Learning Systems for Electric Consumption of Buildings

Abstract

Individual appliances' electricity consumption is automatically disaggregated from a single custom metering system on the main feed to an occupied residential building. A data acquisition system samples voltage and current at 100 kHz, then calculates real and reactive power, harmonics, and other features at 20Hz. A probabilistic eventdetector using the generalized likelihood ratio (GLR) matches human-labeled events to the time-series of features. Machine-learning classification was most successful with the 1-nearest-neighbor algorithm, correctly identifying 90% of the laboratory-generated training events and 79% of validation examples. The challenge of obtaining adequate training data for the real-world home leads to the development of the Wire Spy, a wirelessly-networked event detector with an inductive sensor which clamps to the cable of an appliance.