A Forecasting Ensemble Model Based on Transfer Learning for Energy Supplier From Profile Geometries of Wind Turbines
Abstract
Small-scale wind turbines (SSWTs) are promising solutions for decentralized electrification in urban and remote settings, but their performance depends strongly on blade aerodynamics and on reliable power prediction under limited data availability. This study combines Blade Element Momentum (BEM) analysis with supervised machine learning to predict turbine voltage from geometric and operational parameters, while supporting blade design exploration through chord and twist distributions. A compact dataset from wind-tunnel experiments was enriched through data augmentation and complementary laboratory measurements, resulting in 24 samples. Missing values were handled using KNN imputation. Several regression models were benchmarked under repeated cross-validation with hyperparameter optimization, and assessed using MAE, RMSE, and R². Among the evaluated methods, KNN achieved the best generalization on the test set, with RMSE = 3.35 and R² = 0.69, outperforming tree-based ensembles in the low-data regime. These results indicate that locally adaptive, nonparametric models can better capture the feature–response relationship when observations are scarce. Overall, the proposed lightweight and reproducible pipeline demonstrates that BEM-informed features combined with carefully tuned machine learning can provide accurate voltage forecasts for SSWTs.
Keywords
Renewable Energy, Wind Energy, Forecasting, Machine Learning, Ensemble Models