eNILM: Dynamic Device Integration in Non-Intrusive Load Monitoring
Abstract
We introduce eNILM, an extensible Non-Intrusive Load Monitoring (NILM) framework capable of simultaneously integrating multiple new devices at runtime. Leveraging a mixture of experts using density-based outlier detectors, such as Local Outlier Factor and Isolation Forest, and clustering algorithms such as DBSCAN and HDBSCAN, eNILM dynamically adapts to changes in household device composition. The framework eliminates the need for retraining existing classifiers, enhancing its scalability and flexibility. We conducted extensive evaluations on the public datasets FIRED and DARCK, demonstrating that eNILM achieves over 85% accuracy in filtering out unknown devices and over 91% accuracy in clustering unknown devices on average. These results highlight eNILM’s potential to substantially improve the practicality and reliability of NILM systems.
Justus Breyer, M.Sc.
Researcher
Kilian Plachetka
Prof. Dr. Muhammad Hamad Alizai
LUMS, Lahore, Pakistan
