基于机器学习的广东海域模式风暴潮数据订正及评估OA

This study utilized the ERA5 reanalysis wind fields and sea level pressure data from the European Centre for Medium-Range Weather Forecasts(ECMWF)to drive the ADCIRC model,constructing an hourly storm surge dataset for two storm surge processes in Guangdong coastal waters in 2024.To correct the systematic errors in the storm surge simulations,by constructing a feature system comprising 17 physical features including simulated surge,tidal level,wind-induced surge,wind speed,and their derived interaction terms,a machine learning error correction model based on Random Forest regression with clear physical significance was established.Training results demonstrated the model's strong learning capability on 1 377 samples,with an R2 of 0.71 and 0.65,an RMSE of 16.7 and 21.2,and correlation coefficients of 0.86 and 0.81 for the training and test sets,respectively.Case studies showed that the model achieved improvement rates of 11.65%for the high-wind-speed case in September 2024 and 48.08%for the low-wind-speed case in November 2024,with an average improvement rate of 29.87%.The research suggests that this machine learning correction method effectively reduces systematic errors in the ADCIRC model,particularly demonstrating stronger correction capability under low-wind-speed conditions.