Background: Electronic voltage transformer (EVT) is an essential metering device for future substation automation systems. One of the main drawbacks of EVT is its poor long-term stability, which affects its measurement accuracy. This will, in turn, adversely affect the entire protection and control systems it is employed for. Objective: Aiming at reducing the EVT measurement error over long-term operation, an EVT error prediction method combining Prophet, temporal convolutional network (TCN) and selfattention is proposed in this paper. Methods: The proposed method is based on building prophet and TCN error prediction models to estimate preliminary prediction values. On this basis, self-attention is introduced to further extract features and make full use of the useful information in historical data. Then the secondary prediction can be achieved, and the final predicted value can be reported as an output. Results: The proposed method is validated by applying the error data of an EVT in a substation to its historical operation. The results show that the model can effectively predict the error trend of EVT. Conclusion: The prediction results of this method are similar to the fluctuations of the actual values, indicating that it provides a new reliable method for error prediction of EVT.