Enhanced wind power prediction using rotor equivalent wind speed and machine learning models

Accurate wind-power prediction is essential for optimizing wind-farm operation and maintaining grid stability. Traditional approaches based solely on hub-height wind speed often yield substantial errors for large-diameter turbines due to vertical wind-speed variability. This study addresses this limitation by employing the Rotor Equivalent Wind Speed (REWS) method, which accounts for wind-speed distribution across the rotor-swept area to enhance power estimation accuracy. Eleven months of wind data were obtained from the National Renewable Energy Laboratory's National Solar Radiation Database (NREL-NSRDB) for the Goldwind 2S MW turbine located in Lewes, Delaware, USA. The REWS was computed by dividing the rotor into seven annular segments, and the Weibull distribution was used to characterize the wind profile and estimate power generation. Multiple machine-learning regression models were trained and tested in MATLAB using a 90% – 10% data split to ensure seasonal representation. The models included Decision Trees, Support Vector Machines (SVM), Gaussian Process Regression (GPR), Ensemble methods (Bagged and Boosted Trees), and Neural Networks. Hyperparameters were optimized through five-fold cross-validation, with the Fine Tree model achieving the lowest prediction error (RMSE = 1.19 ms⁻¹, R² = 0.83). The predicted REWS values showed an average deviation of 14% from the measured data, resulting in an 18% difference in the estimated monthly energy output. These results demonstrate that combining REWS with interpretable machine learning models provides a reliable, computationally efficient framework for wind-power forecasting, particularly when high-resolution or long-term datasets are unavailable. The findings also highlight the potential of REWS-based modeling for future applications to larger turbines and complex terrain conditions.