Resolution Tricks and Disaggregation Tools for Smart Power Metering

Abstract

A nonintrusive load monitor (NILM) aims to solve the energy disaggregation problem by incorporating power system analysis, signal processing, and machine learning. This thesis addresses two problems present in state-of-the-art nonintrusive load monitoring research. First, the ability of existing nonintrusive load monitoring techniques and data to generalize is very low, so any data collected for model training needs to be domain-specific. For this reason, this work explores the limits of power signal processing used by deployable NILMs. Secondly, load electrical behavior is almost always assumed to be stationary. Thus, this work presents Adaptive NILM, a set of feature space selection and classification tools useful for nonintrusive load monitoring with limited training data when load operation drifts over time. These techniques are synthesized into a new NILM software package that allows for high-level automation of resolution tracking, feature space evaluation, and adaptive classification. A new NILM hardware implementation, capable of wirelessly integrating data from distributed sensors, is described and demonstrated with case studies.