长短期记忆神经网络在厦门风暴潮预报中的应用  

Application of Long Short-Term Memory Neural Network in Xiamen Storm Surge Forecast

作　　者：苗庆生[1] 徐珊珊[1] 杨锦坤[1] 杨杨 刘玉龙[1] 余璇 Miao Qingsheng;Xu Shanshan;Yang Jinkun;Yang Yang;Liu Yulong;Yu Xuan(National Marine Data Information Center,Tianjing 300171,China;Shanghai University,Shanghai 200444,China)

机构地区：[1]国家海洋信息中心,天津300171 [2]上海大学,上海200444

出　　处：《中国海洋大学学报（自然科学版）》2022年第9期10-19,共10页Periodical of Ocean University of China

基　　金：国家重点研究发展计划项目(2016YFC1401900)资助。

摘　　要：利用长短期记忆神经网络(LSTM)模型强大的长短期记忆能力,建立厦门风暴潮增水预报的人工神经网络模型。利用信息流理论确定了影响增水的10种因子,分别利用不同因子组合测试了不同模型的表现,确定了表现最佳的因子组合。基于此因子组合,对比了LSTM模型和常用的BP神经网络模型、SVM模型和线性回归模型,确定了LSTM模型在风暴潮增水上的优势。基于LSTM最佳预测模型预测了1、2、3及6 h风暴潮增水值,并基于三种不同台风路径分析了模型的平均绝对误差、相关系数、有效系数和极值偏差指标。结果显示,LSTM模型在预报风暴潮短期增水有很高精度,可为防灾减灾提供辅助和参考。Using the strong long short-term memory ability of LSTM model, an artificial neural network model for forecasting the storm surge in Xiamen was established. The information flow theory was used to determine 10 factors affecting storm surge, and the performance of different models was tested by different factor combinations, and the best factor combination was determined. Based on this factor combination, the advantage of LSTM model was determined by comparing LSTM model with BP, SVM and linear regression model. Based on the LSTM optimal prediction model, the storm surge values increment values of 1, 2, 3 and 6 h were predicted. Index of mean average absolute error, correlation coefficient, effective coefficient and extreme deviation were analyzed based on three different typhoon paths. Results showed that LSTM had high accuracy in forecasting storm surge water increase, which can provide assistance and reference for disaster prevention and mitigation.

关 键 词：风暴潮 信息流 长短期记忆神经网络(LSTM) 神经网络 预报 

分 类 号：P731.23[天文地球—海洋科学]