A machine-learning-based strategy for online prediction of hotspot temperature in dry-type three-phase transformers

Abstract

Finite element analysis is complex, time-consuming, and unsuitable for online implementation when all the details are considered in the electromagnetic and thermal models. This paper proposes an approach combining finite element analysis with a neural network, which can predict the steady-state hotspot temperature of dry-type three-phase transformers with desired accuracy in various operating conditions only based on the simply measurable ambient and electrical quantities. In the proposed approach, the losses of the transformer’s windings and core are calculated through a detailed electromagnetic analysis and used as input heat sources to perform a precise thermal analysis. A dataset is generated by repeating this procedure for various ambient and operating conditions. Then, a feed-forward neural network is trained based on this dataset, ready to predict the steady-state hotspot temperature only using the real-time measurements of current, voltage, and ambient temperature. In this study, the transformer’s electromagnetic-thermal behavior is simulated in COMSOL Multiphysics, and the temperature prediction algorithm is also implemented in MATLAB. Experimental tests on a prototype transformer confirm the validity of the implemented electromagnetic and thermal models. The numerical evaluations on this prototype and a real-scale transformer also show that the average absolute error of the hotspot temperature predictor does not exceed 1 °C in various ambient, loading, and harmonic distortion conditions, even in cases not seen in the training stage.