Abstract: Downscaling of Numerical Weather Prediction (NWP) is a critical method to obtain surface wind speed forecasts at 100-meter-scale resolution. There are two main types of downscaling: dynamical, which can capture fine scale dynamics of surface wind speeds, and statistical, which compute efficiently with much lower computational cost. But in a bay with complex terrain, neither of them can satisfy the requirements for operational implementation, because of the expensive and inefficient computation of dynamical downscaling, and a lack of 100-meter-scale-resolution data on which to base statistical downscaling. A 100-meter-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-meter scale is obtained from a Large-Eddy Simulation (LES) at the same horizontal resolution nesting with a coarse resolution WRF model (known as WRF-LES). Subsequently, by using a machine learning method, a statistical downscaling model is developed relating an operational NWP input and the dynamically downscaled output. In this paper, the hybrid dynamical-statistical downscaling method is applied to the surface wind speed simulations of the CMA-GD model with a horizontal resolution of 0.03° in some strong wind cases during winters from 2019 to 2023 over the Meizhou Bay of Fujian Province. In the application, the domain of LES with 111-meter resolution is divided by 0.03°, and a Random Forest algorithm (RF) is used to build the statistical downscaling model for every subarea, based on the training data from 2019 to 2022. Evaluations with respect to the surface wind speed observations from the Automated Weather Stations (AWS) over the Meizhou Bay show that the WRF-LES dynamically downscaled data with 111-metre resolution can capture the spatiotemporal intermittency in surface wind speed pulsation better and have less root-mean-square errors (RMSEs) than the CMA-GD simulations. It also shows that as for the dynamical-statistical downscaled surface wind speeds of the CMA-GD simulations over testing set of 2023, the spatial distribution is consistent with the WRF-LES data, and the temporal variation is consistent with the AWS observations as well. In addition, their RMSEs are significantly less than that of bilinear interpolation downscaling from the CMA-GD simulations, with RMSEs in most areas smaller than 2.5 m/s. In conclusion, the proposed dynamical-statistical framework integrating LES and machine learning can effectively downscale kilometer-scale wind speeds to 100-meter scale resolution in complex bays, which may improve the refined weather forecasts in intricate coastal areas.