Wind speed is involved in multiple scales physical phenomena and depends on specific features, that are not always easy to simulate numerically. Alternative solution that combines the physical advantages provided by numerical weather prediction (NWP) simulations and statistical models is investigated for wind speed forecast. Several aspects that influence the wind speed forecast error at synoptic stations in Romania were identified, such as discrepancy between model and true topography, urbanicity or distance to the Black Sea. Calibration models in the framework of Generalized Additive Models (GAM) are developed for the proposed endeavour. A set of models applied to limited area model ALARO were introduced and evaluated. Results showed improved statistical scores compared to raw ALARO output and simple regression model: a decrease of up to 23% for the RMSE score, or 94% for the bias was observed for the model which performed best in terms of annual bias and RMSE. Different impact of terms involved in the calibration model is found. Most important effects in the model are associated with wind speed observations from the 24 past hours and simulated wind speed effect in relation to altitude.

风速涉及多种尺度的物理现象，并取决于特定的特征，而这些特征并不总是容易通过数字模拟的。研究了结合数值天气预报（NWP）模拟和统计模型提供的物理优势的替代解决方案，以进行风速预测。确定了影响罗马尼亚天气站风速预报误差的几个方面，例如模型与真实地形之间的差异，城市化程度或与黑海的距离。在通用可加模型（GAM）框架内开发的校准模型可用于拟议的工作。引入并评估了一组应用于有限区域模型ALARO的模型。结果表明，与原始ALARO输出和简单的回归模型相比，统计得分有所提高：在年度偏差和RMSE方面表现最佳的模型中，观察到RMSE得分最多降低了23％，或偏差最多降低了94％。发现校准模型中涉及的术语的不同影响。该模型中最重要的影响与过去24小时的风速观测以及与高度相关的模拟风速影响有关。