Short‐Term Memory and Regional Climate Drive City‐Scale Water Demand in the Contiguous US

Gaining insights into current and future urban water demand patterns and their determinants is paramount for water utilities and policymakers to formulate water demand management strategies targeted to high water-using groups and infrastructure planning strategies. In this paper, we explore the complex web of causality between climatic and socio-demographic determinants, and urban water demand patterns across the Contiguous United States (CONUS). We develop a causal discovery framework based on a Neural Granger Causal (NGC) model, a machine learning approach that identifies nonlinear causal relationships between determinants and water demand, enabling comprehensive water demand determinants discovery and water demand forecasting across the CONUS. We train our convolutional NGC model using large-scale open water demand data collected with a monthly resolution from 2010 to 2017 for 86 cities across the CONUS and three K & ouml;ppen climate regions-Arid, Temperate, and Continental-utilizing this globally recognized climate classification system to ensure a robust analysis across varied environmental conditions. We discover that city-scale urban water demand is primarily driven by short-term memory effects. Climatic variables, particularly vapor pressure deficit and temperature, also stand out as primary determinants across all regions, and more evidently in Arid regions as they capture aridity and drought conditions. Our model achieves an average R2 ${R}<^>{2}$ higher than 0.8 for one-month-ahead prediction of water demand across various cities, leveraging the Granger causal relationships in different spatial contexts. Finally, the exploration of temporal dynamics among determinants and water demand amplifies the interpretability of the model results. This enhanced interpretability facilitates discovery of urban water demand determinants and generalization of water demand forecasting.