Daily water demand forecasting: Comparing AI models with SHAP-optimized features

Abstract Accurate water demand prediction is critical for infrastructure stability and resource optimization, yet short-term forecasting remains challenging due to high volatility from meteorological, seasonal, and socio-temporal factors (e.g., holidays). To address this, we collected 402 days of urban water demand records augmented with web-scraped meteorological and temporal features. Through SHapley Additive exPlanations (SHAP) analysis, we identified and retained high-impact features (e.g., maximum temperature, day-of-week) while eliminating redundant variables (e.g., minimum temperature, cloudy conditions), achieving a 22% reduction in feature dimensionality with a 0.16 percentage point improvement in MAPE across all AI models. We systematically compared 7 machine learning models and 3 deep learning models against an ARIMA baseline model using four performance metrics. The results indicate that deep learning methods have significant advantages in prediction accuracy, while machine learning models have certain shortcomings in predicting time series. The organic combination of interpretable feature selection in machine learning and precise prediction in deep learning provides actionable insights for utilities.