Enhancing Wind Power Prediction through Machine Learning

Wind energy data exhibits significant volatility, randomness, and irregularity, compounded by geographical and meteorological influences, which add complexity to data management. Consequently, predicting wind power effectively and accurately is a persistent challenge, crucial for minimizing disruptions in smart grid integration and ensuring operational safety and reliability. This study focuses on enhancing prediction accuracy and evaluating whether data measured below the hub height can sufficiently predict wind power generation at the hub, potentially reducing measurement costs. Ten features comprising wind and meteorological parameters are analyzed at various heights (i.e., 10, 30, 50 m) and the hub level. Two-year wind power generation data from the SCADA system corresponding to two selected wind farms in China are used for training, validation, and testing the four machine learning models. Wind speed, direction, and ambient temperature emerged as significant features, whereas ambient pressure and relative humidity had minimal impact on prediction accuracy. Results indicated that measurements at 30 meters produced wind power generation predictions comparable to those at the hub height. Among the models, the random forest algorithm provided the most accurate predictions for both low and high wind speeds, with R2 values between 0.9708 and 0.9942 and RMSE values from 0.0234 to 0.0397. In contrast, the decision tree model showed the poorest performance across all wind speed ranges, with R2 values as low as 0.7438 and RMSE values up to 0.1251.