Accurate prediction of error is significant for detecting a stability problem in a transformer in good time and ensuring the validity of electric power trade. An ultra-short-term transformer error prediction method based on variational mode decomposition (VMD) and an encoder-decoder model based on a convolutional neural network-gated recurrent unit encoder to decoder moder (CNN-GRU-ED) is proposed. First, because of the data analysis problem caused by the increase in the number of features, transfer entropy (TE) is used to reduce the dimensionality of the original features to obtain a highly correlated feature set. Second, the original error sequence with high fluctuation is decomposed into high and low-frequency modal components, in which the critical parameter of VMD is optimized by particle swarm optimization (PSO). Finally, the CNN-GRU-ED model is built for each component, and the predicted results are superimposed to obtain the final predicted value. Experiments with the operating data of a substation show that the performance of the proposed method is better than other models in single-step and multi-step prediction.