Machine-Learning–Assisted Probabilistic Wind Assessment at Sechura, Peru

Accurate characterization of wind resources is essential for reliable energy yield estimation and wind farm planning, particularly in regions with limited long-term measurements. This study presents a machine-learning–assisted probabilistic wind assessment in Sechura, Peru, based on multi-year hourly wind data obtained from the NASA POWER database. A representative Typical Meteorological Year (TMY) was constructed to preserve seasonal and diurnal variability while enabling standardized annual energy production (AEP) calculations. Wind speed distributions were modeled using empirical distributions, kernel density estimation (KDE), the Weibull distribution, and Gaussian mixture models (GMM). Statistical evaluation indicates that KDE and GMM reduce the annual RMSE by more than 50% compared to the Weibull model, achieving coefficients of determination above 0.98. Annual energy production is estimated at approximately 1.88 GWh, with differences below 0.3% among probabilistic models. The corresponding capacity factor is approximately 0.25 for a utility-scale wind turbine. The results demonstrate that advanced probabilistic models substantially improve wind speed representation while having a limited impact on integrated annual energy estimates, highlighting the importance of model selection for variability and seasonal analysis rather than for annual yield estimation.