Abstract: Abstracts Downscaling of numerical weather prediction (NWP) is a critical method for obtaining surface wind speed forecasts at 100 m resolution. There are two main types of downscaling: dynamical, which can capture the fine-scale dynamics of surface wind speeds, and statistical, which can be computed efficiently at much lower computational costs. However, in a bay with complex terrain, neither can meet the requirements for operational implementation due to the high cost and inefficiency of dynamical downscaling and the lack of 100 m-resolution data to support statistical downscaling. A 100 m-scale hybrid dynamical-statistical downscaling technology, which combines the benefits of dynamical and statistical downscaling, is proposed in this study to address the issue. First, a dynamical downscaling dataset of surface wind speed simulations with a horizontal resolution of 100 m is obtained from a large-eddy simulation (LES) at the same resolution. This LES is nested with a coarse resolution weather research and forecasting (WRF) model (known as WRF-LES). Subsequently, using a machine learning method, a statistical downscaling model is developed that relates operational NWP input to dynamically downscaled output. Hybrid dynamical-statistical downscaling is applied to simulate surface wind speed in the China meteorological administration Guangdong (CMA-GD) model at a horizontal resolution of 0.03°, focusing on strong winter wind events from 2019 to 2023 over Meizhou Bay, Fujian Province. In the application, the domain of LES at 111 m resolution is divided into 0.03° subareas, and a Random Forest algorithm is used to build a statistical downscaling model for each subarea, trained on data from 2019 to 2022. Evaluations with respect to the surface wind speed observations from automated weather stations (AWS) over the Meizhou Bay show that the WRF-LES dynamically downscaled data with 111-m resolution can better capture the spatiotemporal intermittency in surface wind speed pulsation and have less root-mean-square errors (RMSEs) than the CMA-GD simulations. Moreover, the surface wind speeds from the CMA-GD simulations, derived by dynamical-statistical downscaling, exhibit a spatial distribution consistent with the WRF-LES data set for the 2023 test set. The temporal variations of these speeds are consistent with the AWS observations as well. In addition, RMSEs are significantly lower than those of bilinear interpolation downscaling from the CMA-GD simulations, with RMSEs in most regions being less than 2.5 m s⁻¹. In conclusion, the proposed dynamical-statistical framework, which integrates LES and machine learning, can effectively downscale kilometer-scale wind speeds to 100 m resolutions in complex bays. This framework may improve the accuracy of weather forecasts in complex coastal areas.