Research On Forecast Of Storm Surge In The Northern South China Sea Based On Machine Learning

Most casualties and property losses caused by tropical cyclones are caused by storm surges.With rising sea levels and increasing coastal populations,storm surges are expected to bring more and greater risks to coastal areas.Conventional numerical methods and empirical methods have mastered part of the rules of storm surges,but this is still a non-linear and non-stationary multivariable problem.It has been difficult to determine the contribution of the interaction between various storm surge components to the time and location of the peak water level of the storm surge.In the past ten years,machine learning technology has shown great flexibility in the field of earth sciences.Whether it is a pure data-driven model or a numerical model combined with machine learning technology,it provides a new perspective for forecasters to break through the forecast bottleneck.This article discusses the application and development of machine learning in storm surge forecasting,existing problems and future development directions.And built a machine learning model to forecast the storm surge in the northern coast of the South China Sea.One of the innovations of this article is to use TCRM(Tropical Cyclone Risk Model)to synthesize a synthetic typhoon database about 2000 years and to filter out specific types of extreme typhoons in a specific area.The purpose is to summarize extreme virtual typhoons to better predict the actual extreme typhoons in the future.Use the parallel ocean model GOMO(Generalized Operator Model of the Ocean)to quickly and accurately simulate the storm surge caused by 30 synthetic typhoon events selected from it.The data set obtained from this process is used to train the LSTM(Long Short Term Model)machine learning model.During the validation process,the root mean square error of each output site is less than 0.06,and the correlation coefficient is above 0.95,achieving a good fit effect.Six input parameters are used in the network:the latitude and longitude of the eye of the typhoon,the central low pressure,the moving speed of the typhoon,the maximum wind speed,and the maximum wind speed radius.A 48-hour post-report experiment and error non-parametric test of Typhoons "Hato" and "Mangkhut" were carried out on Huizhou,Chiwan,Shenzhen,Shanwei,Zhuhai,Dawanshan,Zhapo,Shuidong,Hong Kong,and Guangzhou tide gauge stations along the northern coast of the South China Sea.Under the test of the fine 15-minute forecast time step,the model gives an accurate prediction of the initial moment of the peak water level and the peak extreme value.It can be seen from the comparison of the machine learning forecast result and the GOMO model forecast result in the non-parametric test-.The error probability of-0.3 to +0.3meters is about 70%,which can be used as a surrogate model for GOMO.Because the storm surge forecasts of multiple stations are outputs at the same time,the machine learning model cannot accurately capture the water level increase or decrease process due to the different forecasting difficulties caused by the geographical location of different stations,such as the Dawanshan station in the terrain of the barrier island.Guangzhou Station,which is located in a semi-enclosed basin,is a difficult location to predict.The machine learning model can simulate the typhoon transport process with high accuracy.It can accurately predict the initial moment of the water increase,but it cannot accurately predict when the water will be reduced and the extent of the water reduction.This may be due to the limitation of the synthetic typhoon data.The results show that the forecast system established by machine learning methods has the potential to replace numerical models.The database constructed by using stochastic typhoon models combined with machine learning technology will have strong scalability and operational application value in future storm surge probability forecasts.