Real-Time DBP Prediction in Smart Water Systems: Methodological Frontiers in AI Application

Disinfection byproduct (DBP) control in drinking water treatment is gradually shifting from reactive monitoring to proactive management. Conventional DBP assessment relies largely on periodic sampling and laboratory analysis, which provide accurate but delayed information. This article examines how AI-based DBP prediction can move beyond high internal accuracy toward deployable intelligence in smart water systems. Key barriers include limited model generalizability and a mismatch between research-grade inputs and data available during routine operation. We propose a modular framework that links online water-quality and operational variables with spectroscopic information and periodically updated molecular characterization, while recognizing surrogate-derived inputs as an additional source of uncertainty. We further outline a three-phase deployment pathway from soft sensing and early warning to human-in-the-loop decision support, and ultimately to digital twin-assisted control. Across these stages, deployment-aligned validation, uncertainty assessment, model updating, and explainability are emphasized. The methodological frontier is therefore not simply more accurate prediction, but prediction that remains reliable and operationally useful under changing treatment conditions.