Evaluating the household-level climate-electricity nexus across three cities through statistical learning techniques

As the climate crisis intensifies, rising temperatures and increased frequency of extreme events are likely to strain the electricity system. This will be particularly disastrous if the grid is unprepared for the climate-induced shifts in electricity demand that will result from increased temperatures. Recently, the use of data-driven modeling has emerged as a way to predict these climate-induced changes in electricity demand, however, much of the work has focused on entire sectors or regions. Here, we focus on the impact of climatic variables on hourly household electricity use for air conditioning. Our goal was to determine the best model for predicting the air conditioning use based on climate variables, as well as use that model to extract insights related to the household-level climate-electricity nexus. Using smart meter data from three US cities (Austin, Texas, Ithaca, New York, and San Diego, California), we tested seven different models of varying complexity. Ultimately, Bayesian additive regression trees (BART) was selected as the best model across all three cities (NRMSE ranged between 0.085 and 0.250). Additionally, we found that while the majority of the climate variables were important, relative humidity was the most important variable in each city. Given that air conditioning tends to drive non-base electricity demand in the summer, understanding these nuances in the climate-electricity nexus as it applies to air conditioning is critical for building a resilient grid.