Double-layer stacking optimization for electricity theft detection considering data incompleteness and intra-class imbalance

• Application of the TimeGAN network for Electricity Theft Data Generation: To the best of our knowledge, our research group is the first to apply the TimeGAN network specifically for generating data related to electricity theft. This innovative approach enables the model to learn both the temporal features and the generative representation of electricity theft data by jointly training the embedding and recovery functions. • Introduction of the Macro-Average F1 Score as an Objective Function in K-Fold Cross-Validation: We present the use of the macro-average F1 score as the objective function for K-fold cross-validation. Traditional accuracy and recall metrics often fall short in accurately evaluating model performance, especially with unbalanced sample class distributions. Our approach utilizes the mean of the macro-average F1 score across K-folds, effectively mitigating the impact of unbalanced samples on model evaluation and ensuring a more accurate and reliable assessment process. • Two-Layer Optimization Configuration Model Based on the WOA: This paper proposes a two-layer optimization configuration model inspired by the WOA. This model incorporates time-varying binary transfer functions to dynamically adjust the weights of the exploration and exploitation phases during the optimization process. As a result, it enhances the algorithm’s search performance and convergence speed, demonstrating superior effectiveness compared to traditional optimization algorithms in solving optimization problems. Electricity theft detection is crucial for reducing non-technical losses in electric power enterprises and ensuring fairness in electric power transactions. However, challenges such as incomplete data, a scarcity of electricity theft samples, and limited performance of detection systems hinder accurate identification of theft. To address these issues, this paper presents a method utilizing a two-stage time-series generative adversarial network (TimeGAN) with an integrated two-layer stacking optimization configuration. The first phase tackles data incompleteness through an enhanced version of TimeGAN, employing embedding and recovery layers to reconstruct incomplete power user data. It introduces an analog denoising training method and supervised information assistance to improve the interpolation accuracy. In the second stage, the method addresses inter- and intra-class imbalances in electricity theft detection by employing the K-shape clustering algorithm to identify unique patterns within the theft data. This enables balanced synthesis of theft samples using these patterns as conditional supervisory terms for TimeGAN. To concurrently optimize the combination of the electricity theft detector (ETD) model and its hyperparameters, an integrated two-layer Stacking optimization framework is developed. This framework incorporates a time-varying binary transfer function and an external repository to enhance the performance of the optimization algorithm. Simulation tests performed on 42,372 actual power consumption records demonstrated that the proposed method achieved, on average, 8.23% higher DR scores, 3.45% higher AUC scores, and 5.60% higher Macro-F1 scores compared to the baseline method.