Aplicações de inteligência artificial na proteção de transformadores: revisão comparativa

Abstract

This work presents a comparative review between traditional power transformer protection methods and Artificial Intelligence (AI) based approaches. The main objective is to investigate how intelligent techniques can enhance reliability, selectivity, and response time levels of protection systems. The adopted methodology is exploratory and comparative, based on bibliographic research and technical analysis of recent studies applying Artificial Neural Networks (ANN), Deep Learning, and hybrid models. The analysis results indicate that traditional methods, although robust and standardized, present structural limitations regarding current transformer (CT) saturation and inrush currents with low harmonic content. In contrast, AI applications, especially convolutional neural network (CNN) architectures and models such as Long Short-Term Memory (LSTM), demonstrate superior performance, achieving detection times reduced to the 1 to 5 ms range and a significant reduction in false trip rates. It was observed that the use of autoencoders contributes significantly to system robustness in noisy and saturated environments. It is concluded that Artificial Intelligence acts as an essential complementary evolution for the safe operation of smart grids and modern high-complexity power systems. However, the feasibility of its large-scale implementation still depends on advances in methodological standardization, field validation, and the availability of real and representative databases.