Fast calculation of temperature distribution in oil-immersed transformer windings based on U-net neural network

Abstract

In this paper, a fast computational method based on the training of the U-net neural network is proposed to solve the fluid-thermal coupling problem of transformers. First, the input variables are selected according to the fluid-thermal coupling principle, and the finite volume method is applied to obtain the output results under different operating conditions and make them into training sets and test sets. Second, the training sets are normalized and input into the U-net neural network. At the same time, three hyperparameters that have more influence on the network training are discussed in detail, and the optimal combination of hyperparameters is determined. Finally, the prediction set is fed into the trained model for prediction computation, and the results are subjected to the inverse normalization operation, whose predictive results are consistent with those calculated by the finite volume method. In addition, the computational time is shortened from 300 to 0.07 s, and the prediction variance does not exceed 0.055 K2. The results show that the proposed method can be used to obtain the temperature of immersed transformer windings quickly, which provides an effective tool for real-time temperature simulation of the digital twin of power transformers.