Non-Intrusive Appliance Identification Using Machine Learning and Time-Domain Features

This paper addresses the challenge of nonintrusive load monitoring (NILM), i.e., identifying the combination of turned-on electrical appliances using a single measurement, which is the aggregated power signal. An automated appliance status labelling method based on new event detection is proposed. Time domain statistical features are calculated using various window lengths. The optimum window size is selected and three well-known and widely used machine learning algorithms, i.e., K-Nearest-Neighbors (KNN), Bagged trees, and Boosted trees are used for classification in the context of power consumption disaggregation. The Reference Energy Disaggregation Dataset (REDD) is used in the experiments to evaluate the performance of the classifiers. The simulation results demonstrate the importance of selecting the appropriate window length and optimizing the classifier configuration. Compared to the state-of-the-art, a very competitive performance, coupled with low computational complexity, is achieved with F1-score values over 97%, when considering all the appliances in the REDD dataset.