基于多模型平差接力的长时序风暴增水预测方法

Prediction method of long-term storm surge based on multi-model gradual adjustment

利用神经网络模型预测未来风暴增水时,预测精度会随预测时序的延伸不断降低。基于长短期记忆(Long Short-Term Memory)神经网络模型,以风速、风向、气压和前时序的风暴增水数据作为模型输入,利用多个模型接力预测风暴增水时间序列,减小误差随模型的迭代累积,建立基于多模型平差接力的长时序风暴增水预测方法。对比不同地点和不同台风下的风暴增水预测分析结果,模型在渤黄海北部区域每月的均方根误差为4~7 cm,在黄海中部区域可控制在10 cm以内,能够较准确预测未来24 h的风暴增水。

When predicting storm surge using classical neural network models,the prediction accuracy will continue to decrease with the extension of the time series.To alleviate this problem,we proposed a method of long-term storm surge prediction based on multi-model gradual adjustment.The method is based on the Long Short-Term Memory neural network model,and integrates multiple models to predict the storm surge time series.Data from previous time periods were fed into the model,such as wind speed,wind direction,air pressure,and storm surge data,and the errors accumulated with model iterations were considered.According to the forecast analysis of storm surge in different locations and under different typhoons,the monthly root mean square error of the prediction model is between 4 and 7 cm in the northern part of the Bohai sea and Yellow Sea,and is less than 10 cm in the central part of the Yellow Sea,which indicate the model can predict storm surge in the next 24 hours accurately.