摘要
为改善台风预报精度,基于实时滚动修正预报思路,利用卷积神经网络嵌套长短期记忆神经网络(CNN-LSTM)和误差校正(EC)技术,搭建了珠江河口台风实时预报模型。研究结果表明:“滚动预报”比单次预报有更好的路径和强度预报效果,随着模型滚动时间的延长,预报整体精度有逐渐改善的趋势。路径预报结果的均方根误差比单次预报减小了25.67%,强度预报结果的平均绝对误差比单次预报减小了65.04%;考虑误差校正的CNN-LSTM-EC的路径、强度“滚动预报”效果均优于CNN-LSTM,前者的路径预报误差较后者减小了22.57%,强度预报误差减小2.5%。
In order to improve the accuracy of typhoon forecasting,this paper introduces a real-time rolling corrected typhoon forecasting model in the Pearl River Estuary utilizing Convolutional Neural Network Long Short-Term Memory(CNN-LSTM)neural network and Error Correction(EC)method.The results show that the rolling forecasts have better performances on typhoon's track and intensity than the single-time forecasts.The overall accuracy of the rolling forecasts increases gradually along with the prolong of the rolling time of the model.In comparison with the single-time forecasts,the root mean squared error of typhoon's track rolling forecasts decreases by 25.67%and the mean absolute error of typhoon's intensity rolling forecasts decreases by 65.04%.The real-time rolling corrected forecasts of typhoon's track and intensity based on CNN-LSTM-EC are better than those based on CNN-LSTM.Compared with the latter,the forecasting error of the former decreases by 22.57%on the typhoon's track and by 2.5%on the typhoon's intensity.
作者
邓志弘
刘丙军
张卡
胡仕焜
曾慧
张明珠
李丹
DENG Zhihong;LIU Bingjun;ZHANG Ka;HU Shikun;ZENG Hui;ZHANG Mingzhu;LI Dan(School of Civil Engineering,Sun Yat-sen University,Zhuhai 519085,China;Water Resources and Environment Research Center of Sun Yat-sen University,Guangzhou 510275,China;Guangzhou Hydraulic Research Institute,Guangzhou 510220,China)
出处
《海洋预报》
CSCD
北大核心
2024年第1期94-103,共10页
Marine Forecasts
基金
广州市水务科技项目(GZSWKJ-2020-2)
国家自然科学基金资助项目(52179029、51879289)。
关键词
实时滚动预报
台风
珠江河口
深度学习
误差校正
real-time rolling forecast
typhoon
Pearl River estuary
deep learning
error correction