The accuracy of voltage transformer (VT) measurements is imperative for the security and reliability of power systems and the equitability of energy transactions. The integration of a substantial number of electric vehicles (EVs) and their charging infrastructures into the grid poses new challenges for VT measurement fidelity, including voltage instabilities and harmonic disruptions. This paper introduces an innovative transformer measurement error prediction model that synthesizes Multivariate Variational Mode Decomposition (MVMD) with a deep learning framework integrating Bidirectional Temporal Convolutional Network and Multi-Head Attention mechanism (BiTCN-MHA). The paper is aimed at enhancing VT measurement accuracy under fluctuating load conditions. Initially, the optimization of parameter selection within the MVMD algorithm enhances the accuracy and interpretability of bi-channel signal decomposition. Subsequently, the model applies the Spearman rank correlation coefficient to extract dominant modal components from both the decomposed load and original ratio error sequences to form the basis for input signal channels in the BiTCN-MHA model. By superimposing predictive components, an effective prediction of future VT measurement error trends can be achieved. This comprehensive approach, accounting for input load correlations and temporal dynamics, facilitates robust predictions of future VT measurement error trends. Computational example analysis of empirical operational VT data shows that, compared to before decomposition, the proposed method reduces the Root-Mean-Square Error (RMSE) by 17.9% and the Mean Absolute Error (MAE) by 23.2%, confirming the method’s robustness and superiority in accurately forecasting VT measurement error trends.