Automatic standard building category classification from smart meter data – A supervised learning approach

Abstract

Increased availability of smart meter data offers better insight into buildings’ electricity usage. By classifying smart meter data by building type and presence of heating appliances, we can efficiently gain metadata about the buildings that is useful for research, grid planning, and energy efficiency policy employment. However, current smart meter classification approaches are largely based on limited datasets and building classes, or on unsupervised methods that don’t align with standard building categories and offer limited control over grouping. This article presents a supervised automatic building category classification approach for labelling smart meter data from buildings into standard building categories in the Norwegian building regulations (TEK17), and whether they have electric heating or not. 82 novel physics-based domain features are presented which can be extracted from any hourly electricity smart meter data series from buildings with a duration of months-years. The features are specifically designed to identify the building and heating type of a smart meter data series by capturing patterns such as seasonality, daily usage trends, similarities with standardized building load profiles, temperature dependency, and other domain-specific characteristics. The classification approach is trained and tested on a large dataset of 2724 buildings from 12 different building categories, both residential and non-residential, and correctly identifies the heating type and building category of unseen Norwegian smart meter data from buildings in 84 % of the test cases. The approach is generalizable to meter data from other Norwegian buildings and is also tested on buildings from other climate zones. The proposed method for smart meter data classification is proven to have high accuracy and applicability for extracting metadata for both residential and non-residential buildings in Norway.