Avaliação de métodos de ensemble dinâmico em aprendizado de máquina para previsões de irradiância solar e velocidade do vento

Abstract

In recent years, with increased interest in decarbonizing energy-generating processes through clean alternatives and the consequent search for new energy sources there has been a growing exploitation of solar irradiance and wind speed resources, and good predictability performance is essential given the natural intermittency of these resources. The aim of this thesis is to implement a dynamic ensemble architecture for time horizons ranging from 10 to 60 minutes for data at 10-minute time intervals. Global horizontal irradiance (GHI) and wind speed were calculated using four independent prediction models (Random forest, k-nearest neighbors, support vector regression and elastic net) to compare their performance with two dynamic ensemble methods, windowing and arbitrating, which combine the results of independent models. The autonomous models and dynamic ensemble methods were evaluated using the RMSE, MAE, R² and MAPE error metrics. The results of this work showed that the dynamic ensemble windowing method was the best performing method when compared to the other models evaluated. For both wind speed and solar irradiance forecasts, the windowing model achieved the best error values in terms of RMSE for all the forecast horizons evaluated. Using this approach, the gain in wind speed forecasting was 0.56% when compared to the second best forecasting model, while the gain in GHI forecasting was 1.96%, considering the RMSE metric. The development of an ensemble model capable of providing accurate estimates can be implemented in real-time forecasting applications, helping to evaluate the operation of wind and solar farms.