<![CDATA[The objective of this paper is to improve the forecasting of monthly water consumption under climate variability by combining ensemble and hybrid modelling with a season-aware design. Monthly consumption and meteorological data from 2003 to 2024 were utilized in this study. Four models were evaluated: (i) a stacking ensemble with STL-trend plus residual learning; (ii) a hybrid machine-learning–physics model with differentially-evolved weights; and (iii–iv) season-specific stacked models for wet and dry periods. Robustness was assessed with time-aware validation and residual diagnostics (Shapiro–Wilk, Breusch–Pagan, Durbin–Watson, Ljung–Box). The findings indicate that across models, ensembles captured nonlinear climate–demand variations while maintaining linear structure. The ensemble and hybrid model achieved strong accuracy with low errors while the season-specific models attained high fit (wet R²≈0.998; dry R²≈0.991) with stable residual behavior. Sensitivity to temperature and humidity aligns with expected physical behavior. Precipitation shows a diminishing-returns effect on water use, where moderate rainfall leads to higher consumption, while heavy rainfall tends to reduce demand. The framework innovatively combines decomposition-assisted stacking, physics-informed hybridization, and seasonal ensemble modelling. Overall, the approach provides highly accurate, interpretable, and climate-aware water demand forecasts for tropical regions, offering a practical basis for utility-scale implementation.]]>
