基于差分回归模型和可迁移长短期记忆网络集成的三沙湾水温预测  

Water temperature prediction in the Sansha Bay based on the integration of differential regression model and transportable long short-term memory network

作　　者：赖晓倩 余镒琦 梁中耀 陈火荣 陈能汪 Lai Xiaoqian;Yu Yiqi;Liang Zhongyao;Chen Huorong;Chen Nengwang(State Key Laboratory of Marine Environmental Science,Xiamen University,Xiamen 361102,China;Fujian Provincial Key Laboratory for Coastal Ecology and Environmental Studies,Xiamen University,Xiamen 361102,China;Fishery Resources Monitoring Center of Fujian Province,Fuzhou 350003,China)

机构地区：[1]厦门大学近海海洋环境科学国家重点实验室,福建厦门361102 [2]厦门大学福建省海陆界面生态环境重点实验室,福建厦门361102 [3]福建省渔业资源监测中心,福建福州350003

出　　处：《海洋学报》2023年第4期165-178,共14页

基　　金：福建省海洋经济发展补助资金项目(ZHHY-2019-1);国家重点研发计划(2016YFC0502901)。

摘　　要：水温预测是保障近海渔业生产和环境安全的关键技术。现有的数值模型开发成本大,业务化应用不足。本文提出了一种集成差分回归(Differential Regression,DR)和可迁移长短期记忆网络(Transferable Long Short-Term Memory,TLSTM)的水温预测方法。以厦门湾(源域,数据多)和三沙湾(目标域,数据少)水温为研究对象,根据三沙湾在线监测水温和预报气温数据建立了DR模型,根据厦门湾长时间监测水温数据建立了TLSTM模型,采用变权算法将纯差分回归模型、混差分回归模型和TLSTM模型集成为三沙湾DR-TLSTM模型,对模型性能进行了评估,并与仅根据三沙湾少量监测数据建立的LSTM模型效果进行了对比。结果表明:(1)TLSTM模型的预测精度优于基于目标域少量数据建立的LSTM模型;(2)DR-TLSTM集成模型具有较高的预测精度,未来1~7 d预测的均方根误差为0.13~0.77℃,未来1~3 d预测的均方根误差小于0.4℃;(3)DR-TLSTM集成模型可有效预测水温骤升或骤降趋势,对水温突变点的预测均方根误差为0.29~1.09℃。基于本文建立的DR-TLSTM集成模型,实现了三沙湾渔业水域的水温预警预报业务化信息服务。Water temperature prediction is a key technology to ensure the production of coastal fisheries and environmental safety.The existing numerical models have high development costs with insufficient business applications.This study develops a prediction method of water temperature through integrating differential regression(DR)and transferable long short-term memory(TLSTM).Taking the water temperature of Xiamen Bay(source domain,with a large number of data)and Sansha Bay(target domain,with less data)as the research object,the DR model is established based on the data of monitoring water temperature and forecast temperature in the Sansha Bay,and the TLSTM model is established based on the long-term monitoring data of water temperature in the Xiamen Bay.The pure differential regression model,mixed differential regression model and TLSTM model are integrated into the DR-TLSTM model of Sansha Bay by using variable weight algorithm,and the performance of the model is evaluated,the results are compared with the LSTM model based on only a small amount of monitoring data in the Sansha Bay.The results show that:(1)the prediction accuracy of TLSTM model is better than that of LSTM model based on a small amount of data in the target domain;(2)the DR-TLSTM model has high prediction accuracy,and the root mean square error of prediction in the next 1−7 days is 0.13−0.77℃,and the root mean square error of prediction in the next 1−3 days is less than 0.4℃;(3)the DR-TLSTM model can effectively predict the sudden rise or fall trend of water temperature,and the root mean square error of predicting the sudden change point of water temperature is 0.29−1.09℃.Based on the DR-TLSTM model,the operational information service of water temperature early warning and forecast in the Sansha Bay is realized.

关 键 词：水温预测 回归模型 LSTM模型 迁移学习 变权集成 

分 类 号：P731.11[天文地球—海洋科学]