In this work the uncertainty of wind power predictions is investigated with a special focus on the important role of the nonlinear power curve. Based on numerical predictions and measured data from six onshore wind farms the overall prediction accuracy is assessed and it is shown that due to the power curve the relative forecast error increases by a factor of 1.8-.2.6 compared to the wind speed forecast. This factor can be considered as an effective nonlinearity factor. A decomposition of the commonly known root mean square error is beneficially used to distinguish different error sources related to either on-site conditions or global properties of the numerical weather prediction system. The statistical distribution of the wind speed prediction error is found to be Gaussian in contrast to the the one of power prediction error. Using the power curve and conditional probability density functions of the wind speed the unsymmetric distribution of the power prediction error can be explained and modeled such that it can be estimated even if no measurement data is available.