长短期记忆神经网络(LSTM)对风暴潮数值模拟的优化应用

Application of Long Short-Term Memory neural network for optimization of numerical simulation results of storm surge

陈鸿生 ¹林小刚 ¹林晓珍²

扫码查看

作者信息

1. 自然资源部海洋环境探测技术与应用重点实验室,广东广州 510301;国家海洋局汕尾海洋环境监测中心站,广东汕尾 516600
2. 国家海洋局深圳海洋环境监测中心站,广东深圳 518000
折叠

摘要

利用长短期记忆神经网络和数值模式相结合的方法,设计了两套针对粤东遮浪海洋站点台风风暴潮增水的预报优化方案.与实测资料对比结果显示,长短期记忆神经网络方法可以显著改善数值模式模拟结果的准确性,最大增水和主振过程中增水后报结果的平均绝对误差、平均相对误差和平均改善幅度分别为7.1 cm、8.2%、74%和16.1 cm、34.7%、33%.进一步分析表明,利用台风信息预测数值模拟结果的订正值可以有效改善神经网络方法的不稳定性,比直接预测风暴潮增水值更加准确、可靠.

Abstract

Using a combination of Long Short-Term Memory(LSTM)neural network and numerical model,two sets of prediction schemes for typhoon storm surge at the Zhelang marine station in eastern Guangdong have been designed.Compared with the measured data,the LSTM neural network can significantly improve the accuracy of the numerical model results.The average absolute error,average relative error and average improvement amplitude of the prediction results for the maximum surge and the main oscillation process are 7.1 cm,8.2%,74%and 16.1 cm,34.7%,33%,respectively.Further analysis shows that predicting the corrected value of numerical results using typhoon information can effectively limit the instability of neural network,which is more accurate and reliable in comparison with predicting the storm surge level directly.

关键词

长短期记忆/神经网络/台风风暴潮/数值模拟

Key words

Long Short-Term Memory/neural network/typhoon storm surge/numerical simulation

引用本文复制引用

基金项目

自然资源部海洋环境探测技术与应用重点实验室自主设立课题(MESTA-2022-D008)

出版年

2024

海洋预报

国家海洋环境预报中心

海洋预报

CSTPCDCSCD北大核心

影响因子：0.37

ISSN：1003-0239

参考文献量30

段落导航