基于Swin-Transformer和时空融合注意力机制的ENSO预测  

ENSO prediction based on Swin-Transformer and spatio-temporal fusion attention mechanism

作　　者：张霄智 方巍[1,2,3] 王淏西 Zhang Xiaozhi;Fang Wei;Wang Haoxi(School of Computer Science&School of Software,Nanjing University of Information Science and Technology,Nanjing 210044,China;China Meteorological Administration,China Meteorological Administration Basin Heavy Rainfall Key Laboratory/Hubei Key Laboratory for Heavy Rain Monitoring and Warning Research,Institute of Heavy Rain,Wuhan 430205,China;State Key Laboratory of Severe Weather,Chinese Academy of Meteorological Sciences,Beijing 100081,China)

机构地区：[1]南京信息工程大学计算机学院/软件学院,江苏南京210044 [2]中国气象局武汉暴雨研究所中国气象局流域强降水重点开放实验室/暴雨监测预警湖北省重点实验室,湖北武汉430205 [3]中国气象科学研究院灾害天气国家重点实验室,北京100081

出　　处：《海洋学报》2024年第12期111-121,共11页

基　　金：国家自然科学基金面上项目(No.42475149);中国气象局流域强降水重点开放实验室开放研究基金(No.2023BHR-Y14);南京气象科技创新研究院北极阁开放研究基金(BJG202306);灾害天气国家重点实验室开放课题(2024LASW-B19);江苏省研究生科研与实践创新计划项目(NO.KYCX24_1533)。

摘　　要：厄尔尼诺-南方涛动预测是气候变化研究的热点问题之一。本文将Swin-Transformer模型与时空融合注意力机制相结合,采用1850-2014年CMIP6多模式模拟历史数据、1871-1979年SODA同化数据和1980-2023年GODAS同化数据,构建厄尔尼诺-南方涛动预测模型,即ENSO-STformer。该模型通过在CMIP6和SODA数据集上进行充分的训练,并在GODAS数据上进行评估,结果表明:本文模型在提前11个月的Ni?o3.4指数相关技巧的平均值上分别比CanCM4、CCSM3、GFDLaer04动力预报系统高出5.1%、21.6%和12.4%,同时,在中长期的Ni?o3.4指数相关技巧上显著优于其他深度学习模型,并可以进行长达24个月的有效ENSO预测,此外,在对2015-2016年厄尔尼诺事件模拟中表现出较强的应对春季预报障碍能力。The prediction of El Niño-Southern Oscillation is one of the hot issues in climate change research.This paper combines swin-transformer model with spatio-temporal fusion attention mechanism,and uses CMIP6 multimodel simulation historical data from 1850 to 2014,SODA assimilated data from 1871 to 1979 and GODAS assimilated data from 1980 to 2023 to construct El Niño-Southern Oscillation prediction model—ENSO-STformer.The model was fully trained on CMIP6 and SODA datasets and evaluated on GODAS data.The results show that the average skill of this model in predicting the Niño3.4 index at 11-month lead times exceeds those of CanCM4,CCSM3,and GFDLaer04 by 5.1%,21.6%,and 12.4%respectively.Meanwhile,the Niño3.4 index related skills of the proposed model are significantly better than other deep learning models in the medium and long term.Effective ENSO forecasts can be made for up to 24 months,and the 2015−2016 El Niño event simulation shows strong ability to cope with spring forecast obstacles.

关 键 词：深度学习 ENSO预测 时空融合注意力机制 卷积神经网络 Ni?o3.4指数 

分 类 号：P467[天文地球—大气科学及气象学]