Aplicação de Machine Learning para a previsão da velocidade do vento em diferentes localidades para um horizonte de previsão de 24 horas

Abstract

In line with the accelerated growth of renewable sources in the Brazilian and global energymatrix, especially wind energy, the precise forecasting of wind speed has become essential for theplanning and efficient operation of the national electrical system. In this context, the application ofmachine learning algorithms emerges as an alternative to complement and correct the existinglimitations in traditional numerical weather prediction models. Given this context, the generalobjective of this work is to achieve an improvement in wind speed forecasting by using a machinelearning algorithm for four distinct locations: Senhor do Bonfim (BA), Conde (BA), MossorÛ(RN), and Rio Grande (RS). The forecast data used are from the global Numerical WeatherPrediction models ARPEGE and GFS. With approximately six months of data, a preliminarycomparative analysis of measured and predicted data was conducted, along with a variableanalysis using feature engineering methods, hyperparameter optimization to find the bestcombination, and the training of the machine learning model for ARPEGE and GFS data in orderto compare them with their original forecasts and with the Persistence method. The machinelearning algorithm used was XGBoost, and with the applied methodology, significant results wereachieved using metrics such as RMSE and R2 for evaluation. The best model performances wereobtained in MossorÛ, achieving an improvement of 73.13% for XGBoost ARPEGE and 45.45%for XGBoost GFS in relation to their original forecasts. Reaching a coefficient of determination(R2) above 0.55, both algorithms registered a reduction of more than 40% in the average RMSEwhen compared to the Persistence method in the same period.