Otimização Multiobjetivo via NSGA-II para Estimativa da Energia Potencial de Aerogeradores com Aprendizagem de Máquina

Authors

  • Durval Camargo Rangel de Souza Brito
  • Arthur Públio
  • Bruno Henrique Groenner Barbosa

Keywords:

wind turbine, NSGA-II, multi-objective optimization, machine learning, energy potential, virtual sensor, power curve

Abstract

This paper proposes a methodology for estimating wind turbine energy potential using SCADA operational data and meteorological mast (met mast) measurements, combining multi-objective evolutionary optimization with machine learning. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) simultaneously performs data source selection, meteorological feature selection, and model hyperparameter optimization for XGBoost and MLP (Multi Layer Perceptron) models, explicitly trading off predictive accuracy against model complexity. Applied to a 2.5 MW wind turbine in Bahia, Brazil, the approach yields a Pareto front with diminishing returns beyond three data sources. Selected models achieve RMSE of 0.049 (XGBoost) and 0.053 (MLP) in capacity factor, against 0.064 for a bin-average baseline, with R² of 0.984 and 0.982 versus 0.973. The optimized models estimate that approximately 6% of available energy potential was not converted over the analyzed period, demonstrating the practical applicability of the virtual sensor for energy loss quantification.

Downloads

Published

2026-10-05