An AI explained data-driven framework for electricity theft detection with optimized and active machine learning

Electricity theft is a major problem that causes significant financial losses and inefficient power distribution. Effective theft detection systems play a critical role in detecting fraudulent consumption patterns. However, the performance and generalization of traditional theft detection systems are hindered by issues such as class imbalance, lack of labeled data, suboptimal hyperparameter tuning, and limited model interpretability. To overcome these issues, we propose a novel framework that combines active learning and metaheuristic optimization to enhance theft detection performance. Initially, the proposed framework addresses the data imbalance in the State Grid Corporation of China dataset by employing localized randomized affine shadow sampling. Next, two models are proposed to increase classification accuracy: Active Stochastic Gradient Descent (ASGD) and Cuckoo Stochastic Gradient Descent (CSGD). The ASGD uses entropy-based active learning to prioritize informative samples, whereas CSGD incorporates cuckoo search optimization to improve parameter tuning. The proposed ASGD and CSGD models show significant improvements of 36.67 % and 35 %, respectively, over the baseline SGD in accuracy, demonstrating enhanced performance in electricity theft detection. The experimental results demonstrate that ASGD and CSGD outperform state-of-the-art models with an improvement score of 6.57 % and 7.89 % in accuracy, 7.89 % and 9.21 % in F1-score, and 7.14 % and 8.33 % in the precision-recall area under the curve. Furthermore, the results of the proposed models are validated using a 10-fold cross-validation technique to ensure their reliability. Additionally, the statistical significance of ASGD and CSGD is confirmed using a t-test. Finally, two explainable artificial intelligence methods: local interpretable model-agnostic explanations and Shapley additive explanations, are employed to uncover the interpretability and explainability of the proposed models’ predictions. The proposed framework is useful for detecting electricity consumption anomalies as it enhances both classification performance and model interpretability, ensuring more reliable predictions.