Convolutional Physics-Informed Neural Networks for Fast Prediction of Core Losses in Axisymmetric Transformers

In this paper, solutions of the magnetostatic Maxwell's Equation on parametric axisymmetric transformer geometries are approximated by a convolutional approach on Physics-Informed Neural Networks (ConvPINN). The trained ConvPINN is capable of predicting magnetic vector potentials (MVPs) and magnetic flux densities in a matter of milliseconds for a range of geometries described by a total of 18-20 degrees of freedom (DoF). The combination of ConvPINN with an existing framework for core loss prediction yields a super fast workflow for the approximation of core losses on a wide range of geometric setups. The combination of inference speed and accuracy enables new orders of magnitude for the optimization of transformer designs going forward.