Predictability of energy use in homes

Predictability of home energy usage forms the basis of many home energy management and demand-response systems. While existing studies focus on designing more accurate prediction algorithms, a comprehensive energy management solution requires a broad understanding of prediction accuracy at different granularities, for example appliance and home, as well as different time horizons, for example an hour, day, or week into the future. In this paper, we undertake an analysis of predictability of power draw of appliances and whole-home energy consumption at four different time horizons: an hour, a quarter-day, a day, and a week in the future. Our analysis presents two research contributions. Our first contribution is a diverse dataset, GreenHomes, that includes appliance power draw and whole-home energy consumption data from seven homes across three states in the United States over a two-year period. Our second and primary contribution is a set of insights into the predictability of home energy usage. We show that simple statistic-based algorithms perform as well as sophisticated machine learning algorithms and time-series based predictors. These simple algorithms can considerably reduce the computational need for large-scale predictive analysis of home energy data. We also show that appliance-level power draw is more predictable than whole-home energy consumption at shorter time horizons while home-level energy consumption is more predictable at longer time horizons. Finally, we show that there is large variation in predictability across homes. This variation may be attributed to home type and points to the need for personalized energy management systems.