Detecting anomalies in energy consumption is critical for efficient energy management, fault detection, and sustainability. However, the challenge of class imbalance, where normal consumption data vastly outweighs anomalous instances, presents significant difficulties in building accurate predictive models. This paper conducts a comparative analysis of class imbalance handling techniques for deep models in detecting anomalies in energy consumption data. Specifically, controlled experiments are used to evaluate the performance of deep learning models, such as convolution neural networks (CNN), long short-term memory (LSTM) and BiLSTM deep algorithms as well as synthetic data generation (SMOTE), cost-sensitive learning, and generative adversarial networks (GAN) tailored to address the imbalance issue. Through a comprehensive empirical study using a real-world energy dataset, we assess the models' effectiveness based on area under the curve (AUC), precision, recall, F1-score, and their ability to generalize across different levels of imbalance. This research contributes to improving model selection for practitioners facing the class imbalance challenge in the energy sector.