Abstract:
In this paper, a fast calculation method based on U-net neural network training was proposed to solve the problem of long time for calculating the temperature rise of oil immersed transformer windings using traditional numerical methods, which can quickly predict the temperature rise and hot spots of transformer windings. For a 35 kV oil-immersed transformer, the training sets required for deep learning under different operating conditions was generated by using Fluent software. After determining the optimal combination of hyperparameters, the computational efficiency of the transformer temperature field was significantly improved. Finally, a fiber-optic test temperature measurement platform was established to verify the effectiveness of the algorithm. By using the results obtained from Fluent software as reference, the relative errors of the U-net neural network for the inner and outer B-phase low-voltage winding and the inner high-voltage winding were around 0.24%, 0.21% and 0.39%, and the single calculation time was shortened from 10 854 s to 0.05 s, and the average error between the prediction results and the test temperature was 4 ℃ at the maximum and 2 ℃ at the minimum. The results show that the method can be used to obtain the temperature of oil-immersed transformer windings quickly, and it can meet the real-time simulation requirements of oil-immersed transformer temperature and hot spot digital twin technology.