Leveraging interpretable machine learning models with hyperparameter optimization algorithms for domestic water demand forecasting in arid climates

Kuwait, an extremely arid country, lacks the surface water reserves that many other nations rely on. However, optimized and explainable machine learning (ML) models have not been used to predict water consumption in extremely arid regions. Here, three ML models were used to model Kuwait’s monthly water demand; Bayesian Optimization (BO) was integrated to enhance the performance of the ML models. Furthermore, the SHAP technique was employed to interpret the outcomes of the ML models. The study used historical water-demand data from 1993–2018 and therefore does not capture post-2018 urban water-consumption behavior. The analysis results demonstrated a coefficient of determination (R2) of 0.97 and a mean absolute error (MAE) of 353.31 million imperial gallons (MIG) during the validation period. Seasonal evaluations revealed that the ML models were effective during the summer season. These findings demonstrate the potential of optimized ML for urban water-demand forecasting in arid regions.