Unsupervised Identification of Electrical Loads from Aggregate Power Measurements

Non-Intrusive Load Monitoring (NILM) provides appliance level operating information from a single power sensor at the service entry point to a facility. This promises significant benefits to the smart grid including fine grained energy usage information, condition based maintenance diagnostics, and anomaly detection. However, in order for these systems to gain wide adoption the algorithms used to dissagregate the bulk power signal into individual appliance transients need to handle a wide variety of load types as well as a constantly changing load environment. Classical supervised learning approaches struggle to collect enough training data and do not have the flexibility to identify new appliances or track appliances that change or degrade over time. This paper introduces an unsupervised learning model that can be used to dynamically identify and track loads in an aggregate power signal with no a-priori knowledge of load characteristics.