A Generative Prediction Model Based on TimeVAE and Crossformer for Predicting Multi-Channel Data of Transformer Bushing

Transformer bushings play a crucial role in power systems. Despite their importance, there has been limited research specifically on the state prediction of transformer bushings. Moreover, currently available models only predict univariate data of bushings without considering the physical connection to other relevant data, which would give rise to a compromised prediction. To achieve a more accurate joint prediction of three critical transformer states: dielectric loss, end shield current, and equivalent capacitance, this work builds upon the state-of-the-art predictive network Crossformer. In addition to Crossformer, a data generation model named TimeVAE is introduced to enlarge the scale of the training dataset. This combination enabled us to achieve accurate multivariate predictions for transformer bushings. To evaluate the effectiveness of the proposed predictor, we conducted comparative experiments with attention-based frameworks (Autoformer, Pyraformer, Informer, Reformer, Transformer), which are considered efficient for multivariable prediction tasks. The results show that our proposed model outperforms these models in four key metrics: mean square error, mean absolute error, root mean squared error, and coefficient of determination.