| The coastal areas of Fujian Province are affected by storm surge disaster every year.It’s important and necessary to make more accurate and rapid forecasts of storm surge.Deep learning has strong nonlinear mapping capabilities and can quickly calculate the results after training,which provides new possibilities for storm surge prediction.In this paper,deep learning methods were used to explore the storm surge prediction.First of all,gridded hypothetical typhoon storm surge database established by the 2008 Fujian Storm Surge Overbank Warning Assistant Decision System was used to construct the most basic Multi-Layer Perception(MLP)single-time storm surge forecast model,and initially explore the experimental effects of using hypothetical typhoons for deep learning.Secondly,a series of storm surges were generated by a simple continental shelf model to illustrate the limitations of the MLP single-time forecast method and the effectiveness of the Long and Short Time Memory(LSTM)time series forecast.Finally,based on simulation storm surges dataset generated by the numerical model which was supervisor group’s achievement of the Seventh Five-Year Plan,the MLP single-time forecast model and the LSTM storm surge forecast models with 6h,12h,24h ahead of time were constructed for five stations along the coast of Fujian.And the historical storm surges were used for hindcast verification to illustrate the feasibility of building the LSTM water increase forecasting model method for the Fujian coast.The main conclusions are as follows:(1)For the storm surge caused by the linear typhoon path,at the convergence point of the typhoon path,the result of the MLP single-time storm surge was unstable.However,the LSTM model trained through time series had smaller forecast error and a smoother storm surge curve.(2)For the storm surge with non-linear typhoon paths generated by the simple continental shelf model,MLP single-time storm surge forecast and LSTM time-series storm surge forecast both had small error,and the prediction effect of the latter was better.(3)For the hypothetical typhoon storm surge based on the actual topography along the coast of Fujian Province,the LSTM forecast error of the five stations was less than 0.2m.LSTM had learned the relationship between the time series typhoon factors and the storm surge.The hindcast test results of the historical typhoon on the deep learning model showed that establishing a more reliable deep learning model of storm surges along the coast of Fujian would need more hypothetical typhoon samples. |
