基于浮标44013数据的有效波高预测研究  

The study of significant wave height forecasting based on six classical machine learning models

作　　者：邓泽贵 李醒飞[1,2,3] 杨少波 DENG Ze-gui;LI Xing-fei;YANG Shao-bo(State Key Laboratory of Precision Measuring Technology and Instruments,Tianjin University,Tianjin 300072,China;Pilot National Laboratory for Marine Science and Technology(Qingdao),Qingdao 266003,China;Qingdao Institute for Ocean Engineering of Tianjin University,Qingdao 266200,China)

机构地区：[1]天津大学精密测量技术与仪器国家重点实验室,天津300072 [2]青岛海洋科学与技术试点国家实验室,山东青岛266003 [3]天津大学青岛海洋技术研究院,山东青岛266200

出　　处：《海洋湖沼通报》2022年第1期31-41,共11页Transactions of Oceanology and Limnology

基　　金：天津市重点研发计划科技支撑重点项目(18YFZCSF00620);天津市重点研发计划院市合作项目(18YFYSZC00120)~~。

摘　　要：随着海事活动和海洋资源开发的不断增加,精准的海浪预报变得越来越重要。其中有效波高(SWH,Significant Wave Height)作为海浪的主要参数之一,精准的有效波高预测不仅可以给各项海事活动提供必要的海洋气象预报,还可以为波浪能高效利用提供重要的参考依据。由于受到复杂海洋环境和自然混沌行为的影响,有效波高的精准预报存在诸多困难。近年来,机器学习促进了许多预测任务的发展,因此本文研究了线性回归(LR,Linear Regression)、支持向量回归(SVR,Support Vector Regression)、神经网络(ANN,Artificial Neural Networks)、K近邻(KNN,K-Nearest Neighbor)、决策树(DT,Decision Tree)、随机森林(RF,Random Forest)六种经典机器学习模型预测有效波高的性能。实验结果表明:(1)在预测未来一小时有效波高时,线性回归、支持向量回归、神经网络都能给出较好的预测结果;而K近邻、决策树、随机森林表现较差。(2)在预测未来一天日平均和日最大有效波高时,支持向量机、神经网络、随机森林三种模型预测指标相近且优于其余三种模型;K近邻的预测结果与观测值之间的偏差仍然最大。(3)线性回归和K近邻算法速度最快,而神经网络耗时最长。相比于使用单一特征,使用多种特征通常能够提升模型的预测性能,但模型训练时间也随之增加。with the increasing of marine activities and exploitation of marine resources,the wave forecasting becomes more and more important.Accurate prediction of significant wave height(SWH)which is one of the most important parameters of wave can not only provide essential sea condition for various marine activities,but also provide important reference for the effective use of wave energy.Due to the influence of complex ocean environment and natural chaos,it is very difficult to forecast SWH accurately.In recent years,the rapid development has promoted many prediction tasks.Therefore,this paper aim to investigate the performance of SWH forecasting of six classical machine learning models which are Linear Regression(LR),Support Vector Regression(SVR),Artificial Neural Networks(ANN),K-Nearest Neighbors(KNN),Decision Tree(DT)and Random Forest(RF).The experimental results show that:(1)for the one-hour ahead SWH forecasting,LR,SVR and ANN could give better prediction results.By contrast,KNN,DT and RF performed worst.(2)For one-day ahead daily mean and daily maximum SWH forecasting,the prediction metrics of SVR,ANN and RF were close,which were more excellent than other’.The deviation between the prediction results of k-nearest-neighbor and the observation values is still largest.(3)The training time of LR and KNN was obviously shortest,but the training time of ANN was longest.Using multiple features could improve the performance of the machine learning models comparing with using single feature,but the time consumption of the models will be more expensive.

关 键 词：有效波高 机器学习 预测评估 

分 类 号：P731.33[天文地球—海洋科学]