Energy fraud in electricity consumption, particularly non-technical losses caused by electricity theft, presents significant financial and operational challenges for utility companies. This study compares various classification and anomaly detection methods applied to the real-world electricity consumption dataset from State Grid Corporation of China (SGCC) for electricity theft detection (ETD). The investigation focuses on handling missing values using Linear Interpolation, also addresses class imbalance with Synthetic Minority Over-sampling Technique. Classification methods, including Random Forest and Light Gradient Boost, are compared to anomaly detection methods One-Class SVM, Isolation Forest and AutoEncoders. Performance evaluation uses metrics like Recall and F1-Score. Preliminary results from the comparison between the two approaches favor classification with class balancing; however, this represents a non-realistic scenario for the energy distribution company. Currently, we are evaluating strategies to improve the performance of anomaly detection methods (i.e., one-class classification), as these can be applied in more realistic scenarios.