Análises experimentais de algoritmos de aprendizagem de máquina na classificação de distúrbios elétricos

Abstract

In the last years, the electric power quality analysis came back to the fore with the advance in many computational areas and the serious issues related to the deterioration of it, that can cause damages to the life, lifetime reduction of the electrical devices, and, consequent environment and economic impacts. In this sense, it was verified that the interest in researches related to the classification of the disturbances that cause the deterioration of the energy quality returned to grow, mainly the creation of systems that can implement it in a real environment. Then, this work aims to verify, experimentally, many strands related to the process of electrical disturbance classification and implementation of embedded systems with the purpose of implement the classification process. Therefore, it was analyzed experimentally the accuracy of different machine learning algorithms (obtaining as a better result the accuracy of the Random forest, both in noisy and not noisy signals), efficiency of the algorithms in embedded environments (in which the Decision Tree was the best), and the analysis of the classification of multiple disturbances (in which 22 disturbances were analyzed and the Random Forest get an accuracy above 90% in not noisy environment). Besides that, it was implemented a benchmark2 of simulated disturbances and make it available for the community (as well as the source code in Python), aiming to stay easier make comparisons among different proposes of different authors.