Wind turbine output power prediction by a segmented multivariate polynomial-XGBoost model

Abstract

ABSTRACT Accurate prediction of wind turbine output power is important to ensure safe, stable, and efficient wind power generation. In this study, a hybrid prediction model based on a combination of segmented multivariate polynomial and extreme gradient boosting (XGBoost) under different operating conditions is proposed. First, the wind farm supervisory control and data acquisition (SCADA) data are processed using isolated forest for outliers, in order to improve model prediction accuracy. Then, random forest is used for feature selection, where wind speed, rotor speed, pitch angle, wind direction and temperature are found to be the most relevant features related to the output power. Next, based on the analysis of the wind turbine characteristic curves, its operating state is divided into three phases: constant power (CP), constant speed (CS), and maximum power point tracking (MPPT). Different polynomials are established for each phase, the CP phase is of order 5, the CS phase is of order 7, and the MPPT phase is of order 4. Finally, the XGBoost model is trained for power prediction with polynomial new features for each phase. The segmented multivariate polynomial-XGBoost model has a mean absolute percentage error (MAPE) of 0.04, a root mean square error (RMSE) of 49, and a coefficient of determination R2 of 0.99. Compared to the single XGBoost model and the unsegmented polynomial-XGBoost model, the MAPE decreases by 42% and 33% respectively, while RMSE decreases by 50% and 12% respectively, resulting in high prediction accuracy.