AI-enabled CPS operational framework with daily-retrained model bank for chemical resource optimization in drinking-water treatment.

This study presents an AI-enabled cyber-physical system (CPS) operational framework for chemical resource optimization in drinking-water treatment. The framework integrates daily model-bank retraining, 24-h-ahead hourly coagulant dose concentration setpoints, influent-flow-based feed-rate conversion, SCADA-PLC actuation, and command-monitor tag verification within a pilot-scale closed-loop control system. A side-by-side pilot-scale comparison was conducted for approximately 15 weeks using two parallel treatment trains: a conventional control train operated with twice-daily jar-test-based setpoints and a CPS-based experimental train operated with hourly model-based setpoints. The CPS generated 1860 hourly model-selection and setpoint-generation records during the analyzed period. Total diluted coagulant consumption decreased from 536.62 L in the conventional train to 490.83 L in the CPS train, corresponding to an 8.53% reduction (45.79 L). A paired analysis of 77 complete operating days confirmed that the reduction was statistically significant (paired t-test, p = 0.0042). Based on 1870 paired hourly settled-water turbidity records, the CPS train achieved 99.25% compliance with the 1 NTU process-level optimization benchmark, compared with 96.95% for the conventional train, while no statistically significant difference in paired daily mean turbidity was observed between the two trains (Wilcoxon signed-rank test, p = 0.173). Model-bank analysis showed that Random Forest was selected most frequently (n = 1339), while regime-stratified analysis indicated generally stable prediction performance across the observed raw-water turbidity range. These results demonstrate the potential of AI-enabled CPS operation to translate predictive modeling into field-level chemical resource management while maintaining overall process-level turbidity performance. Future improvements should incorporate event-triggered retraining, real-time turbidity surge detection, and rate-of-change constraints to improve robustness under abrupt source-water disturbances.