Aprendizado de Máquina aplicado na previsão da geração de energia elétrica de uma usina solar fotovoltaica no Ceará

Abstract

Given the growing of solar photovoltaic (PV) power in the Brazilian power mix, and considering that PV generation is intermittent, there is great need of PV generation forecasting models for better operational planning of both the PV power plant itself and the power grid. Thus, this paper proposes the application of Machine Learning models to forecast the power generation of a 160 MW PV plant located in the state of Ceará (Brazil). The aim is to forecast the daily power output of the power plant over 365 days based on the historical power data and the weather data of the PV power plant. To accomplish the task, distinct prediction models were tested, such as: sequences recognition, Artificial Neural Networks (ANN), XGBoost and hybrid approaches. The performance of the implemented models was evaluated using error metrics, and the XGBoost model achieved the most accurate results as regard to the prediction error and the execution time followed by the ANN. The Pattern Sequence-based Forecast (PSF), which is more transparent than an ANN or even XGBoost, has proved competitive, having the best performance among the models trained only with historical data of power generation.