To accurately locate electricity theft and reduce the economic losses caused by electricity theft to the power system, a multi-classification electricity theft detection model based on the entropy weight method Stacking E_Stacking(stacking based entropy) ensemble learning was proposed. First, based on the collinear characteristics of electricity consumption information, the VIF(variance inflation factor) was used as a standard to reduce the dimensionality of the data to reduce data complexity. Then k-fold cross-validation was embedded during model training to effectively prevent model overfitting. This model contained two layers of learners, a primary learner, and a meta-learner, it could fully combine the advantages of the two-layer learner and combine the learned complementary features and discriminative features to further improve detection performance. Finally, the Irish dataset and a portion of UCI(University of California Irvine) datasets were used to verify the model, and the results were better than several common methods at present, which demonstrates that the model has certain effectiveness and robustness.