Fast and Accurate Design of Single-Phase Transformers Using Bayesian Neural Networks

Abstract

The accurate design of single-phase transformers is essential for achieving high efficiency, reliability, and cost-effectiveness in power applications. However, traditional analytical approaches often depend on simplifying assumptions that limit precision, particularly when estimating core losses and leakage flux. This study introduces a Bayesian Neural Network (BNN)–based method for the precise design of small single-phase power transformers. The Finite Element Method (FEM) is used to generate detailed electromagnetic and performance data under various operating conditions, forming the training dataset for the BNN model. The trained BNN effectively captures complex nonlinear relationships between design parameters and performance indices, enabling fast and accurate prediction of transformer characteristics without repeated FEM simulations. The proposed approach significantly reduces design time, enhances prediction accuracy, and minimizes dependence on empirical trial-and-error techniques. Additionally, an experimental setup is developed to determine the operating point of an existing single-phase transformer, which is subsequently redesigned using the BNN methodology. This study provides a reliable and efficient framework for academic research and education in applying machine learning (ML), including artificial neural networks (ANNs), to the accurate design of electrical equipment.