Sustainable coagulant optimization via inverse filtration control and ensemble learning in a full-scale drinking...

Excessive reliance on conservative coagulant dosing to buffer raw water fluctuations in drinking water treatment plants (DWTPs) frequently leads to chemical redundancy and elevated carbon emissions. This study proposed a novel strategy shifting the primary turbidity control boundary downstream to the sand filter, termed Inverse Filtration Control (IFC), to optimize coagulant efficiency. This strategy leveraged the inherent particulate retention capacity across the treatment continuum and optimized the nonlinear coagulant demand. This research developed a cascaded Boosting-based generalized regression neural network (Boosting-GRNN) to reliably execute this strategy and effectively decouple global water quality trends from high-frequency fluctuations. Leakage-free chronological evaluation over the continuous future block yielded R² values of 0.9462 and 0.9156 for coagulant dosage and sand filter effluent turbidity predictions, respectively. A retrospective model simulation, based on 1,704 daily operational records (2021-2025) and evaluated over the continuous future block, estimated a potential mean coagulant dosage reduction of 10.1% at a sedimentation turbidity target of 1.03 NTU. Under the historical annualization assumptions, this corresponds to approximately 271 t of PAC and 298 t of CO₂-eq per year. SHAP analysis further indicated distinct patterns of model feature reliance across the two predictive modules, providing a transparent basis for evaluating the digital twin framework. This study bridged the gap between isolated unit operations and coordinated inter-unit control, and offered a transparent, verifiable digital twin roadmap for sustainable and low-carbon water utility operations.