This paper constructs a multi-feature fusion training set containing approximately 2 million valid samples by collecting maximum wind speed forecast data from the operational numerical prediction model (CMA-GD) for the year 2023, combined with micro-topographic sample data from South China. Based on this training set, a Light-GBM regression model is developed. Experiments on intense typhoon cases show that the Light-GBM regression model can effectively correct the forecast errors of maximum wind speeds that directly output by the numerical model in most cases. The root means square errors of the maximum wind speed forecast for Typhoon “Sula” decreased from 2.58 to 1.49 within 48 h, and from 2.95 to 1.66 within 72 h. And the wind power forecasting model established using the above wind speed forecasting data achieved smaller deviations in power forecasting.