Explainable Anomaly Detection for Grid Monitoring using Probabilistic Load Forecasting

Anomaly detection (AD) plays a crucial role in smart grids. The integration of power-intensive, internet-connected technologies in the residential sector gives rise to new risk scenarios, which can compromise grid stability if not promptly detected. Forecasting-based AD offers a real-time capable solution for grid operators and provides interpretable results. However, commonly used deterministic threshold-based techniques struggle with heavily fluctuating data at high temporal resolution. In this paper, we propose a novel probabilistic real-time AD framework. The two-stage approach combines probabilistic very-short-term load forecasting with a quantile-based scoring function. By considering the predictive distribution of power demand based on external features, our method characterizes the inherent uncertainty in power demand patterns, allowing for clearer distinction between normal variations and critical anomalies. Our method is tested on real aggregate household data, from which we can detect abnormal demand patterns of individual households. In addition, our method is evaluated with synthetic anomalies using classification metrics and the receiver-operator characteristic (ROC), demonstrating superior performance over conventional deterministic methods. By leveraging probabilistic forecasting for AD, our method offers a straightforward, transparent, yet accurate and robust tool for smart grid nowcasting. The proposed real-time framework can provide grid operators with a warning system to anticipate and mitigate critical situations timely, thereby enhancing power system resilience.