Uso de aprendizagem de máquina e otimização multiobjetivo para estimar energia potencial de aerogeradores

Abstract

The global energy transition has driven the growth of wind energy, particularly in Brazil, where it accounts for 12.8% of the national electricity mix. The efficient operation of wind farms requires sophisticated strategies to maximize energy extraction, given the inherent intermittency of the wind resource and the operational challenges of field environments. This work proposes an artificial-intelligence-based methodology for estimating the potential energy of wind turbines, using exclusively meteorological data from met masts and operational data from SCADA systems. The approach integrates multimodal data filtering and multi-objective optimization via NSGA-II for the simultaneous selection of data sources, meteorological variables, and machine learning model hyperparameters, exploring the trade-off between predictive accuracy and operational complexity. The methodology was validated on two GE 2.5 MW wind turbines located in Xique-Xique, Bahia, under distinct operational conditions: W301, operating under relatively free-stream wind, and W310, subject to pronounced wake effects. The dataset comprises 31 SCADA variables and meteorological measurements from 8 nearby met masts, totaling 51,564 samples collected between January 2024 and January 2025. Data filtering yielded 65% of usable samples for W301 and 63.5% for W310, removing periods of shutdowns, curtailments, and degraded operation. The NSGA-II algorithm revealed a consistent geospatial hierarchy in source selection, autonomously converging to three-met-mast configurations covering inflow conditions, intermediate flow evolution, and local conditions near the target turbine. For W301, optimized models achieved R 2 of 0.984 (XGBoost) and 0.982 (MLP), compared to 0.973 for the reference model. For W310, gains were even more expressive, with R 2 of 0.973 and 0.972 against 0.933 for the baseline, demonstrating that the methodology delivers proportionally greater value under complex operational conditions. Applying the final models to the complete historical dataset enabled estimation of energy losses between 4.7% and 6.5% of the available potential, with direct implications for maintenance planning and contractual negotiations. The proposed approach overcomes limitations of traditional methods based on IEC 61400-12-1, providing estimates of the true physical potential of the turbine without embedding operational restrictions as normal behavior.