Scalable Online Machine Learning for Water Quality Optimization and Carbon Source Reduction in Wastewater Treatment Plants

Abstract

Effective real-time control of wastewater treatment plants (WWTPs) remains challenging due to highly variable influent conditions, while regulation of external carbon sources (ECS) is critical for sustaining nitrogen removal and meeting increasingly stringent effluent standards. This study developed a systematic framework integrating online learning (OL) and reinforcement learning (RL) to optimize effluent quality and ECS dosing in a full-scale WWTP. Based on one year of high-frequency sensor data, Transformer (TF) was identified as the best-performing model among six machine learning methods, with a mean R² of 0.91 across multiple effluent indicators. SHAP was employed to identify key features, which informed the design of a key-indicator-based time-decayed online learning (KTOL) strategy, enabling the Transformer model to maintain stable predictive performance under dynamic operating conditions. Building on this, RL enabled real-time ECS regulation. During a 10-day practical test, KTOL-TF-RL outperformed conventional feedback control by improving effluent compliance and reducing ECS usage by 21.8%. Monte Carlo simulations demonstrated robustness against sensor distortion and malfunction. Overall, the KTOL-TF-RL framework demonstrates the potential of intelligent control as a promising pathway for achieving cost-effective and resilient wastewater treatment.