Traditional deep learning-based approaches struggle to effectively address long-distance sequence dependencies in anomaly electricity consumption detection. Hence, we innovatively combine the Transformer and BiLSTM and propose a Residual Attention-based TransBiLSTM model (RA-TransBiLSTM) in this paper. Initially, Transformer and BiLSTM are combined to capture long-distance dependencies in time series. Subsequently, a residual attention mechanism is introduced to differentially weigh features, significantly enhancing the model's capability to represent and process diverse features. Considering the common issue of scarce anomaly data in electricity consumption anomaly detection, we further propose a negative sample generation method based on generative adversarial network to effectively balance the dataset's positive and negative sample ratio. Experimental results on the HUE dataset demonstrate that, compared to other time series analysis methods, RA-TransBiLSTM achieves optimal performance.