融合人工智能气象模型的热带气旋路径集成预报技术研究

打开文本图片集

Consensus forecast technology for tropical cyclone tracks by integrating AI weather prediction models

WANG Xuan， QI Liangbo （ShanghaiCentralMeteorologicalObservatory，Shanghai2Ooo3O，China）

AbstractTo improve the consensus forecast of tropical cyclone （TC） tracks over western North Pacific， this study constructs 6 consensus schemes by combining three artificial intelligence （AI） weather prediction models（Pangu，FuXiand FengWu）with numerical weather prediction（NWP）models （including deterministic and ensemble forecasts）using three consensus methods：simple multi-model average（AVG），selective ensemble average（SEAV） and selective ensemble change-weighted average （SECW）.These schemes are evaluated for their 120-h TC track forecasts from 2023 to 2024.The results areas follows.（1） Individual AI models exhibit clear advantages over NWP in TC track forecast（most notably at longer lead times）.Compared to the ECMWF ensemble mean，FengWu shows comprehensive superiority，while the other two AI models perform worse at 24h and 48h ，but better at 72-120h Against pure NWP consensus forecasts，FengWu has slightly higher errors at 24-72h ，but significant advantages at 96h and 120h . （2）The integrated schemes combining AI and NWP leverage the strengths of both approaches， substantially outperforming the ECMWF-EPS ensemble mean. They generally surpass both individual AI models and pure NWP consensus forecasts，and exceed pure AI consensus forecasts within 72h.At96hand 120h ，however，the pure AI consensus forecasts show the most significant advantages.（3） Compared to the operational consensus forecasts from the Shanghai Typhoon Institute， the three proposed integrated schemes—incorporating additional ensemble and AI model members— significantly reduce track errors and enhance forecast stability，with the scheme performing the best.This study demonstrates that AI models have promising application potential in TC track consensus forecast.

KeywordsAI weather prediction model; TC track forecast； consensus forecast; forecast evaluation

0 引言

得益于数值天气预报（numericalweatherprediction，NWP）模式的改进、卫星观测数据的增加、资料同化技术的发展以及集合预报和多模式集成预报的应用，过去30年西北太平洋热带气旋（tropicalcyclone，TC）的路径预报能力显著提升，2015—2022年官方TC预报机构 24h.48h.72h 路径预报平均误差的区间分别为 65～100km，110～ 160 km、175～255 km[1]

相对于单个预报模式，多模式集成预报的优势被广泛认可，已经成为国内外主流气象机构开展TC路径预报的重要依据[2]。（剩余11420字）