LAUSR

Abstract Mini-grids are the lowest-cost solutions for electrifying many homes and businesses in rural communities with low energy access. Estimates of the electricity demand of unelectrified customers are a crucial input to selecting mini-grid sites, projecting revenue, and sizing system components to provide adequate capacity while minimizing capital costs. Typical customer survey-based demand estimates for these communities — where there are no historical data — are not reliable, typically overpredicting demand. Here, we test a data-driven approach to demand prediction using survey and smart meter data from 1378 Tanzanian mini-grid customers. We found that models incorporating customer survey data into their predictions consistently out-performed a baseline model that did not. Our best-performing model, the LASSO, predicted daily electricity demand with a median absolute error of 66% and 37% for individual connections and mini-grid sites, respectively. Quantitative measures of variable importance show that most survey data are not useful for estimating demand. These results suggest that surveys should prioritize thorough inventories of prospective customers' currently-owned appliances instead of detailed demographic information or self-reported habits and plans. Pairing shortened questionnaires with smart meter data from preexisting mini-grids can improve estimates of initial customer electricity demand significantly compared to standard field practices.