Real-Time Magnetic Flux Prediction for Transformer Short-Circuit Fault Diagnostics With Deep Neural Network and Autoencoder

Abstract

Accurate prediction of magnetic flux distribution in power transformers under winding short-circuit faults is essential for online monitoring and digital-twin-based diagnostics. Existing data-driven surrogate models typically rely on standard deep neural networks (DNNs) and linear dimensionality reduction techniques, which limits their ability to capture the nonlinear spatial characteristics of electromagnetic fields. This article proposes a latent-space surrogate modeling framework that integrates finite element-generated magnetic-field data with an autoencoder (AE)-based nonlinear latent representation. The AE effectively compresses high-dimensional flux fields into a compact feature space that preserves dominant flux-distribution patterns, while a DNN regression model maps electrical inputs to the latent features for fast magnetic-field reconstruction. Three surrogate architectures—DNN+ principal component analysis (PCA), DNN+AE(DNN), and DNN+AE [long short-term memory (LSTM)]—are evaluated using transient short-circuit simulations. Results show that the proposed DNN+AE(DNN) achieves the best overall accuracy in the final flux prediction task, outperforming both the PCA baseline and the LSTM-based AE. The method attains a leakage-flux mean absolute percentage error (MAPE) of 3.04% and delivers up to a $51\times $ speedup relative to conventional finite element analysis (FEA). These findings demonstrate the effectiveness of nonlinear representation learning for fast and accurate magnetic flux prediction in transformer fault analysis.