Time Will Tell: Using Smart Meter Time Series Data to Derive Household Features and Explain Heterogeneity in Pricing Programs

The recent nationwide adoption of smart meters provides a new source of rich data about individual household electricity consumption. Data science techniques can extract a variety of high temporal resolution, household-specific features from the hourly electricity time series itself and in combination with other readily available relevant information, like weather or census data. This allows us to observe or estimate important characteristics of household electricity use that were previously unobservable. These characteristics and household features have the advantage of representing the actual choices and behaviors of households, which can differ substantially from stated preferences and subjective information from traditional survey or interview methods. The use of data-derived numerical features is common in machine learning but not in traditional engineering and econometric models. In this paper, we use this technique to help answer a question that is important to program implementers: who is likely to respond to my program? In other words, program implementers typically use rules of thumb to identify target households (e.g., top 25% of usage from monthly bills), and program evaluation typically only identifies the overall average effect. Understanding the heterogeneity of program response can help shed light on how and why different households respond in different ways, allowing implementers to focus on specific groups, tailor programs to speak to the way that households actually behave, and predict the effectiveness of future programs for portfolio planning purposes. We identify household specific features that explain heterogeneity in response to experimental time-of-use and critical-peak-pricing electricity rates. The experiment was performed using randomized controlled trials with treatment groups encouraged to enroll into new rates. The household responses of interest are metrics related to energy consumption during peak hours. We use numerical features derived from pre-treatment smart meter data as covariates in an instrumental variable regression, and we find, promisingly, that they can explain considerable heterogeneity of treatment outcomes. These results lay the groundwork for using smart meter data along with data science techniques to improve program uptake, evaluation, design, and targeting.