Abstract
Non-Intrusive Load Monitoring (NILM), an important application of machine learning, frequently misidentifies activities of unknown devices, resulting in incorrect energy consumption estimates. This paper proposes an innovative filtering step between event detection and classification of event-based NILM to exclude events from unknown devices. This approach incorporates confidence-based classifiers, clustering, ensembling, and density-based techniques, notably Local Outlier Factor and One-Class SVM. The best techniques reduce false positives (over 93%) for unknown devices while preserving most events from known devices (less than 7% loss). This significant advancement enhances event-based NILM system accuracy, offering more reliable energy monitoring for real-world applications, and thereby contributes to broader energy conservation efforts in the context of climate change.
Justus Breyer, M.Sc.
Researcher
Asst. Prof. Dr. Muhammad Hamad Alizai
LUMS, Lahore, Pakistan
Sandeep Samant
