ISO-DBSCAN: An Unsupervised Approach to Anomaly Detection with NILM Data in Residential Buildings

Abstract

The development of an efficient energy management solution is a challenging research area, particularly in addressing unusual appliance behavior, safety, and predictive maintenance in residential settings. Many studies have been limited to labeled datasets and the identification of hidden clusters in the datasets has been impractical. This study introduces ISODBSCAN, a hybrid unsupervised model that combines density-based spatial clustering of noise applications (DBSCAN) and isolation forest to detect anomalies in NILM data without labeled inputs. Using data from House 4 in the REDD dataset, we pre-process the time series energy readings and extract segments representing appliance behavior. DBSCAN is used to identify dense clusters of normal usage, while Isolation Forest further analyses these results to isolate outliers that deviate from expected patterns. The model achieves a silhouette score of 0.46, outperforming conventional methods like K-Means and Gaussian Mixture Models (GMM), which struggle with irregular cluster shapes and require predefined parameters. ISO-DBSCAN offers both robustness and interpretability, effectively isolating anomalies and mapping them back to specific time windows, thereby enhancing transparency and facilitating real-time monitoring. This approach is particularly beneficial in residential settings where energy behavior varies and labels are unavailable. The model’s performance is sensitive to DBSCAN parameters, such as minPts. Future work includes integrating automated hyperparameter tuning techniques, such as Bayesian optimization, expanding the approach to additional households and data sets, and improving real-time usability through dashboards and alert systems.