614481&idx=1&sn=e45cb91a0dc5f51db7e80761d639a474&chksm=c11e105ef66999488545e230a32ef8c771768c8b488024d5b039bba9ed12974339f795cbe42e&scene=21#wechat_redirect",[142],[162],{"type":23,"value":163},"MindSpore AI科学计算系列 | 化学深度学习模型ChemGPT的性能评估公式拟合",{"type":17,"tag":18,"props":165,"children":167},{"id":166},"mindspore-ai科学计算系列-gnn-mom基于昇思mindspore-elec的图残差学习电磁求解器",[168],{"type":17,"tag":138,"props":169,"children":171},{"href":150,"rel":170},[142],[172],{"type":23,"value":173},"MindSpore AI科学计算系列 | GNN-MoM基于昇思MindSpore Elec的图残差学习电磁求解器",{"type":17,"tag":18,"props":175,"children":177},{"id":176},"mindspore-ai科学计算系列-vae基于mindspore-elec的特征域mt反演提升大地电磁反演的精度和分辨率",[178],{"type":17,"tag":138,"props":179,"children":182},{"href":180,"rel":181},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247613627&idx=1&sn=ad3dd2c6f59f0626a93afb93c736123e&chksm=c11e2cf4f669a5e22f5c7f33c3088e3efa800f13cb1cfac88f583df387787f3087f941167dd3&scene=21#wechat_redirect",[142],[183],{"type":23,"value":184},"MindSpore AI科学计算系列 | VAE基于MindSpore Elec的特征域MT反演，提升大地电磁反演的精度和分辨率",{"type":17,"tag":18,"props":186,"children":188},{"id":187},"mindspore-ai科学计算系列-allegro提升昇思分子势能预测准确性将上线mindspore-chemistry",[189],{"type":17,"tag":138,"props":190,"children":192},{"href":140,"rel":191},[142],[193],{"type":23,"value":194},"MindSpore AI科学计算系列 | Allegro提升昇思分子势能预测准确性，将上线MindSpore Chemistry",{"title":7,"searchDepth":196,"depth":196,"links":197},4,[],"markdown","content:technology-blogs:zh:2969.md","content","technology-blogs/zh/2969.md","technology-blogs/zh/2969","md",1776506124856]