Traditionally a wind turbine’s power curve is used to model the long-term energy yield of the wind turbine and afterwards assess the performance of the turbine (power curve verification). But the current power curve is typically univariate: only dependent on the wind speed and slightly adjusted for site density, turbulence intensity and wind shear. This method limits the amount of atmospheric conditions taken into account, which (could) affect turbine performance. Therefore the Power Curve Working Group introduced
the concept of an inner power curve (for common conditions) and an outer power curve (for all other conditions). While this is an improvement, it would be desirable to have not 1, not 2, but a multitude of power curves for different conditions.

This multitude of power curves, for a range of atmospheric conditions, is the goal of this thesis. Using a non-parametric machine learning method, a neural network, the feasibility of modeling a wind turbine’s power curve from historical data is studied. This resulted in a model which provides a refinement on the warranted power curve (provided by the manufacturer) for certain conditions of 7 atmospheric parameters. These parameters are wind speed, turbulence intensity, wind shear, wind veer, temperature, pressure and humidity, but could in theory be extended to more parameters.

This model resulted in three sub-models, which: 1) assess individual parameter impact/sensitivity on turbine performance, 2) provide for a company tool to continue power curve verification after a few years and 3) provide a general model to refine long-term energy yield estimations for future wind farms. Despite
the somewhat limited model and currently available datasets to train the models from, both predictive models show improved accuracy compared to the traditional power curve. Besides validation with theory, the impact assessment model allowed for investigation of the impact of certain atmospheric conditions on
the performance of certain wind turbines.