EGBAD: Ensemble graph-boosted anomaly detection for user-level multi-energy load data

Anomaly detection is crucial for data-driven applications in integrated energy systems. Traditional anomaly detection methods primarily focus on one single energy load, often neglecting potential spatial correlations between multivariate energy time series. Meanwhile, addressing the imbalanced nature of user-level multi-energy load data remains a significant challenge. In this paper, we propose EGBAD, an Ensemble Graph-Boosted Anomaly Detection framework for user-level multi-energy load that leverages the advantages of graph relational analysis and ensemble learning. First, a dynamic graph construction method based on multidimensional scaling (MDS) is proposed to transform multi-energy load data into graph representations. These graphs are subsequently processed using graph convolutional network (GCN) to capture the spatiotemporal correlations between multi-energy load time series. In addition, to improve detection robustness under class imbalance, the entire training process is embedded within a Boosting ensemble learning framework, where the weight assigned to the minority class is progressively increased at each boosting stage. Experimental results on publicly real-world datasets demonstrate that the proposed model achieves superior anomaly detection accuracy compared to most baseline methods. Notably, it performs especially well in scenarios characterized by extreme data imbalance, achieving the highest recall and F1-score for anomaly detection.