深度学习提高有效波高预报精度的实验分析  

Experimental Analysis on Improving Prediction Accuracy of Significant Wave Height by Deep Learning

作　　者：叶佳承 于华明 葛晶晶[3] 李松霖 万江岳 Ye Jiacheng;Yu Huaming;Ge Jingjing;Li Songlin;Wang Jiangyue(College of Oceanic and Atmospheric Sciences,Ocean University of China,Qingdao 266100,China;Sanya Oceanographic Institution,Ocean University of China,Sanya 572025,China;Unit 31110 of the Chinese People′s Liberation Army)

机构地区：[1]中国海洋大学海洋与大气学院,山东青岛266100 [2]中国海洋大学三亚海洋研究院,海南三亚572025 [3]解放军31110部队

出　　处：《中国海洋大学学报（自然科学版）》2024年第3期1-8,共8页Periodical of Ocean University of China

基　　金：国家重点研究发展计划项目(2018YFB1502801,2018YFB1502802);山东省重点研究发展计划项目(2019JZZY020713)资助。

摘　　要：海浪预报对于指导海上生产活动具有重要作用,随着观测技术和人工智能技术的快速发展,海浪观测数据大量积累,采用大数据结合深度学习的方法是当今数字时代提高海浪预报精度的有效手段。本研究采用深度学习中的长短时记忆网络(Long short-term memory, LSTM)模型建立海浪预报订正模型,基于董家口港区现场海浪观测数据,对海浪数值模式的预报结果开展订正实验,使用均方根误差(RMSE)与相关系数(COR)作为订正效果的评价指标。实验结果表明:将平均周期与有效波高作为输入变量进行模式训练,得到的LSTM模型可令海浪预报的RMSE由0.26 m降至0.14 m,预报精度提高了46%;将平均周期纳入输入变量对LSTM订正模型的订正效果有明显改善,主要体现在对相关系数的改善,可以使COR提高10%;与传统订正方案相比,LSTM能够通过加入相关变量提高订正能力。本研究成果可广泛应用于其他海域的海浪数值预报结果订正,对提高海浪预报精度具有重要现实意义。Wave prediction plays an important role in guiding offshore production.With the rapid development of observation technology and artificial intelligence technology,a plenty of wave observation data is accumulated.Deep learning is an effective means to improve the accuracy of wave prediction in the digital era.This paper applies the Long short-term Memory network(LSTM)deep learning to establish the wave prediction correction model.Based on the field wave observation data in Dongjiakou Port area,the paper corrects the wave numerical model prediction results and adopts Root Mean Square Error(RMSE)and Correlation coefficient(COR)as evaluating indicator of the correction effect.The experimental results show that:by using average period and significant wave height as input variables to train model,The LSTM model can reduce the RMSE from 0.26 m to 0.14 m and improve the prediction accuracy by 46%.Incorporating the average period into the input variable can improve the correction effect of the LSTM correction model,particularly in the improvement of COR,which can improve COR by 10%.Compared with the traditional correction scheme,LSTM can improve the correction ability by adding relevant variables,which has stronger potential.The results of this study can be widely applied to the correction of numerical wave prediction results in other sea areas.It has important and practical significance for using wave observation data and improving the accuracy of wave prediction.

关 键 词：海浪预报 深度学习 观测数据 平均周期 

分 类 号：P731.22[天文地球—海洋科学]