A blind event-based learning algorithm for non-intrusive load disaggregation

• Published in: International journal of electrical power & energy systems Vol. 129; p. 106834
• Main Authors: Qureshi, Moomal, Ghiaus, Christian, Ahmad, Naveed
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.07.2021; Elsevier
• Subjects: Computer Science; Electric power; Engineering Sciences; Gaussian mixture models; Modeling and Simulation; Non-intrusive load monitoring; Real-time disaggregation; Unsupervised learning; Non-intrusive load monitoring; Gaussian mixture models; Real-time disaggregation; Unsupervised learning; online disaggregation; unsupervised; real-time disaggregation
• ISSN: 0142-0615; 1879-3517; 1879-3517
• DOI: 10.1016/j.ijepes.2021.106834

•A blind selection method for a threshold-based event identification.•Evaluation of unsupervised learning using GMM for energy disaggregation.•Identification of individual power levels in the data and constructing appliance models without any user intervention.•Comparison with other state of the art methods.•Evaluation of GMM based technique against a mean-shift clustering approach.•The results of the disaggregation are verified on a private dataset as well as public dataset. Non-intrusive loading monitoring (NILM) provides a smart solution to the problem of electrical energy monitoring of households at the appliance level. In blind disaggregation, the power level of each appliance is not known a priori. In this paper, we propose an event-based blind disaggregation algorithm that uses Gaussian mixture models (GMM) for clustering to automatically detect two-state appliances from the aggregate data. The benefit of using Gaussian mixture models over other clustering methods is that they can automatically learn the statistical distributions present in the data. This is beneficial, especially when the appliances have similar power consumptions. Since Gaussian mixture models do not determine the number of clusters automatically, we use Bayesian information criteria (BIC) to determine the number of clusters. The blind disaggregation method is tested with data from a real house collected by a smart meter which samples the aggregate consumption at 3.4 kHz and also from Reference Energy Disaggregation Dataset (REDD) public data, sampled at a frequency of 1 Hz. We compared the performance of our algorithm with other unsupervised methods and found comparable performance. We also compared the Gaussian mixture model with mean shift clustering in a blind disaggregation. We saw an improved performance by using Gaussian mixture models instead of mean-shift clustering in various accuracy measures.