Detecting user-side electricity theft accurately has long been a challenge for power supply companies, with traditional theft detection methods having certain limitations. Addressing the highly imbalanced positive and negative samples in the field of theft detection, and the poor performance of single classification models, this study proposes a theft detection method based on an improved Rotation Forest algorithm. The Rotation Forest algorithm uses Principal Component Analysis (PCA) for feature extraction, training each base classifier with all principal components of the original training set. Building upon the classical Rotation Forest algorithm, improvements are made in three aspects: balancing the positive and negative samples in the subset using the Synthetic Minority Oversampling Technique (SMOTE) algorithm, further sampling the training subset using Bootstrap sampling in the Bagging algorithm, and selectively integrating base classifiers based on accuracy. A case study using actual user data from a region in East China demonstrates that the proposed theft detection method achieves better results in multiple evaluation metrics compared to single classification models and existing ensemble learning strategies.