基于LSTM网络的海水温度剖面预报研究
Research on ocean temperature profile forecasts based on Long Short-Term Memory neural network
范培勤 1过武宏 1唐帅 1张驰 1曲泓玥2
作者信息
- 1. 海军潜艇学院,山东青岛 266199
- 2. 青岛海洋科学与技术试点国家实验室,山东青岛 266000
- 折叠
摘要
基于海水温度历史观测和海洋模式数值预报数据,利用长短期记忆(LSTM)神经网络开展了海水温度剖面短时预报方法研究.以南海17°46.91′N,112°03.24′E处海水温剖面预报为研究对象,利用模式预报数据和观测数据,构建了观测、预报、观测和预报混合3个样本数据集.基于LSTM神经网络模型,建立了由编码器-解码器组成的多对多海水温度剖面序列预报模型,并开展了模型的训练和性能验证分析工作.结果表明:该模型具有较高的预报精度,对小样本问题的处理具有良好的稳定性;采用深度分层预报,可有效提高模型预报精度,改善模型泛化能力;与只使用观测数据集的预报相比,混合样本集预报误差下降明显,为基于小样本观测数据的海洋环境要素预报研究提供了一种思路.
Abstract
Based on ocean temperature historical observations and ocean model data,the short-term forecasting method for ocean temperature profile is studied using Long Short-Term Memory(LSTM)neural network.For the ocean temperature profile forecasts at(17°46.91'N,112°03.24'E)in the South China Sea,three sample sets of observation,forecasts,mixed observation-forecasts are constructed using the observational and forecasting data.Based on the LSTM neural network,a many-to-many ocean temperature profile forecasting model composed of encoder and decoder is established,and model training and verification are carried out.The results show that the model has a high forecasting accuracy and a good stability for processing small sample problems.Utilizing deep layered forecasting method can effectively improve the forecasting accuracy and generalization ability.The forecasting error of mixed sample set decreases significantly in comparison with that of observation sample set,which provides an idea for marine environment forecasts of small sample problems.
关键词
深度学习/长短期记忆神经网络/海水温度/短时预报Key words
deep learning/LSTM/ocean temperature/short-term prediction引用本文复制引用
基金项目
基础计划加强重点基础研究项目(2020-JCJQ-ZD-144-00)
基础计划加强重点基础研究项目(2020-JCJQ-ZD-144-03)
出版年
2024
国家科技期刊平台
NSTL
万方数据