Discriminant analysis classification of residential electricity smart meter data

The objective of this study is to apply machine learning classification to predict building characteristics from electricity smart meter data for the purpose of building stock characterization. Given that there are no publicly available large-scale residential electric smart meter data sets with detailed building characteristics, an open-source virtual smart meter (VSM) data set is used. The VSM data consists of electricity consumption profiles for 200,000 homes with 21 known characteristics, which are used to train predictive models with linear discriminant analysis (LDA). The classification accuracy (CA) is determined for a variety of scenarios where the meter data aggregation and period are varied. The CA depends on the parameter to be classified (the class), the number of data points per building (the features) and the number of buildings used for classification. Reliable classification results are obtained when the number of buildings exceeds the number of features by a significant margin. An application of the developed predictive models to a small data set of 30 real houses illustrates the usefulness of the method but also the challenges in achieving a generalized model with virtual data.

Uncontrolled Keywords

Residential, Smart meter data, Supervised machine learning, Linear discriminant analysis, Building stock characterization, Classification studies