基于LSTM神经网络方法的粤东近岸海温预报  

Sea temperature forecasting based on LSTM neural network along the coast of eastern Guangdong

作　　者：林小刚 王兆毅[2] 李竞时 庞运禧 罗荣真[1] 闫桐[3] LIN Xiaogang;WANG Zhaoyi;LI Jingshi;PANG Yunxi;LUO Rongzhen;YAN Tong(Shanwei Marine Environmental Monitoring Center,State Oceanic Administrator,Shanwei 516600,China;National Marine Environmental Forecasting Center,Beijing 100086,China;State Key Laboratory of Tropical Oceanography(South China Sea Institute of Oceanology,Chinese Academy of Science),Guangzhou 510301,China)

机构地区：[1]国家海洋局汕尾海洋环境监测中心站,广东汕尾516600 [2]国家海洋环境预报中心,北京100081 [3]热带海洋环境国家重点实验室(中国科学院南海海洋研究所),广东广州510301

出　　处：《海洋预报》2022年第5期27-36,共10页Marine Forecasts

基　　金：自然资源部南海局海洋科学技术局长基金(180222)。

摘　　要：利用粤东沿岸各海洋站点的海温观测数据以及区域大气模式的气象预报数据,基于长短期记忆的神经网络方法,通过分析筛选训练数据,建立了高效的粤东近岸24 h海温预报方法。与观测资料进行对比,该方法在粤东近岸海域24 h海温预报的均方根误差和平均绝对误差分别为0.45℃和0.32℃,在珠江口沿岸的误差更小。进一步分析表明,气象要素与海温变化值的相关性比与高通滤波后的海温相关性更好。对海温变化值进行预报,而后叠加海温初始值,可以得到更加准确且稳定的预报结果。The sea temperature observation data of marine stations along the coast of eastern Guangdong and meteorological forecast data of a regional atmospheric model is analyzed, selected and trained based on the Long Short-Term Memory neural network, and an efficient 24-hour nearshore sea temperature forecasting method along the coast of eastern Guangdong is established in this paper. Compared with observation, the root mean square error and the mean absolute error of 24-hour sea temperature forecast in the coastal waters along eastern Guangdong are 0.45 ℃ and 0.33 ℃, respectively, and the error is even smaller in the Pearl River Estuary. Further analysis shows that the correlation between meteorological factors and the variation of sea temperature is more significant than that with the sea temperature after high pass filtering. Therefore, more accurate and stable forecasting results could be achieved when the sea temperature variation is forecasted and superimposed with its initial value thereafter.

关 键 词：神经网络 长短期记忆网络 粤东近岸 海温预报 

分 类 号：P731.31[天文地球—海洋科学]