Desenvolvimento de proteção diferencial restrita a terra utilizando técnicas de análise de dados e aprendizagem de máquina

Abstract

In power transmission and distribution systems, the most technically, operationally, and financially significant equipment is the power transformer. This device requires reliable and secure protection, typically achieved through differential protection. This method operates based on the vector difference of currents measured at the input and output of the equipment to be protected. However, for delta-wye grounded connected power transformers, an internal fault near the neutral generates low line currents, which prevents the occurrence of a vector difference between currents and, consequently, the protection does not operate. Given this limitation, new approaches to protect or signal such issues are required. This work aimed to develop a Restricted Earth Fault (REF) differential protection scheme using data analysis techniques and machine learning algorithms. A database of 50,435 cases was used, containing signals of four currents in the frequency domain, measured on the wye side of the transformer. These currents represent phase and neutral currents, and the frequency components studied included DC (0 Hz), fundamental (60 Hz), and harmonic components up to the 5th order (300 Hz). The cases were subjected to three feature selection methods, ranking the most significant attributes for detecting internal faults near the neutral. Subsequently, five machine learning algorithms were applied as classifiers, with parameter tuning using grid search and cross-validation. As a result, the decision tree and random forest algorithms achieved accuracies of 99.51% and 99.59%, respectively, and demonstrated stable behavior based on standard deviation and variance analyses. Thus, these algorithms showed potential for the development of a Restricted Earth Fault classifier, providing an initial analysis for future research.