Assessing urban water consumption dynamics using machine learning and spatial variables under normal and pandemic conditions: A case study

Abstract

This study developed a spatially explicit framework to estimate household water consumption (HWC) and the number of household water subscribers in each consumption class (NHWSCC) using land surface temperature (LST), land value, and elevation. Three machine-learning (ML) models were applied: symbolic regression (SR), an extreme learning machine (ELM), and an ELM hybridized with the artificial bee colony algorithm (ELM-ABC). The spatial predictors supported the generation of HWC maps for normal and pandemic-period conditions in Tabriz, Iran. Model performance was generally more stable at grid resolutions of 600–700 m, although the achievable resolution varied among consumption classes. The models captured spatial patterns for many low- and intermediate-consumption classes but performed less well for the highest-consumption categories. SR sensitivity analysis identified LST as the most influential predictor of NHWSCC. Differences in predictor sensitivity between the study periods suggest changes in the relationships between spatial characteristics and household water use. Observed annual consumption increased in lower-land-value areas and decreased in higher-land-value areas during the pandemic period. These findings support the use of readily available spatial proxies for within-city water-consumption mapping, while recognizing the limitations of the predictors and validation design.