A Deep-Learning Surrogate Model for the Fast Calculation of Temperature Distribution of ±500 kV RIP Bushings

In response to the problems of time-consuming simulation of temperature field of resin impregnated paper (RIP) bushings on valve side of converter transformer, which result from a high degree of freedom and complex coupling mechanism, a deep learning surrogate model (DLSM) suitable for fast prediction of bushing temperature is proposed. The model consists of two parts: proper orthogonal decomposition (POD) reduced order model based on back propagation neural network (BPNN) and convolutional neural network (CNN), which predict the outside and inside part of bushing respectively to improve the accuracy. Subsequently, the temperature fields calculated by the finite element method (FEM) and DLSM are compared, and the average relative error is less than 0.7%. In addition, the influence of different train sizes and input perturbations on the precision of the model is analyzed, which is good when at least 108 samples are trained. The prediction results of the model with input perturbation within 5% are similar to the FEM results. Finally, the time for the model to predict 30 test conditions is 1.5 s, which is much less than the time of finite element simulation. The research results show that DLSM has high precision, robustness and prediction speed. The method in this paper provides a feasible solution for online monitoring and digital twins of bushings.