In this paper we assess a combined physical/statistical approach for the downscaling of model wind speed. The key factor in this approach is the decomposition of the total error (model - observation) into a small-scale representation mismatch (RM) and a large-scale model error (ME). The RM is caused by the difference between the grid box mean conditions of the model and the locally valid conditions. For wind speed, the RM is primarily determined by the difference in roughness between the model and the location. In the first step of the combined approach, the physical method (based on surface layer theory) adjusts the model output for the roughness characteristics at several observation sites. For these local wind estimates the RM is strongly reduced but the ME remains. To reduce this ME, the local wind estimates, together with the corresponding observations, are used in one pool to derive one linear regression equation. With local roughness length information derived from land-use maps, this regression equation can then be applied to model output to produce high-resolution wind speed fields. Using a three-year dataset, the combined approach is validated at six independent stations in The Netherlands (with different RM). In this way, we have shown that for observation-free locations the combined approach results in a significant improvement in skill compared to standard model output as well as the physical method only. The method can be optimized for special conditions, such as high wind speed cases.