Seasonal-Adaptive Feature Engineering for Water Consumption Prediction: A Cross-Regional Stability Analysis

Abstract Forecasting water consumption is difficult because it depends on complex relation between demand and weather variables. Many traditional methods use fixed correlation and do not consider seasonal changes or non-stationary data. In this work, we propose a seasonal-adaptive feature engineering framework to solve these problems. The framework uses adaptive correlation thresholds, systematic construction of temperature features, and stability checking for better performance in different seasons. It also evaluates feature stability, applies adaptation protocols for new regions, and studies delayed effect and interaction of weather variables. Experiments show lower mean absolute error (13.8% less than baselines) and also competitive results with foundation models (4.3–6.5% better), but with much lower cost and higher feature stability (29.9% more). A last filtering step increases operational reliability. Cross-regional tests in five climate zones of Spain show adaptation time between 1.8 and 4.2 days. Results prove that adaptive feature engineering improves accuracy and robustness for water demand forecasting and helps utilities in resource planning.