Prognostic analysis of power transformers using high frequency models and Complex-Valued neural networks

This paper proposes a prognostic approach capable of identifying different failure mechanisms in electrical power transformers starting from their high-frequency equivalent circuits. The main objective is to define a general procedure for detecting malfunctions in power transformers before they can lead to catastrophic consequences. Indeed, these malfunction conditions are characterized by variations in the components of the equivalent lumped-parameter model, and this paper shows that they can be correctly recognized through a complex-valued neural network. Therefore, the main contributions of the work consist of the experimental characterization of a power transformer in the 20 Hz - 1 MHz range considering the effects of ambient temperature variations, the parametric analysis of malfunctions and the definition of a procedure based on Power Line Communication technologies and machine learning to detect anomalous situations. The prognostic analysis is organized into four successive phases: experimental measurements, calculation of lumped parameters, simulation of the equivalent circuit and prognostic analysis using complex-valued neural networks. The measurements are performed on a low-voltage isolation transformer in Delta/Y configuration and this allows the overall procedure to be extended to medium voltage transformers used in distribution networks. The results show that malfunctions can be recognized with an average accuracy level of 97%.