A generalizable approach to imbalanced classification of residential electric space heat

Changes in climate and energy technologies motivate a greater understanding of residential electricity usage and its relation to weather conditions. The recent proliferation of smart electricity meters promises an influx of new datasets spanning diverse cities, geographies, and climates worldwide. However, although analytics for smart meters is a rapidly expanding field of research, issues such as generalizability to new data and robustness to data quality remain underexplored in the literature. We characterize residential electricity consumption patterns from a large, uncurated testbed of smart electricity meter data, revealing challenges in adapting existing methodologies to datasets with different scopes and locations. We propose a novel feature—the proportion of electricity used below a temperature threshold—summarizing a household’s demand-temperature profile that is productive for identifying electric primary space heating in a smart meter data set of Chicago single-family residences. Weighted logistic regression using the proportion of electricity consumed below a selected low temperature mitigates difficulties of the dataset such as skew and class imbalance. Although the limitations of the dataset restrict some approaches, this experiment suggests advantages of the feature that can be adapted to study other datasets beyond the identification of space heating. Such data-driven approaches can be valuable for knowledge distillation from abundant, uncurated smart electricity meter data.