Assessing the feasibility of temperature response models for residential energy consumption behavior analysis

• We compare different methods for modeling temperature response functions. • Lagged dependent variable models perform the best from assumption testing. • Estimated model parameters are similar across methods. • We recommend using stationary lagged dependent variable models. Temperature response functions (TRFs) are models used to predict a household’s daily energy consumption from outdoor temperature and other control variables. In recent energy equity work, TRFs are modeled as piecewise linear regressions and interpreted for inference instead of prediction. Inference of regression parameters assumes the models help identify when and how intensely households use climate controls. Utilities apply this knowledge to better understand which consumers show energy limiting behaviors (e.g., reducing usage to reduce cost). However, past studies have not evaluated whether TRF models satisfy regression assumptions. By testing these assumptions for different model formulations, we learn how to adapt the model, which can deepen our understanding of energy disparities. We analyze data from approximately 100,000 households in Tallahassee, Florida, US and 350,000 households from the US mid-Atlantic region. We found that including a lagged dependent variable as a predictor significantly improves conformity with all linear regression assumptions compared to all other formulations. Further, most parameter estimates are similar across methods, and most differences observed are expected due to the model formulations. For these reasons, we suggest using the stationary lagged dependent variable method for modeling TRFs. Overall, our work can help energy utilities better evaluate energy assistance programs and the quality of TRF analyses.