Development of Innovative Predictive Control Strategies for Nutrient Removal

This paper describes the development and deployment of an innovative hybrid nutrient removal controller that integrates machine learning with mechanistic modeling. Automated data pipelines facilitate the ingestion of SCADA, laboratory (LIMS), and meteorological data. Rigorous data cleaning and imputation processes ensure the high quality of input data. A soft sensor, using an auto-calibrated SUMO model, estimates dynamic influent flow and concentration profiles. Machine learning forecasters extend these profiles by 24 hours, while emulators swiftly simulate responses for optimization. The controller was deployed at four facilities, including Clean Water Services, where primary effluent orthophosphate loads were predicted with a mean average percent error of 12.20%. The flexible and modular structure of the controller, which integrates mechanistic and machine learning elements, exemplifies its impressive predictive optimization capabilities.