气象人工智能预报模型上新升级——五大模型分工协作赋能千行百业 □ 本报记者 郭静原 在全球变暖背景下，气候形势日趋复杂多变，高温、暴雨、强对流等灾害的突发性与破坏性显著增 强，既对气象预报预警的精准度、时效性提出更高要求，也推动气象服务加速向智能化转型。日前，中 国气象局发布新的气象人工智能科学模型"风源"，并实现"风清""风雷""风顺"3款气象人工智能预报模 型的同步升级。加上前不久刚刚发布面向气象服务领域的千亿参数语言模型"风和"，5大模型分工协 作，正以更贴心的气象预报预警能力，加速赋能千行百业高质量发展。 提升精准度 近年来，极端天气气候事件频发，"报得快""报得准"的需求和难度也随之增加。中国气象局党组书 记、局长陈振林指出，人工智能技术以其高效的计算和多源数据融合能力，正成为连接气象预报、灾害 预警和应急响应的关键纽带，是突破传统预报局限的"金钥匙"。 聚焦雷暴、短时强降水等灾害性天气的临灾预警，"风雷"模型的表现尤为亮眼。其回波预报产品可 在几分钟内预测对流系统的新生与消散，强回波预报质量提升超过25%。 国家气象中心风雷敏捷攻关团队负责人张小雯介绍，升级推出的"风雷"定量降水预报模型，已在多 ...

Core Insights - The article discusses the advancements in artificial intelligence (AI) models for meteorological services in response to increasing climate variability and extreme weather events, highlighting the launch of new AI models by the China Meteorological Administration (CMA) [2][3][4] Group 1: AI Model Developments - The CMA has released the AI model "Fengyuan" and upgraded three other models: "Fengqing," "Fenglei," and "Fengshun," enhancing the precision and timeliness of weather forecasts [2][3] - The "Fenghe" model, with a trillion parameters, is designed to provide personalized weather services across various sectors, including transportation and agriculture [3][4] Group 2: Enhanced Forecasting Capabilities - The upgraded "Fengshun" model focuses on key meteorological factors such as daily maximum/minimum temperatures and solar radiation, significantly improving the ability to respond to agricultural weather disasters [4][5] - The "Fenglei" model has shown a 25% improvement in the quality of severe weather predictions, particularly for thunderstorms and short-term heavy rainfall [5][6] Group 3: Integration of AI in Meteorology - AI technology is becoming a crucial link between weather forecasting, disaster warning, and emergency response, addressing the challenges posed by extreme weather events [5][6] - The "Fengyuan" model can directly analyze real-time observational data from satellites and weather stations, aiming to enhance the accuracy of weather predictions [6][7] Group 4: Future Directions - The CMA's strategy for 2025-2035 emphasizes the integration of numerical forecasting with AI, aiming to improve the simulation and prediction capabilities of Earth system interactions [7] - The AI models are expected to provide higher resolution forecasts and longer prediction periods, significantly enhancing disaster response and decision-making processes [7]

Core Viewpoint - The article discusses the advancements in artificial intelligence (AI) models for meteorological services in China, highlighting the launch of new AI models by the China Meteorological Administration (CMA) to enhance weather forecasting and disaster warning capabilities in the context of increasingly complex climate conditions. Group 1: AI Model Developments - The CMA has released the "Fengyuan" AI model and upgraded three other models: "Fengqing," "Fenglei," and "Fengshun," to improve weather forecasting and warning capabilities [1][2] - The "Fenghe" model, with a trillion parameters, is designed for personalized weather services across various sectors, including transportation and health [2] Group 2: Applications and Benefits - The upgraded "Fengshun" model focuses on agricultural, renewable energy, and water resource needs, adding over ten key meteorological factors such as daily maximum and minimum temperatures [3] - The "Fenglei" model has shown a 25% improvement in the quality of severe weather predictions, particularly for thunderstorms and heavy rainfall [4] Group 3: Technological Integration - AI technology is becoming a crucial link between weather forecasting, disaster warning, and emergency response, addressing the challenges posed by extreme weather events [4] - The "Fengyuan" model can directly analyze real-time observational data from satellites and weather stations, enhancing the accuracy of global weather forecasts [5] Group 4: Future Directions - The CMA's strategy for 2025-2035 emphasizes the integration of numerical forecasting with AI to improve simulation and prediction capabilities of Earth's systems [7] - The AI models are expected to significantly enhance forecasting precision and efficiency, with the "Fengqing" model capable of generating 15-day global weather forecasts in just three minutes [7]

在我们看不见的数字世界里，通过"妈祖（MAZU）"全民早期预警云平台，中国气象人工智能正横 跨亚洲、非洲、大洋洲，把气象全民早期预警中国方案输送到世界各地，在地球上空织就一条"气象丝 路"，为全球防灾减灾、应对气候变化贡献中国智慧、中国经验。 跨越山海，气象人工智能方案在40多个国家试用 每天，地球各个角落，都可能在经历极端天气。 在阿富汗，一场强对流来袭，局地伴有冰雹；在乌干达，接连两场强对流突袭，伴随局地短时强降 雨…… 强对流天气尺度小、突发性强、生命史短、影响大，是气象预报的世界性难题。依靠中国气象人工 智能，阿富汗、乌干达均提前捕捉到信号，迅速发布预警，为转移群众、减少损失提供支撑。 这些走出国门的中国气象人工智能，有一个共同的名字"妈祖"。 "妈祖"是中国气象局响应联合国倡议推出的全民早期预警方案，缩写"MAZU"有着严谨的科学内涵 ——M（Multi—hazard，多灾种）、A（Alert，预警）、Z（Zero—gap，零差距）、U（Universal，普 惠）。 "妈祖"通过建设云端早期预警平台助力各国应对气候变化挑战。平台上每种人工智能都是一个"工 具箱"，针对不同国家需求，提供定制化"工具 ...

每天，地球各个角落，都可能在经历极端天气。 在阿富汗，一场强对流来袭，局地伴有冰雹；在乌干达，接连两场强对流突袭，伴随局地短时强降 雨…… 强对流天气尺度小、突发性强、生命史短、影响大，是气象预报的世界性难题。依靠中国气象人工智 能，阿富汗、乌干达均提前捕捉到信号，迅速发布预警，为转移群众、减少损失提供支撑。 这些走出国门的中国气象人工智能，有一个共同的名字"妈祖"。 在我们看不见的数字世界里，通过"妈祖（MAZU）"全民早期预警云平台，中国气象人工智能正横跨亚 洲、非洲、大洋洲，把气象全民早期预警中国方案输送到世界各地，在地球上空织就一条"气象丝路"， 为全球防灾减灾、应对气候变化贡献中国智慧、中国经验。 跨越山海，气象人工智能方案在40多个国家试用 服务场景越来越广泛，有云端版、基础版、专业版、旗舰版、城市早期预警版、农业早期预警版等，适 配多种气象灾害和城市、农业等各类气象风险场景。 "目前，'妈祖'已经在巴基斯坦、埃塞俄比亚、所罗门群岛、吉布提、蒙古国5个国家实现落地部署与实 时应用，在亚洲、非洲、大洋洲的43个国家和地区开展在线试用。"中国气象局国际合作司司长曾沁介 绍。 中国气象局坚持"妈祖"人工 ...

编者按： 人才蔚起，千帆竞逐；创新潮涌，引领未来。"十四五"以来，面向世界科技前沿、面向经济主战场、面 向国家重大需求、面向人民生命健康，中国气象局党组谋篇布局，陆续组建一批覆盖气象科研业务服务 核心领域重点创新团队，聚焦气象关键核心领域"卡脖子"问题持续攻关，着力打造布局合理、开放高 效、支撑有力、充满活力、与业务服务深度融合的科技创新团队体系，加快实现高水平气象科技自立自 强。 当前，首批创新团队建设期满，在气象核心技术突破、业务服务能力提升及科研成果转化等方面成效显 著，为推动气象高质量发展注入强劲动力。为展现这些团队的奋斗足迹与卓越成就，弘扬科学家精神， 激发更多气象科技工作者勇攀高峰，中国气象报开设专栏，带您走近各支重点创新团队，探寻他们在风 云变幻中矢志不渝的创新故事。 8月，第12届世界运动会在四川成都举办。恰逢汛期，复杂多变的天气、密集的户外赛事，对气象服务 保障提出极高要求。在赛事期间的短期预报环节，人工智能全球中短期预报系统"风清"成为核心支撑力 量，精准预测短中期尺度的温度、湿度等要素，为每日赛事调度提供"场馆级"精细预报。 而这双精准洞察风云变幻的"慧眼"，源自一支成立仅三年的年轻团队 ...

（总台央视记者 刘璐璐） 责编：张青津、卢思宇 中国气象局今天上午举行新闻发布会，会上发布《地球系统预报发展战略（2025—2035年）》。 地球系统预报是将地球各个圈层作为一个相互耦合的整体，在统一框架下对大气、海洋、陆地、冰冻圈 和生物圈等多圈层进行综合模拟和预测的预报体系。与传统预报相比，二者基础理论和技术体系一脉相 承，同样依托观测资料、物理规律和高性能计算支撑。但传统预报主要聚焦大气、侧重天气现象刻画， 而地球系统预报通过融合数值模式、人工智能和数字孪生等先进技术，实现对多圈层相互作用及其反馈 机制的整体表征，不仅拓展了预报对象的覆盖范围，也更加注重风险形成、演变和影响评估，能为防灾 减灾、应对气候变化和服务经济社会发展提供更有力的科技支撑。 《战略》系统规划了八大重点任务。 《战略》的目标：到2035年，要建成自主可控、国际先进的地球系统预报体系。这个目标将通过两个阶 段稳步推进。未来五年内，重点实现新一代预报模式的业务化运行，并建成气象人工智能模型的统一基 础框架。推动各项技术的深度融合，发布第一代地球系统数字服务平台。在2030年至2035年间，全面建 成统一框架的地球系统预报体系，显著提升对 ...

天气预报从来不是简单的"看云识天气"，而是一场与自然规律的持久博弈。面对短临强对流天气 的"猝不及防"、气候变化带来的极端天气频发等难题，气象工作者从未停下攻坚克难的脚步。他们组建 科研团队，进行数值预报模式研发，让预报时效从小时级向分钟级延伸；他们与陆海空天一体化观测共 同成长，让智慧气象服务平台实现预报预警信息精准推送……这背后是他们无数个日夜的钻研与坚守。 为了让天气预报更准一些，他们在平凡的岗位上书写着不平凡的人生，为人民群众的美好生活贡献 力量。 坚持百分百努力 天气预报可不可以更准？ 国家气候中心首席预报员高辉的答案是："天气预报不可能100%准确，但我们在付出100%的努 力。" 节奏是跳跃的，压力是常态的，责任更是重大的。高辉这样描述首席预报员的工作状态："很多时 候，思维的切换必须在瞬间完成。短短一个上午，可能刚与华南的同事会商完台风动态，紧接着就要转 向华北研判高温干旱，或是和西南的同事会商秋雨。" 而他和天气打交道，源于一场滞留。1998年盛夏，长江流域发生罕见特大洪涝灾害。高辉陪着因铁 路受损返乡无望的舍友，在宿舍里听了半个月的风雨声。"那时我就想，如果能提前一个月，甚至更早 预见这种 ...

Core Insights - The article discusses the introduction of the "Variable Adaptive Mixture of Experts (VA-MoE)" model by research teams from the University of Hong Kong and Zhejiang University, which aims to enhance weather forecasting by allowing for incremental learning without the need for complete retraining when new variables or stations are added [1][2]. Group 1: Model Overview - VA-MoE utilizes a phased training approach and variable indexing embedding mechanism to guide different expert modules to focus on specific types of meteorological variables, significantly reducing computational costs while maintaining accuracy [1][2]. - The model is designed to address the challenges posed by the asynchronous nature of meteorological data collection, which often requires full retraining of existing AI models when new variables are introduced [2]. Group 2: Research Highlights - The research results have been accepted at the ICCV25 conference under the title "VA-MoE: Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting" [3]. - The study employs the ERA5 dataset, which includes continuous meteorological observation data from 1979 to the present, ensuring a comprehensive experimental foundation [5]. Group 3: Experimental Design - The dataset is divided into high-altitude variables for initial training and ground variables for incremental training, simulating the dynamic expansion of variables in actual observations [6][10]. - The initial training phase uses 40 years of data (1979-2020), while the incremental training phase uses 20 years of data (2000-2020) to adapt to new variable introductions [8]. Group 4: Performance Evaluation - VA-MoE has been shown to outperform similar models in upper-air variable forecasting, achieving significant improvements even with reduced data and parameters [7][20]. - The model's two-stage training strategy allows it to maintain or improve accuracy when new ground variables are introduced, demonstrating its capability to learn new information without losing previously acquired knowledge [20][21]. Group 5: Industry Implications - The advancements in AI-driven weather forecasting, as exemplified by VA-MoE, indicate a shift towards more efficient and adaptable modeling techniques that can better handle the complexities of meteorological data [22]. - The ongoing collaboration between academia and industry is expected to further drive innovations in weather modeling, enhancing predictive capabilities and operational efficiencies [23][25].