Long term wind energy forecasting using machine learning techniques

Ensuring the reliability of wind energy as a dependable source requires overcoming challenges posed by the inherent volatility and stochastic nature of wind patterns. Long-term forecasting provides strategic advantages in managing energy generation projects, enabling the development of effective portfolio management strategies. The primary objective of this study was the development of forecasting methods to support strategic decision-making within the scope of wind energy operations, specifically targeting the Pindaí Wind Complex and its commercial dispatch. The study integrated Big Data analytics, data engineering, and computational techniques through the application of machine learning algorithms: including eXtreme Gradient Boosting, Multilayer Perceptron, Support Vector Regression, Ridge Regression, and Random Forests, aiming to generate forward-looking projections of the complex’s energy production for the year 2023. To this end, five supervised machine learning techniques were modeled and implemented. These techniques were grounded in their respective mathematical and structural formulations, and the empirical foundation for modeling was provided by historical power generation data from the Pindaí Wind Complex, combined with high-resolution realized and forecasted meteorological data retrieved via the Open-Meteo API. The models are trained using historical monthly generation data from the Pindaí Wind Complex, which has an installed capacity of 79.9 MW and is located in the northeastern region of Brazil, along with meteorological data from reanalysis models, such as air temperature, relative humidity, precipitation, surface pressure, wind speed at 10 m, wind speed at 100 m, and wind gusts. These methodologies are applied to forecast monthly wind generation for the year 2023, and the outputs are systematically compared using evaluation metrics to determine the most suitable modeling approach. The results highlight the superiority of the Multilayer Perceptron, Support Vector Regression, and eXtreme Gradient Boosting models, which achieved Kling-Gupta Efficiency (KGE) of 0.89, 0.89, and 0.90, mean absolute scaled error (MASE) of 0.29, 0.31, and 0.18, root mean square errors (RMSE) of 0.56, 0.59, and 0.35, and mean absolute errors (MAE) of 0.48, 0.52, and 0.29, respectively.