基于串联深度神经网络的Chl-a浓度短期预报方法研究  

Study on short-term prediction method of Chl-a based on cascade DNN model

作　　者：何恩业[1] 李尚鲁 杨静[1] 季轩梁 高姗[1] 王丹[1] HE En-ye;LI Shang-lu;YANG Jing;JI Xuan-liang;GAO Shan;WANG Dan(Key Laboratory of Marine Hazards Forecasting,National Marine Environmental Forecasting Center,Ministry of Natural Resources,Beijing 100081 China;Marine Monitoring&Forecasting Center of Zhejiang Province,Hangzhou 310007 China)

机构地区：[1]国家海洋环境预报中心自然资源部海洋灾害预报技术重点实验室,北京100081 [2]浙江省海洋监测预报中心,浙江杭州310007

出　　处：《海洋预报》2021年第4期1-10,共10页Marine Forecasts

基　　金：国家重点研发计划(2016YFC1401605、2016YFC1401800);广东省海洋遥感重点实验室(中国科学院南海海洋研究所)开放课题(2017B030301005-LORS2011)。

摘　　要：以浙江海洋保护区2019年5月生态浮标监测数据为基础,对叶绿素a(Chl-a)与各理化因子进行Pearson相关性分析,发现研究海域的Chl-a与溶解氧和pH呈显著正相关(P=0.01),与硝氮和磷酸盐呈显著负相关(P=0.05)。在此基础上,建立了一种串联深度神经网络(DNN)的Chl-a短期预报模型,该模型以5层神经网络为基本单元,采用前后串联方式构建了拥有6个隐层的DNN。实验结果显示:DNN模型能够较为准确地预测Chl-a浓度短期变化趋势,24 h和48 h预报结果的RMSE分别为1.25μg/L和2.43μg/L,MAE分别为1.03μg/L和1.99μg/L,相比于浅层网络预测精度更高。Based on the monitoring data of ecobuoys in Zhejiang marine protected area in May 2019,this paper analyses the correlation between Chl-a and physicochemical factors.Statistics shows that Chl-a is positively correlated with dissolved oxygen and pH at the level of P=0.01,while it is negatively correlated with nitrate and phosphonate at the level of P=0.05.In addition,a Chl-a short-term prediction model is established,which constructs a cascade deep neural network(DNN)with 6 hidden layers in series by taking 5-layer neural network as the basic unit.The experimental results show that the cascade DNN model can accurately predict the shortterm variation trend of Chl-a with higher prediction accuracy compared to the shallow neural network.The RMSE of 24 h and 48 h prediction is 1.25μg/L and 2.43μg/L,respectively.The MAE of 24 h and 48 h prediction is 1.03μg/L and 1.99μg/L,respectively.

关 键 词：DNN 神经网络 深度学习 串联神经网络 叶绿素A 

分 类 号：X55[环境科学与工程—环境工程]