Machine-learning-based tropical cyclone wind field model incorporating multiple meteorological parameters

Abstract

Multiple hazards caused by tropical cyclones (TCs), such as heavy rains and strong winds, result in substantial property losses and casualties worldwide each year. TC wind field models, describing the development of the wind hazard, are key within early warning realizations and associated risk assessments. Different to conventional parametric, analytical or meteorological numerical models, this study aims to develop a machine-learning-based approach for modeling TC wind fields by incorporating multiple meteorological parameters. The wind field model considers linear and nonlinear modeling respectively, where the input data includes various meteorological parameters such as surface pressure gradient (SPG), geopotential (GEO), boundary layer height (BLH), and forecast surface roughness (FSR). The output data is the TC wind field data of the Regional and Mesoscale Meteorology Branch (RAMMB) extracted by image recognition method, and assimilated with the wind field from the fifth generation of the European Center for Medium-Range Weather Forecasts (ECMWF) atmospheric reanalysis dataset ERA5. In the linear model, various combinations of parameters are considered, yet always yielding unsatisfactory results. The best results in the linear model were obtained using all four parameter combinations, where the root mean square error (RMSE) was 2.60 m/s and the coefficient of determination $R^2$ value was 0.44. To increase performance, three nonlinear machine learning methods—Fully Connected Deep Neural Networks (FC-DNN), Convolutional Neural Networks (CNN), and Transformer—are introduced to the training process. Comparing the wind field continuity, RMSE and $R^2$ between the three models, it is found that the Transformer outperforms all other models, with $R^2$ value of 0.877 and an RMSE of 2.23. As a final step, the trained Transformer model was used to predict the evolution of wind speed of the Typhoon Lekima (1909), in what could serve as effective model validation.