Numerical atmospheric models are widely used as a meteorological data source when planning the locations of new wind farms. However, before relying on model output for decision-making, it must be verified against observations. Due to commercial restrictions on the availability of observation data, previous studies on atmospheric model validation for wind energy applications are often limited to a single model or a small geographical region. This work performs a large-scale validation of modelled winds at wind turbine heights from seven model datasets against data from more than 500 observation campaigns across Europe. Principal component analysis is used to identify spatial, diurnal, and seasonal patterns of wind speed and direction biases. The results of the analysis show that all seven models exhibit similar spatial and temporal patterns of wind speed bias. Models generally show a more positive wind-speed bias in the Central European Plain and a more negative bias in mountainous regions, namely Southern Europe and the Scandinavian Mountains. Moreover, the temporal patterns of biases also differ between these regions, and wind direction bias shows the same temporal and spatial patterns as the wind speed bias. We show that these wind speed and direction biases can be explained by differences in terrain height between the models and the real world. The magnitude of the wind speed bias ranges from 0.1 to 0.9 ms^{<span>&minus;</span>1} per 100 m of elevation difference, depending on the season and time of day. Two WRF model simulations with different terrain source data are performed, and the modelled winds are compared to provide more robust support for the hypothesis. The results of this work suggest that improving terrain representation in the models can help improve their performance.&nbsp;