Predictive segmentation of energy consumers

This paper proposes a predictive segmentation technique for identifying sub-groups in a large population that are both homogeneous with respect to certain patterns in customer attributes, and predictive with respect to a desired outcome. Our motivation is creating a highly-interpretable and intuitive segmentation and targeting process for customers of energy utility companies that is also optimal in some sense. In this setting, the energy utility wants to design a small number of message types to be sent to appropriately-chosen customers who are most likely to respond to different types of communications. The proposed method uses consumption, demographics, and program enrollment data to extract basic predictive patterns using standard machine learning techniques. We next define a feasible potential assignment of patterns to a small number of segments described by expert guidelines and hypotheses about consumer characteristics, which are available from prior behavioral research. The algorithm then identifies an optimal allocation of patterns to segments that is feasible and maximizes predictive power. This is formulated as maximizing the minimum enrollment rate from across the segments, which is then expressed as solving a mixed-integer linear-fractional program. We propose a bisection-based method to quickly solve this program by means of identifying feasible sets. We exemplify the methodology on a large-scale dataset from a leading U.S. energy utility, and obtain segments of customers whose likelihood of enrollment is more than twice larger than that of the average population, and that are described by a small number of simple, intuitive rules. The segments designed this way achieve a 2–3× improvement in the probability of enrollment over the overall population.