Técnicas de Inteligência Artificial aplicadas à otimização multiobjetivo de equipamentos eletromagnéticos saturáveis: um estudo de caso em reatores a núcleo saturado

Abstract

This thesis proposes and validates a multi-objective optimization methodology for the design of large-scale Saturated Core Reactors (SCRs), based on replacing three-dimensional Finite Element simulations with Artificial Neural Networks (ANNs) trained as surrogate models. The work begins by contextualizing the design problem of electromagnetic equipment and identifying the research gap concerning the application of ANN-based surrogate models to three-dimensional saturated magnetic structures. A reference reactor, rated at 500 kVAr in a Twin-Trippler topology, was designed, modeled in ANSYS Maxwell software, and experimentally validated through the fabrication of a physical prototype and factory acceptance electrical tests, confirming the accuracy of the adopted computational model. From a simplified model, shown to be equivalent to the detailed manufacturing model, a parametric sweep using Finite Element simulations was conducted to generate a training dataset. A comparative study among machine learning techniques demonstrated the superiority of ANNs in capturing the strongly nonlinear behavior of the equipment under deep saturation, achieving coefficients of determination (R²) above 0.96 for the reactive power and core loss objective functions. The trained ANNs were then employed to exhaustively explore a design space of more than 247 million combinations, enabling the construction of the Pareto frontier between core losses and volume within minutes of processing. The selected optimal design, validated through detailed Finite Element simulation, achieved reductions of 12.86% in magnetic core losses and 14.37% in core volume relative to the reference equipment, while preserving the rated reactive power absorption capacity within acceptable design margins. The results confirm the effectiveness of the proposed methodology as a computationally efficient and experimentally validated design tool for electromagnetic equipment operating under deep magnetic saturation.