A clustering-based supervised approach for anomaly detection in building energy consumption

Since buildings account for a significant portion of global energy consumption, energy efficiency is a critical research topic. This study proposes a novel clustering-based approach for detecting building energy anomalies. Unlike the single global model approach in the literature, the proposed method first clusters buildings according to their consumption patterns using the K-means algorithm; then, independent anomaly detection is performed for each cluster. Isolation Forest and Local Outlier Factor are used as unsupervised methods, while XGBoost, Random Forest, and LightGBM are used as supervised methods. Feature engineering was performed using building-level statistical features during the clustering phase. Experiments were conducted on the LEAD 1.0 dataset. The results show that models trained on buildings with similar consumption patterns provide significant performance improvement compared to models trained on the entire dataset. While the highest F1 score obtained by the global model was 0.572, the proposed cluster-based approach increased this value to 0.903. Among the models, XGBoost and LightGBM stood out as the best-performing models across clusters, with LightGBM achieving the highest F1 score of 0.903 in Cluster 2. Further analyses revealed that model performance improved significantly as data homogeneity increased, strengthening the effectiveness of the clustering approach.