摘要
海表面高度异常(SSHA)是海洋系统中重要的参数之一,论文首次利用深度学习中的长短期记忆神经网络(LSTM)对海表面高度异常(SSHA)进行预测。将SSHA的预测当作时间序列预测问题,采用stacked-LSTM,建立海表面高度异常预测模型。该模型能够捕捉SSHA序列变化的规律,处理序列变化长期依赖问题。通过实验探索了stacked-LSTM预测模型的最佳参数设置,并使用中国南海海域的CORA再分析海表面高度异常数据进行验证,在预测未来24h、48h、72h、96h、120h的SSHA值上准确率(平均值±标准差)分别达到了90.10±10.64%、84.68±14.34%、78.29±17.37%、72.65±18.96%、66.41±20.91%,并与ANN、RNN、TCN等模型进行了对比,该基于数据驱动的模型能够运行在PC终端,为海洋工程提供移动服务。
Sea surface height anomaly(SSHA)is an elemental factor in ocean environment and marine engineering.In this work,a deep learning method,named stacked-LSTM,is proposed to predict SSHA.Specifically,SSHA prediction is treated as a time series forecasting problem,and stacked-LSTM can mine the discipline hidden in short time series,and tackle long-term de⁃pendence of series changes.Data experiments is conducted on SSHA dataset of china ocean reanalysis(CORA)in the south china sea.As results,the method achieves average predicting accuracy plus/minus standard deviation of coming 24h,48h,72h,96h and 120h by 90.10±10.64%,84.68±14.34%,78.29±17.37%,72.65±18.96%,66.41±20.91%,respectively.The method performs bet⁃ter than several state-of-the-art machine learning methods,including artificial neural network(ANN),recurrent neural network(RNN),time convolutional network(TCN).The data driven SSHA predicting method can be run on PC terminal or service,thus providing mobile service to marine engineering.
作者
江璟瑜
徐丹亚
韩宁生
王子赫
JIANG Jingyu;XU Danya;HAN Ningsheng;WANG Zihe(College of Computer and Science Technology,China University of Petroleum,Qingdao 266580;Southern Marine Science and Engineering Guangdong Laboratory(Zhuhai),Zhuhai 519080)
出处
《舰船电子工程》
2021年第2期97-99,共3页
Ship Electronic Engineering
基金
国家自然科学基金重大研发计划(编号:41890851)
国家自然科学基金项目(编号:61873280,61873281,61672033,61672248,61972416)
山东省泰山学者专项基金项目(编号:tsqn201812029)
山东省自然科学基金项目(编号:ZR2019MF012)
中央高校建设基金项目(编号:18CX02152A,19CX05003A-6)资助。
