Surface winds are characterized by large spatial and temporal variances as a result of the earth’s surface. High-resolution models are therefore of particular importance for a precise prediction. Modern regional models offer a suitably resolution, but require very high computing power and are limited in space and time. This master thesis investigates the applicability of a Deep Residual Neural Net- work in the generation of high-resolution surface wind predictions over Europe. The model was trained in a supervised manner using data from a global weather model with a horizontal resolution of ≈ 12.5 km as predictors and data from a regional model with a horizontal resolution of ≈ 2.2 km as predictands. The goal is to down- scale wind fields from a global weather model to predict important topographically induced wind phenomena on a small scale, regardless of the availability of a regional model. The performance of the model is evaluated quantitatively and qualitatively both on the training area and on unseen areas. For the first time, the possibility of adding wind gusts as an additional output parameter to the mean wind is being investigated. The importances of various meteorological input parameters is analyzed using the feature perturbation method. The results of the work show that the model can uses information from various meteorological input parameters to predict high-resolution surface wind fields. In all metrics, the results are better than in comparison with the interpolated data of the global weather model, both for the mean wind and for the wind gusts. Topographi- cally induced wind phenomena such as mountain and valley wind circulation, downs- lope winds, channeling and blocking situations are adapted to the high-resolution topography. Limitations exist for very complex phenomena that depend on a large number of meteorological parameters and are only rarely represented in the training data set. This is especially the case when the model is used in areas outside the training area.