Integrating socio‑environmental and demographic factors in machine learning forecasts of water consumption

Abstract

Reliable forecasting of water consumption is essential for water resources management because it enables policymakers and utilities to balance supply and demand effectively. This study examines seasonal (three-month horizon) water consumption in Isfahan Province, Iran, using a modeling table of 528 seasonal observations spanning 24 subscriber categories over the period 2016–2021. The predictor set includes subscriber category, year, season, temperature, rainfall, a COVID-19 indicator, death rate, and birth rate. The analysis compares five machine learning algorithms, namely Random Forest (RF), Decision Tree (DT), Gradient Boosting (GB), eXtreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost), under a random 80/20 holdout split with 7-fold cross-validation on the training partition. All five models achieve high predictive accuracy, but their relative performance depends on metric choice, computational cost, and sensitivity to low-consumption subscriber categories.