Desenvolvimento de algoritmo para detecção de arco elétrico série no lado cc em sistemas de geração fotovoltaica utilizando máquina de vetores de suporte

Abstract

In photovoltaic systems, series electric arc faults are the main causes of fires and are strongly associated with the aging and wear of insulating materials, damaged cables, loose connections, connector corrosion, among other possible reasons. In this context, a machine learning algorithm based on Support Vector Machine was developed to detect this phenomenon on the DC side of photovoltaic systems. The proposed learning model uses training and testing data obtained from tests conducted in a solar power plant. The acquired data include events such as inverter start/stop, partial shading, maximum power point tracking noise, pre-arc noise, series arc, string connection/disconnection, and current anomalies due to external disturbances. For the creation of this model, aimed at classifying time series into arc and non-arc categories, two attributes were considered. One is derived from calculating the standard deviation of the pre-processing of the current signal, and the second comes from calculating the standard deviation of the Fast Fourier Transform applied to the current signal. The detection algorithm demonstrated good performance; in the best cases, it achieved an accuracy of 97.3%, capable of identifying series electric arcs and differentiating them from other events that can cause anomalies in the current. Besides accuracy, other indicators such as the confusion matrix and Receiver Operating Characteristic and Precision-Recall curves were used to evaluate the model’s performance.