Aprendizado de máquina aplicado na previsão da geração de energia elétrica de uma usina solar fotovoltaica

Abstract

Given the growing integration of photovoltaic (PV) solar power generation in Brazilian and world power grids and considering that solar power generation is intermittent, there is an urge to build prediction models to forecast the solar power output, in order to allow better planning and operation of the power plant and the electrical grid itself. Thus, this paper proposes the application of Machine Learning to forecast the power generation of a 160 MW PV plant located in the state of Ceará. It is proposed the task of forecasting the PV power output at hourly intervals for the next 365 days (one year), given the previous year's time series of PV power outputs and weather data collected from the plant. To this end, 13 distinct prediction models were implemented, being different from each other in either methodology or training method. The models implemented in this paper are based on: sequences recognition, Artificial Neural Networks (ANNs), and XGBoost, as well as there is the use of hybrid models, coupling two distinct methods. Then, the implemented models had their performance evaluated using error-based metrics, which showed that the XGBoost model achieved the most accurate results. The main contribution of this work was the identification of the models with the best performance in the proposed task, including the scenario in which the weather data was unavailable.