《海洋预报》| 人工智能气象模型台风路径预报评估和集成应用研究

摘要：评估了2023—2024年3个人工智能（AI）气象预报模型（盘古、风乌和伏羲）与欧洲中期天气预报中心（ECMWF）高分辨率数值模式EC在西北太平洋的路径预报性能（计38个台风），对比分析了不同预报时效、不同路径趋势、不同发展强度、不同发展阶段各方法的优劣，在此基础上，对AI模型和数值预报模式的集成应用方案进行了测试。结果表明：AI模型在短期和中长期预报中均展现出优势，其中风乌模型在中短期（24～72 h）表现最优，伏羲模型在长时效稳定性更佳；在西行路径台风预报中，AI气象模型展现出显著优势，其中72～120 h预报时效的平均路径误差较EC降低约100 km；强台风（TY级及以上）的路径预报中，AI模型在72 h以上时效误差增长率比EC低15%～20%；集成方案测试表明，针对西行类台风采用纯AI集成（加权集成Ⅱ）可降低误差，非西行类需保留EC以平衡不确定性，对不同路径趋势采用不同集成方案可望取得更高预报技巧。

关键词：人工智能气象模型 台风路径预报 检验评估 集成应用

Abstract：This study evaluates the track forecast performance of three AI-based meteorological prediction models(i. e., "PANGU", "FENGWU", and "FUXI") against the high-resolution numerical model from the European Centre for Medium-Range Weather Forecasts(EC) over the Northwest Pacific for 38 typhoons during 2023—2024. A comparative analysis is conducted across different forecast lead times, track tendencies, intensities, and initial stages. Based on the evaluation, an ensemble application strategy integrating AI models and numerical prediction system is tested. The results show that: AI models exhibit advantages in both short-term and medium-to-long-term forecasts, with the FENGWU model performing optimally in the medium-short term(24～72 h) and the FUXI model demonstrating better stability in long lead times(>96 h); AI-based meteoro-logical models show significant advantages in forecasting typhoons with a westward track, with the average track error reduced by approximately 100 km compared to EC for the 72～120 h lead time; for track forecasts of intense typhoons(above TY grade), the error growth rate of AI models is 15%～20% lower than that of EC for lead times >72 h; Tests on the integrated scheme indicate that a pure AI integration(weighted integration Ⅱ) for westward-track typhoons can reduce errors, while non-westward-track typhoons require retaining EC to balance uncertainties. Adopting different integration schemes for varying track trends is expected to achieve higher forecast skill.

Key words：artificial intelligence meteorological model; typhoon track forecast; statistic verification； multimodel integration

读者也可以通过公众号菜单栏

“相关知识”板块进入《当期目录》和《过刊浏览》

查阅详细期刊内容

未来两天海洋环境预报

35个美丽海岛环境预报