Automated customer segmentation based on smart meter data with temperature and daylight sensitivity

Abstract

Utilities increasingly leverage knowledge on their customer's household characteristics in their energy efficiency programs. Examples of such characteristics include the number of persons per household, their employment status, or the type of dwelling they live in. This information allows utilities to personalize energy efficiency campaigns, which increases participation rates and in turn leads to larger energy savings and higher customer retention. However, gathering this information through surveys is costly and cumbersome. We therefore investigate the possibility to automatically infer household characteristics from electricity consumption data measured by an off-the-shelf smart meter. In this paper, we develop a method to determine the sensitivity of a household to outdoor temperature and the times of sunset/sunrise, and use this information to improve the performance of our household classification system. We further investigate the relevance of different features for such a system. Our evaluation is based on smart meter data collected at a 30-minute granularity in more than 4000 Irish households over a period of 75 weeks. The results show that we can improve accuracy by up to 2.3 percentage points using temperature and daylight coefficients. The characteristics floor area, type of dwelling, and percentage of installed energy-efficient light bulbs particularly benefit from temperature and daylight coefficients. Finally, we investigate the impact of the data granularity on the classification performance and show that semi-hourly or hourly data is required, as it performs on average 6.6 percentage points better than using daily consumption averages.