A data-tagging and temporal-association framework for robust and explainable anomaly detection in power communication systems

With the growth of smart grids, power communication networks face challenges in detecting diverse and hidden anomalies due to business heterogeneity, temporal dependencies, and load volatility. Existing statistical and machine learning methods often suffer from poor interpretability, high false alarms, and limited robustness. This paper proposes a novel data-tagging and temporal-association-based anomaly detection framework. By constructing a multi-dimensional tagging system that integrates business, node, and temporal features, raw load data are transformed into semantic-tagged objects. A sparse and simplified preprocessing reduces noise and dimensionality, while an ensemble detection method combining local density, isolation forest, and anomaly transformer improves detection performance and interpretability. Experiments on real-world power station datasets show the proposed method achieves superior accuracy, recall, and stability, effectively detecting both point-wise and segment-wise anomalies hidden in periodic patterns. SHAP analysis further reveals key feature contributions, enhancing transparency and aiding decision-making.