Wind Power Generation in Peru: Exploratory Analysis and Prediction Using Outlier Detection and Supervised Learning Algorithms

This study presents a comprehensive analysis of wind power generation forecasting in Peru using machine learning algorithms and outlier detection techniques. The research employed Random Forest Regression, Gradient Boosting Regression, and Support Vector Regression (SVR) to predict wind power generation across 13 companies using a two-year dataset (May 2023 - April 2025). A 7-day sliding window approach was implemented to forecast daily generation, with models trained on temporal features, lagged values, and rolling statistics. Exploratory data analysis revealed significant variability in generation capacity across companies, with monthly generation ranging from 359 GWh to 825 GWh, showing 42.0% growth between study years. The capacity factor decreased from 33.0% in Year 1 to 22.5% in Year 2, indicating rapid capacity expansion. Random Forest demonstrated superior performance with R2 = 0.438, MAE = 3,700.43 MWh, and RMSE = 4,734.9 MWh, followed by SVR (R² = 0.423). Company-specific analysis showed heterogeneous performance, with ENGIE facilities achieving the highest predictability (R² > 0.45), while some companies exhibited negative R2 values. Outlier detection using IQR methods improved model stability by reducing RMSE = 226.63 to 198.51 MWh, though slightly decreasing R2 = 0.418 to 0.327.