A robust multi-location evaluation of a machine learning framework for wind power forecasting

Abstract

Wind is a highly effective and environmentally friendly renewable energy source. As the global development of wind farms continues, accurate wind power prediction has become essential for ensuring consistent energy production. Machine learning (ML) significantly advances wind power forecasting, improving the reliability and efficiency of wind power systems. This study presents an analysis of ML algorithms applied to four datasets from different geographical locations. The initial step involved the elimination of outliers using the Z-score and IQR methods to maximize the performance of the regression. The three algorithms (XGBoost, XGB, RFR, and Support Vector Regression) with RBF, polynomial, and linear kernels were trained on the same features and evaluated using R 2 , and MAE. XGBoost provided the most effective results with R 2 values 0.99 in all the locations and MAE from 11.10 to 15.94. RFR performed satisfactorily also, but the R 2 values were 0.99 in three sites, but in site 4, the ( R 2 = 0.83), and a much higher MAE of (600.81). The linear kernel was the best among the SVR models, as it attained R 2 values 0.99 and a much lower MAE on all data locations. RBF and polynomial kernels were lagging, with lower R 2 and higher MAE values. These findings highlight XGBoost and linear-kernel SVR as the best to use in wind power forecasting on diverse datasets with high accuracy levels and low error rates that can be used to improve wind farm energy production.