The existing evaluation methods based on large-dimensional random matrices cannot reliably evaluate transformers due to the influence of primary voltage fluctuations, this paper proposes a method for evaluating the error condition of voltage transformers based on measurement consistency of same phase. The paper use kernel principal component analysis (KPCA) and reconstruction algorithm to establish evaluating indicators with higher sensitivity. Considering the similarity of the evaluating indicators of same phase, a difference evaluating indicator of voltage transformers based on measuring same phase is established to diminish the influence of stationary fluctuations and judge the polarity of errors by judging whether the difference evaluating indicators are positive or negative. The simulation experiments and field application results suggest that 0.2% error of voltage transformers can be estimated using the improved evaluation method based on large-dimensional random matrices.