基于LSTM的北极海冰范围多步预测策略研究  

Research on multi-step prediction strategies of Arctic sea ice extent based on Long Short-Term Memory

作　　者：王漫漫 邹斌[2,3] 石立坚 曾韬[2,3] 张颖 路敦旺[1,2,3] WANG Manman;ZOU Bin;SHI Lijian;ZENG Tao;ZHANG Ying;LU Dunwang(National Marine Environmental Forecasting Center,Beijing 100081,China;National Satellite Ocean Application Service,Beijing 100081,China;Key Laboratory of Space Ocean Remote Sensing and Application,Ministry of Natural Resources,Beijing 100081,China;Key Laboratory of Marine Hazards Forecasting,National Marine Environmental Forecasting Center,Ministry of Natural Resources,Beijing 100081,China)

机构地区：[1]国家海洋环境预报中心,北京100081 [2]国家卫星海洋应用中心,北京100081 [3]自然资源部空间海洋遥感与应用研究重点实验室,北京100081 [4]国家海洋环境预报中心自然资源部海洋灾害预报技术重点实验室,北京100081

出　　处：《海洋预报》2025年第1期11-22,共12页Marine Forecasts

基　　金：国家重点研发计划项目(2021YFC2803300、2022YFC2807003)。

摘　　要：已有研究对北极海冰范围开展单步预测,而多步预测及其策略研究有待进一步探索。使用1979-2022年的北极月平均海冰范围数据,采用长短期记忆网络(LSTM)深度学习方法,结合递归(Recursive)、直接(Direct)、多输入多输出(MIMO)和Seq2Seq策略实现对未来12个月北极海冰范围的多步预测。结果表明:24个月为模型的最佳输入长度;与另外3种基本的多步预测策略相比,Seq2Seq策略对12个月北极海冰范围预测的准确性更好,均方根误差为3.30×10^(5)km^(2)。While previous researches have primarily focused on single-step prediction of Arctic sea ice extent,multi-step prediction and strategy are yet to be explored.This study utilizes monthly average Arctic sea ice extent data spanning from 1978 to 2022 and employs Long Short-Term Memory to implement multi-step predictions of Arctic sea ice extent for the next 12 months using four strategies:Recursive,Direct,Multi-input Multi-output,and Seq2Seq.The results show that a model input length of 24 months performs optimally.When compared to the other three basic multi-step prediction strategies,the Seq2Seq strategy demonstrates superior accuracy in forecasting Arctic sea ice extent over the next 12 months,with an root mean square error of 0.33 million square kilometers.

关 键 词：北极海冰范围 长短期记忆网络 多步预测策略 Seq2Seq策略 

分 类 号：P731.32[天文地球—海洋科学]