Non-Intrusive Load Monitoring: Disaggregation of Energy by Unsupervised Power Consumption Clustering

There is a growing trend in monitoring residential infrastructures to provide inhabitants with more information about their energy consumption and help them to reduce usage and cost. Device-level power consumption information, while a functionality in newer smart appliances, is not generally available to consumers. In electricity consumption disaggregation, Non-Intrusive Load Monitoring (NILM) refers to methods that provide consumers estimates of device-level energy consumption based on aggregate measurements usually taken at the main circuit panel or electric meter. The traditional NILM approach characterizes changes in the power signal when devices turn on or o, and it infers the consumption of different devices present in the home based on these changes. Generally, these NILM methods require training and models of the devices present in the home in order to function properly. Because of these challenges, much of the NILM literature does not address the actual energy disaggregation problem but focuses on detecting events and classifying changes in power. In this dissertation, we propose a relaxation to the traditional NILM problem and provide an unsupervised, data-driven algorithm to solve it. Specifically we propose Power Consumption Clustered Non-Intrusive Load Monitoring (PCC-NILM), a relaxation that reports on the energy usage of devices grouped together by power consumption levels. In order to solve the PCC-NILM problem, we provide the Approximate Power Trace Decomposition Algorithm (APTDA). Unlike other methods, APTDA does not require training and it provides estimated energy consumption for different classes of devices.