Continually evolving time-constant networks enable proactive control of coagulation for low-cost drinking water treatment

Effective chemical dosing control in drinking water treatment plants (DWTPs) is essential for maintaining effluent quality while minimizing resource consumption under increasingly dynamic and uncertain conditions. This study proposes a data-driven control framework integrating Liquid Time-Constant neural networks (LTC) model with reinforcement learning (RL) for real-time optimization of polyaluminum chloride (PAC) dosage. The LTC model achieved high predictive accuracy (R² = 0.936 for turbidity, R² = 0.918 for pH) with compact structure and low memory usage (888.67 MB). Robustness evaluations show stable performance under 10% sensor noise (R² > 0.90) and tolerance to 20% data loss, with favorable transferability across distinct DWTPs (R² > 0.85). Its incremental learning enables continuous adaptation with 94.9% less training data and 43.3% faster convergence than periodic retraining. Explainability analysis identifies PAC dosage and influent turbidity as dominant predictors with complex interactions. In offline surrogate-based control simulations driven by real operational data, the RL-based dosing strategy reduced PAC consumption by 13.4% versus conventional strategies, while maintaining predicted effluent turbidity below 0.1 NTU. These findings indicate the potential of the framework as a safety-constrained decision-support strategy for improving chemical efficiency, stability, and intelligent decision-making in DWTP operations.