Scholarlink

<p>Designing an effective Electricity Theft Detection (ETD) system is a challenging task due to challenges such as severe class imbalance, high-dimensional consumption data, and high false positive rates. Furthermore, existing studies have largely overlooked the integration of Explainable Artificial Intelligence (XAI) within ETD frameworks. To address these limitations, this study presents an interpretable hybrid deep learning approach for ETD. The proposed framework combines comprehensive data preprocessing, class balancing, feature selection, robust classification, and post-hoc interpretability. A Temporal Convolutional Network (TCN) is utilized to learn discriminative representations from high-dimensional time-series data, while a Graph Convolutional Network (GCN) leverages relational information to accurately classify consumers as legitimate or fraudulent. To enhance model transparency, Shapley Additive Explanations (SHAP) are incorporated to deliver intuitive, human-understandable explanations of prediction outcomes. Extensive experiments conducted on a real-world dataset demonstrate that the proposed framework achieves benchmark performance, attaining 95.70% accuracy and a 95.90% F1-score, and consistently outperforming state-of-the-art baselines. Moreover, evaluation using the recently introduced MARS metrics further confirms the robustness and reliability of the framework. Overall, this work provides a high-performing and interpretable ETD solution, offering both methodological contributions and practical benefits for enhancing smart grid security.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable fraud detection in smart grids using enhanced deep learning approach

  • 期刊论文

  • 发表日期：

    2026年1月12日

查看全文
  • Muhammad Zeeshan Younas,
  • Muhammad Shahid Iqbal Malik

摘要

Designing an effective Electricity Theft Detection (ETD) system is a challenging task due to challenges such as severe class imbalance, high-dimensional consumption data, and high false positive rates. Furthermore, existing studies have largely overlooked the integration of Explainable Artificial Intelligence (XAI) within ETD frameworks. To address these limitations, this study presents an interpretable hybrid deep learning approach for ETD. The proposed framework combines comprehensive data preprocessing, class balancing, feature selection, robust classification, and post-hoc interpretability. A Temporal Convolutional Network (TCN) is utilized to learn discriminative representations from high-dimensional time-series data, while a Graph Convolutional Network (GCN) leverages relational information to accurately classify consumers as legitimate or fraudulent. To enhance model transparency, Shapley Additive Explanations (SHAP) are incorporated to deliver intuitive, human-understandable explanations of prediction outcomes. Extensive experiments conducted on a real-world dataset demonstrate that the proposed framework achieves benchmark performance, attaining 95.70% accuracy and a 95.90% F1-score, and consistently outperforming state-of-the-art baselines. Moreover, evaluation using the recently introduced MARS metrics further confirms the robustness and reliability of the framework. Overall, this work provides a high-performing and interpretable ETD solution, offering both methodological contributions and practical benefits for enhancing smart grid security.