With the rapid digitalization of power systems and the wide spreading deployment of advanced metering infrastructure, utilities are now confronted with massive volumes of fine-grained power marketing data. Effectively identifying abnormal behavior in these data streams is crucial not only for revenue assurance, but also for safeguarding the operational security of modern smart grids. An attention-enhanced hybrid deep learning framework is introduced for anomaly detection in power marketing data. The proposed model integrates Convolutional Neural Networks (CNNs) CNNs to extract local morphological features and short-term trends, Long Short-Term Memory (LSTM) networks to capture temporal dynamics, and an attention mechanism that adaptively emphasizes the most informative features for distinguishing normal and abnormal records. To further improve convergence speed and generalization performance, the entire architecture is optimized using the RIME metaheuristic algorithm. Experiments conducted on real operational data from a provincial power marketing system show that the proposed RIME-CNN-LSTM-Attention model attains an accuracy of 93.67% and an F1-score of 94.75%, outperforming a range of conventional baseline methods. The results highlight the promise of combining architectural innovation with advanced intelligent optimization techniques to tackle the increasingly complex anomaly detection tasks in contemporary smart grid environments.
Zhao et al. (Thu,) studied this question.