Sistema de regras para acompanhamento de performance em usinas fotovoltaicas empregando técnicas de aprendizagem de máquina

Abstract

Photovoltaic power generation continues growing its share on energy matrices in Brazil and worldwide, justifying the development of new technologies that allow better efficiency and performance monitoring. Given that, this work proposes a performance assessment rule system for photovoltaic plants, based on expected power estimation. Machine Learning Techniques were also applied aiming to approximate a photovoltaic system’s expected energy, based on historical time-series of electric power and environmental variables (e.g.: irradiance, ambient temperature, wind speed and air humidity). Using real data collected from a photovoltaic plant located in Brazil’s Northeast region, filters and preprocessing techniques are applied to improve the Machine Learning models compared on the task of regression: Multilayer Perceptron, k-Nearest Neighbors and Random Forest. The performance of the energy models obtained by the Machine Learning techniques are compared to a physical model available in PVlib, a Python library for photovoltaic systems simulation. The results show a greater accuracy for the models obtained with the Machine Learning, with Random Forest presenting 57.42 as Root Mean Square Error on the test data. The proposed system for identifying low performance occurrences is described in its implementation details and presented by a graphical user interface. A case study is conducted to validate the proposed system, obtaining 667 low performance occurrences detected for 32 inverters in a 31 days period.