GRMC: A Calendar-Based Gated Residual Mamba Classifier for Interpretable Electricity Theft Detection on Low-Resolution Daily Meter Data

Electricity theft is a major source of non-technical losses in power grids, particularly in regions where only low-resolution daily meter data are available. Detecting theft under such constraints requires models that are both accurate and interpretable. This paper proposes the Gated Residual Mamba Classifier (GRMC), a novel framework that combines the recent Mamba state-space model with gated residual connections and bidirectional pooling. To enhance interpretability and capture real consumption behavior, daily electricity records are transformed into an 8-channel calendar-aware time-series representation, embedding weekly, monthly, and seasonal usage patterns. Using the large-scale State Grid Corporation of China (SGCC) dataset (42,372 users, 1035 days), GRMC is evaluated against widely used machine learning models (Random Forest, XGBoost, LightGBM) and deep learning baselines (CNN, Mamba). GRMC achieves 94.74% accuracy, 94.87% F1-score, and 98.45% AUC, consistently outperforming competing approaches. To the best of our knowledge, this is the first study that integrates calendar-aware contextual features with the Mamba architecture for electricity theft detection. Comparative analysis with recent state-of-the-art methods further confirms GRMC's advantage in recall and robustness to seasonal and behavioral variations. These results demonstrate the potential of GRMC as a reliable and interpretable solution for electricity theft detection in smart grid environments with limited data resolution.