Smart meter data analytics: prediction of enrollment in residential energy efficiency programs

Massive rollout of residential smart meters has spurred interest in processing the highly granular data available from these devices. Whereas the majority of smart meter data analytics is devoted to characterization of household electric appliances and their operational schedules, little work has been done to leverage these data to predict household propensity to enroll in energy efficiency and demand response programs. The state-of-the-art methodology for household enrollment prediction involves measurable household characteristics (e. g., age, household income, education, presence of children, average energy bill) and a multivariate logistic regression that connects these predictor variables with the probability to enroll. Unfortunately, the prediction accuracy of this method is just slightly better than 50%, and the required household data are not freely available to utilities/program contractors. We developed a new method for prediction of household propensity to enroll using only hourly electricity consumption data from households' smart meters, collected over twelve months. The method implements advanced machine learning algorithms to reach an unprecedented prediction accuracy of about 90%. This level of accuracy was obtained in our study of a US West Coast behavior-based residential program.