Predicting Wind Turbine Power Output Based on XGBoost

Chang Liu, Jiandun Li, Haitao Wang

Abstract: The prediction of wind power is crucial to ensuring the reliability and economic efficiency of wind power generation systems, as well as to maintaining balance and efficient operation of power systems. However, due to the non-stationary and chaotic nature of wind speeds, predicting wind power is a challenging task. Recently, various solutions have been proposed, e.g., SARIMA-based models and BP neural network-based models, which have successfully predicted periodicity and short-term wind power generation, but their performances are limited. In this paper, we select the top-eight most significant attributes from a public wind power dataset, i.e., wind direction, hub temperature, bearing shaft temperature, gearbox bearing temperature, gearbox oil temperature, rotor speed, reactive power and active power. We then train eight supervised machine learning models, i.e., Linear Regression, Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), K-Nearest Neighbors (KNN), Ridge and Lasso Regression, Decision Tree, and Gradient Boosting, to predict the wind power output of the next 70 days. Experimental results showed that the XGBoost model outperforms others (R-squared score = 0.96, accuracy = 95.39%, MAE = 39.43, and cross-validation score = 0.98). Compared to the state-of-the-art performance achieved by the Random Forest model, XGBoost has improved the prediction accuracy by 4.69% points.