GFS、CMA-GFS、WRF模式与AI气象大模型对台风“麦德姆”预报能力对比分析

Comparative analysis of GFS, CMA‑GFS, WRF and AI weather models for Typhoon Matmo forecast

摘要

【目的】评估 2025 年台风“麦德姆”（2521）路径、强度及结构预报中多模式性能差异，为业务预警与模式改进提供参考。【方法】基于中国气象局（CMA）最佳路径数据与欧洲中期天气预报中心再分析数据（ERA5），对比两类全球数值模式、两类区域降尺度模式及两种人工智能（AI）气象大模型的预报结果，采用平均绝对误差（MAE）和均方根误差（RMSE）定量评估 24、48、60 与 96 h 预报时效的路径与强度误差，并结合暖心、动力结构及大尺度环流诊断分析误差机制。【结果】在 24～60 h 路径预报中，FuXi 模型精度最优，48 h MAE 为 25 km；96 h RMSE 增至 417.4 km。CMA-GFS（China Meteorological Administration Global Forecast System ）模型短时效偏差较大，但随预报时效延长误差增长较缓；GFS 模型（Global Forecast System）短期预报较好，长时效易偏移。WRF-GFS（Weather Research and Forecasting model driven by the Global Forecast System）与 WRF-CMA-GFS（Weather Research and Forecasting model driven by the China Meteorological Administration Global Forecast System）短时效误差偏大，但长时效稳定性突出。在强度预报方面，Pangu 和 FuXi 模型全程显著低估强度，无法刻画近海快速降压过程；WRF-CMA-GFS 对快速增强与峰值强度的模拟最优，且预报质量显著依赖初始场。【结论】GFS 与 CMA-GFS 在路径与强度预报中表现出显著差异：路径预报上二者呈短–长时效能力反转特征，即 GFS 短期占优但长时效误差骤增，CMA-GFS 短期偏差明显但长时效最稳定；强度预报表现则与该特征相反。WRF 模式长时效预报性能突出，且质量显著依赖初始场。本文通过“麦德姆”这一近海增强个例的多模式对比，揭示了不同模式与 AI 模型在不同时效中的差异化表现及结构模拟的关键作用，研究结果可为台风多模式集成优选、时效策略制定提供量化依据，为 AI 模型融入物理约束、区域模式优化初始场提供改进方向，对提升近海增强台风预报业务能力具有一定应用价值。

Abstract

[Objective] This study evaluates the performance differences among multiple models in forecasting the track, intensity, and structure of Typhoon Matmo (2521), aiming to provide references for operational warning and model improvement. [Method] Based on the China Meteorological Administration (CMA) best track data and European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis data (ERA5), this study compared the forecast results of two global numerical models, two regional downscaling models, and two artificial intelligence (AI) meteorological large models. The mean absolute error (MAE) and root mean square error (RMSE) were used to quantitatively evaluate track and intensity errors at 24-hour, 48-hour, 60-hour, and 96-hour forecast horizons. Error mechanisms were also diagnosed through analyses of thermal structure, dynamic structure, and large-scale circulation. [Result] For 24-60-hour track forecasts, the FuXi model demonstrated the highest accuracy, with a 48-h MAE of 25 km; however, the 96-hour RMSE increased to 417.4 km. The China Meteorological Administra- tion Global Forecast System (CMA-GFS) model exhibited larger errors in the short-range forecast period, but its error growth rate slowed as the forecast horizon extended; conversely, the Global Forecast System (GFS) per- formed well in the short term but tended to deviate in the long term. The Weather Research and Forecasting (WRF) models driven by the Global Forecast System (WRF-GFS) and Weather Research and Forecasting model driven by the China Meteorological Administration Global Forecast System (WRF-CMA-GFS) models had relatively large errors in the short range forecast period but demonstrated outstanding stability in the long rangeforecast pe- riod. Regarding intensity forecasts, Pangu and FuXi significantly underestimated intensity throughout the entire track and failed to capture rapid pressure drops near the coast. WRF-CMA-GFS performed best in simulating rap- id intensification and peak intensity, although its forecast quality was significantly dependent on initial conditions. [Conclusion] GFS and CMA-GFS exhibit significant differences in track and intensity forecasts. Regarding tracks, the two models show a reversal of short-term versus long-term capabilities: GFS performs better in the short term but shows a sharp increase in errors over the long term, while CMA-GFS exhibits significant short-term devia- tions but is the most stable over the long term. Intensity forecasts, however, show the opposite pattern. The Weather Research and Forecasting WRF model demonstrates outstanding long-range forecasting performance, with intensity forecasts generally outperforming track forecasts, though forecast quality remains significantly de- pendent on initial conditions. Through a multi-model comparison of Typhoon Matmo—a rapid offshore intensifi- cation case-this study reveals the differentiated performance of various models across different time scales and highlights the critical role of structural simulation. These findings provide quantitative evidence for optimizing typhoon multi-model ensembles and formulating time-scale-dependent strategies, offer directions for integrating physical constraints into AI models and optimizing initial conditions for regional models, and hold practical value for enhancing operational forecasting capabilities for rapidly intensifying typhoons near the coast.

关键词

台风“麦德姆” / 数值预报模式 / AI 气象大模型 / 路径预报 / 强度预报

Key words

Typhoon Matmo / numerical weather prediction models / AI weather models / track forecasting / intensity forecasting

参考文献

[1] Zhang G. Warming—induced contraction of tropical con— vection delays and reduces tropical cyclone formation[J].Nat Commun, 2023, 14: 6274.

[2] Shan K Y, Lin Y L, Chu P S, et al. Seasonal advance of intense tropical cyclones in a warming climate[J].Nature, 2023, 623(7985): 83-89.

[3] Liu Y H, Guan S D, Lin I I, et al. Effect of storm size on sea surface cooling and tropical cyclone intensification in the western north Pacific[J].J Clim, 2023, 36(20): 7277-7296.

[4] Saha S, Moorthi S, Wu X R, et al. The NCEP climate fore— cast system version 2[J].J Clim, 2014, 27(6): 2185-2208.

[5] 沈学顺, 苏勇, 李兴良, 等. 中国业务数值天气预报自主发展中的理论和方法创新[J].气象, 2025, 51(11): 1293-1320.

[6] 张进, 孙健, 沈学顺, 等. CMA—GFS V4.0 模式关键技术研发和业务化[J].应用气象学报, 2023, 34(5): 513-526.

[7] 韩威, 尹若莹, 李俊, 等. 静止气象卫星大气廓线目标观测及在高影响天气预报中的影响[J].中国科学: 地球科学, 2025, 55(4): 991-1005.

[8] Bi K F, Xie L X, Zhang H H, et al. Accurate medi— um—range global weather forecasting with 3D neural net— works[J].Nature, 2023, 619(7970): 533-538.

[9] Chen L, Zhong X H, Zhang F, et al. FuXi: a cascade ma— chine learning forecasting system for 15—day global weather forecast[J].npj Clim Atmos Sci, 2023, 6: 190.

[10] Sun X Y, Zhong X H, Xu X Z, et al. A data—to—forecast machine learning system for global weather[J].Nat Commun, 2025, 16: 6658.

[11] Shi Y, Hu R, Wu N G, et al. Comparison of AI and NWP models in operational severe weather forecasting: a study on tropical cyclone predictions[J].J Geophys Res Mach Learn Comput, 2025, 2(2): e2024JH000481.

[12] Bonavita M. On some limitations of current machine learning weather prediction models[J].Geophys Res Lett, 2024, 51(12): e2023GL107377.

[13] 中华人民共和国应急管理部 . 国家防灾减灾救灾委员会办公室 应急管理部发布2025年10月全国自然灾害情况 [EB/OL]. (2025—11—11)[2026—03—12].https://www. mem. gov.cn/xw/yjglbgzdt/202511/t20251111_585250. shtml.

[14] Ying M, Zhang W, Yu H, et al. An overview of the China meteorological administration tropical cyclone database[J].J Atmos Ocean Technol, 2014, 31(2): 287-301.

[15] Skamarock W C, Klemp J B, Dudhia J, et al. A description of the advanced research WRF Model version 4[R].NCAR tech. note ncar/tn—556+ str, 2019, 145(10.5065).

[16] Lam R, Sanchez—gonzalez A, Willson M, et al. Learning skillful medium—range global weather forecasting[J].Sci— ence, 2023, 382(6677): 1416-1421.

[17] Yamaguchi M, Ishida J, Sato H, et al. WGNE intercom— parison of tropical cyclone forecasts by operational NWP models: a quarter century and beyond[J].Bull Am Meteor Soc, 2017, 98(11): 2337-2349.

[18] Cha D H, Wang Y Q . A dynamical initialization scheme for real—time forecasts of tropical cyclones using the WRF model[J].Mon Weather Rev, 2013, 141(3): 964-986.

[19] Zhang D L, Chen H . Importance of the upper—level warm core in the rapid intensification of a tropical cyclone[J].Geophys Res Lett, 2012, 39(2): 2806.

[20] Charlton—perez A J, Dacre H F, Driscoll S, et al. Do AI models produce better weather forecasts than physics—based models A quantitative evaluation case study of Storm Ciarán[J].npj Clim Atmos Sci, 2024, 7: 93.

[21] Sahu P L, Sandeep S, Kodamana H . Evaluating global machine learning models for tropical cyclone dynamics and thermodynamics[J].J Geophys Res Mach Learn Comput, 2025, 2(2): e2025JH000594.

[22] 高拴柱, 吕心艳, 王海平, 等. 热带气旋莫兰蒂(1010)强度的观测研究和增强条件的诊断分析[J].气象, 2012, 38(7): 834-840.

[23] EMANUEL K A . An air—sea interaction theory for tropical cyclones. part I: steady—state maintenance[J].J Atmos Sci, 1986, 43(6): 585-605.

[24] 王志烈, 胡坚, 丁一汇 . 西北太平洋风场对台风路径的影响[J].应用气象学报, 1991, 2(4): 362-368.

[25] 吴胜蓝, 漆梁波, 余晖 . 登陆台风热力不对称结构对路径偏折的影响[J].气象, 2025, 51(6): 675-685.

[26] 陈联寿 . 热带气旋研究和业务预报技术的发展[J].应用气象学报, 2006, 17(6): 672-681.

[27] Davis C, Wang W, Chen S S, et al. Prediction of landfall— ing hurricanes with the advanced hurricane WRF model[J].Mon Weather Rev, 2008, 136(6): 1990-2005.

基金资助

国家自然科学基金项目(42375159)

粤西海洋气象灾害理论与应用研究重点实验室开放项目(2025KSYS009)

广东海洋大学大学生创新训练项目(CXXL2026012)