白杨,李威,邵祺.基于经验正交函数和机器学习的南海海面高度异常预测[J].海洋通报,2020,(6):
基于经验正交函数和机器学习的南海海面高度异常预测
A prediction model of Sea Surface Height Anomaly based on Empirical Orthogonal Function and machine learning
投稿时间:2020-02-04  修订日期:2020-04-26
DOI:10.11840/j.issn.1001-6392.2020.06.005
中文关键词:  海洋涡旋  主成分  经验正交函数  惯性预报  气候态预报
英文关键词:ocean eddies  principal components  Empirical Orthogonal Function  persistent forecast  climatology
基金项目:国家重点研发计划 (2016YFC1401800;2018YFC1406206;2017YFC1404103);国家自然科学基金 (41876014;11801402)
作者单位E-mail
白杨 天津大学 海洋科学与技术学院天津 300072 1324180078@qq.com 
李威 天津大学 海洋科学与技术学院天津 300072 liwei1978@tju.edu.cn 
邵祺 天津大学 海洋科学与技术学院天津 300072  
摘要点击次数: 171
全文下载次数: 72
中文摘要:
      海面高度异常 (SSHA) 作为重要的海洋要素,对研究海洋温盐剖面、海洋涡旋等海洋动力现象具有重要意义。然而,传统的海洋预测技术存在着预测时效过短、预测过程复杂等诸多问题,现有的机器学习预测方法也只针对几个点或区域 进行平均,忽略了很多重要信息。因此,本文提出了一种基于经验正交函数和 BP 神经网络 (一种机器学习方法) 的 SSHA预测模型 (EOF-BPNN) 来实现对起报时刻后 30 天的南海 SSHA 预测。首先,对 1993 年 1 月 1 日—2013 年 12 月 31 日的逐日 SSHA 数据进行距平归一化预处理,构建相关系数矩阵,并对该矩阵进行 EOF 分解,获取主成分。然后将主成分输入 BP神经网络进行训练,实现对主成分的预测。最后将主成分预测值与相应的空间模态结合,获取 SSHA 预测值。结果表明,相较于惯性预报和气候态预报,EOF-BPNN 模型不仅能够提供提前 30 天的较为精确的 SSHA 和相应的涡旋演化过程预报,且在整个南海区域拥有更高的 SSHA 相关系数,证明了 EOF-BPNN 模型具有较好的预测性能。
英文摘要:
      As a vital marine element, the Sea Surface Height Anomalous (SSHA) is of great significance for studying marine temperature-salinity profile, ocean eddies and other marine dynamic phenomena. However, for traditional ocean prediction techniques, there are many problems in the SSHA prediction, such as too low prediction aging and complicated prediction procedure, the current machine learning forecasting methods are also aimed at a few points or regional averages which neglect much vital information. Thus, a prediction model of SSHA based on Empirical Orthogonal Function and BP Neural Network which is a machine learning method (EOF-BPNN) is proposed in this article to predict SSHA 30 days in advance in the South China Sea (SCS), and a series of processes are conducted. In order to solve memory demand owing to the large scale data, the PCs can be obtained by using the data preprocessing and EOF analysis. Then, the PCs 30 days in advance can be forecast by using the BP Neural Network. Further, the PCs prediction can be combined with corresponding spatial patterns to reconstruct the SSHA prediction. The results have demonstrated that the EOF-BPNN model is not only more capable of predicting the SSHA 30 days in advance, but also provide better accuracy and capture of eddies than persistence, climatology, the greater anomaly correlation coefficients )(CCs) and smaller root mean squared errors (RMSEs) are shown that the EOF-BPNN model is efficient for SSHA prediction in the entire SCS region.
查看全文    下载PDF阅读器
关闭