基于多尺度深度学习对南海海表温度预报的研究  

Forecast of sea surface temperature in the South China Sea based on multi-scale deep learning model

作　　者：张宇 许大志 俞胜宾 邢会斌 管玉平[4,5] Zhang Yu;Xu Dazhi;Yu Shengbin;Xing Huibin;Guan Yuping(South China Sea Marine Forecast and Hazard Mitigation Center,Guangzhou 510310,China;Key Laboratory of Marine Environment Survey Technology and Application,Ministry of Natural Resource,Guangzhou 510310,China;Southern Marine Science and Engineering Guangdong Laboratory(Zhuhai),Zhuhai 519000,China;State Key Laboratory of Tropical Oceanography,South China Sea Institute of Oceanology,Chinese Academy of Sciences,Guangzhou 510301,China;College of Marine Science,University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]自然资源部南海预报减灾中心,广东广州510310 [2]自然资源部海洋环境探测技术与应用重点实验室,广东广州510310 [3]南方海洋科学与工程广东省实验室(珠海),广东珠海519000 [4]中国科学院南海海洋研究所热带海洋环境国家重点实验室,广东广州510301 [5]中国科学院大学海洋学院,北京100049

出　　处：《海洋学报》2024年第5期27-36,共10页

基　　金：自然资源部海洋环境探测技术与应用重点实验室自主设立研究课题(MESTA-2021-D003);广州市基础与应用基础研究项目(202201011271);国家海洋局南海预报中心自主立项项目(SCSMF-FR-2021-07);国家自然科学基金青年科学基金项目(42206027);中国-东盟国家蓝色伙伴关系建设项目(99950410)。

摘　　要：海表温度是海洋最重要的物理量之一,提供了气候系统的基本信息,准确地预报海表温度有着广泛而重要的应用。近年来,基于人工智能的海温预报方法开始流行,并展现出巨大的潜力。基于卷积长短时记忆神经网络(ConvLSTM),本文研究了多尺度输入场对南海北部二维海表温度预报结果的影响。文章采用多元集合经验模态分解方法(MEEMD)将日均海表温度分解成多个尺度的空间主模态,并以不同的组合训练ConvLSTM模型进行预报实验。结果表明,采用前4个海表温度主模态数据训练模型时,预报1~7 d海表温度的均方根误差约为0.4~0.8℃,比仅用原始海表温度训练时减小了0.2~1.2℃;平均绝对百分比误差为1%~6%,减小了0.5%~10%;空间相关系数为99.5%~96.5%,提高了0.5%~3.5%。而且,随机实验也进一步证明该方法具有较高的普适性。基于深度学习的预报模型,需结合海温的物理特性,选择合适的数据进行训练,才能进一步提高其预报精度。本文初步探究了人工智能方法与物理概念在海温预报中的融合,可为以后的研究提供一定的参考。Sea surface temperature(SST) is one of the most important physical variables of the ocean,which provides the basic information of the climate system.Accurately SST forecasting system has a comprehensive and essential application.In recent years,AI-based SST forecasting methods have become popular and shown great potential.Based on the convolutional long and short-term memory neural network(ConvLSTM),this paper studies the influence of multi-scale input fields on SST prediction in the northern South China Sea.Multi-dimensional ensemble empirical mode decomposition method(MEEMD) is used to decompose the average daily SST into the spatial eigenmodes of differentiated scales.Input different combinations of eigenmodes into ConvLSTM for training and prediction experiments.Results show that when using all four SST eigenmodes,the RMSE of the predicted SST in 1-7 days is 0.4-0.8℃,decrease 0.2-1.2℃ compared with the original SST alone;the MAPE is 1%-6%,decrease 0.5%-10%;the spatial correlation coefficient is 99.5%-96.5%,improve 0.5%-3.5%.Moreover,the randomized experiments also further proved the method has a high universality.The prediction model based on deep learning needs to select the appropriate training data in order to further improve its prediction accuracy.This paper preliminarily explores the integration of artificial intelligence methods and physical concepts in SST prediction,which can provide some reference for future research.

关 键 词：海温预报 深度学习 ConvLSTM MEEMD 

分 类 号：P731.31[天文地球—海洋科学]