A Surrogate Model for Steady-State Temperature Rise Calculation of Oil-Immersed Transformer Windings Integrating Model Order Reduction and Neural Networks

Abstract

The limited accuracy and long computation time of temperature-field solutions for oil-immersed transformers impede rapid perception of their internal state. This paper proposes a tailored hybrid physics–data surrogate model for steady-state temperature fields of oil-immersed transformers that deeply integrates Proper Orthogonal Decomposition (POD) with a specifically modified neural network architecture. First, a full-order model of the oil-immersed transformer is built, and temperature-field simulations are performed under various operating conditions. Second, Proper Orthogonal Decomposition (POD) is applied to extract characteristic modes from the simulated temperature-field snapshots, transforming the nonlinear temperature field into a linear combination of modes and modal coefficients. Finally, a Gaussian-activated fully connected neural network with local approximation capability is constructed, and a physics-informed loss function embedding POD modal importance is designed to approximate the nonlinear mapping between transformer operating parameters and modal coefficients, thereby completing the surrogate model construction. For the temperature-field solution under a given condition, the field can be rapidly reconstructed simply by linearly weighting the modes with modal coefficients predicted by the surrogate model. Validation shows that the proposed surrogate model can compute a single steady-state temperature-field distribution within 0.0129 s, and the maximum computation errors for global temperature and hotspot temperature do not exceed 3.5%, with the numerical results further validated by optical fibre temperature measurement experiments, meeting the requirement for real-time temperature-field perception in on-site operation and maintenance.