Residential classification using limited sample size consumption data with associated metadata
University of New Brunswick
The ongoing deployment of residential smart meters in numerous jurisdictions will inevitably lead to a significant amount of electricity consumption data becoming available. This information presents a valuable opportunity to utilities for mining and analysis of load patterns. Household load shapes can reveal significant differences among large groups of households in the magnitude and timing of their electricity consumption. However, limitations in dataset sample size and the associated descriptive data (i.e. metadata) makes it difficult for residential classification problems. It is often desirable to classify residential customers based on their energy consumption profiles and give meaningful insight into the amount of energy these customers use over a specified period. The research work presented in this thesis focusses on specific classification requirements of two limited sample size datasets which have differing metadata associated with each. For Dataset 1, clustering techniques, statistical analysis and machine learning (ML) algorithms were used to analyze the difference between residential consumption profiles based only on lot size and home price. Strategic groupings such as more expensive homes on large lots versus lower cost homes on small lots have shown to have some separability in terms of load profiles. Dataset 2 has only heating source type associated with it and the main focus was to differentiate homes that utilize one of three heating types: electric baseboard (BB), air-source heat pumps (HP) or mini-split heat pump (MS). This will help provide information on expected consumption patterns for the presence of a given heating type. Statistical and ML techniques were again applied to this problem and performance is assessed using numerous load profile and weather features.