To enhance the operational efficiency of power grids and mitigate losses, this study proposes an abnormal electricity consumption detection method grounded in the Time Series Generative Adversarial Network (Time GAN) and the CNN-Transformer-Bi LSTM architecture. Initially, the raw dataset is preprocessed through data cleansing and missing value imputation. Subsequently, to address the scarcity and low prevalence of abnormal samples in real-world data, a Time-GAN data augmentation approach integrated with a multi-head self-attention mechanism is put forward. The incorporation of this mechanism effectively elevates the quality of generated abnormal data. Next, a hybrid CNN-Transformer-Bi LSTM model is utilized. By reducing the dimensionality of CNN kernels, it extracts local short-term periodicities in sequential data. Transformer is designed to capture long-range global dependencies, and Bi LSTM is employed to extract time-dependent features of column-based data. Their synergy enhances the model’s capacity to learn sample characteristics. Validation on real-world datasets demonstrates that the proposed method outperforms existing models in comprehensive metrics such as accuracy (ACC) and the area under the curve (AUC) value.