Fourier Feature-Enhanced Neural Networks for Wind Turbine Power Modeling

Accurate prediction of wind turbine power output is essential for optimizing renewable energy generation, enhancing grid integration, and improving the efficiency of wind farms. However, the inherent non-linearities of wind speed–power relationships, combined with abrupt cut-in, rated, and cut-out effects, pose a significant modeling challenge. In this study, we investigate the use of artificial neural networks (ANNs) to model the power curve of a $1\mathrm{kW}$ wind turbine, using an open-access dataset of real operational measurements recorded at 10 min intervals over the course of 2011. In particular, we compare a conventional multilayer perceptron (MLP) trained on raw wind speed inputs with a Fourier-feature-encoded MLP designed to mitigate spectral bias—the tendency of neural networks to favor smooth, low-frequency patterns over sharp, high-frequency variations. Experimental results show that the Fourier-enhanced MLP substantially improves predictive performance, reducing the mean absolute error (MAE) by more than $65\%$ and achieving an $R^2$ score of $0.999$. The proposed approach demonstrates that Fourier feature encoding enables neural networks to capture sharp non-linearities in wind-turbine power curves, representing one of the first applications of this technique to wind-turbine power-curve modeling.