Application of artificial neural networks to effluent phosphate prediction in struvite recovery

In advanced wastewater treatment plants (AWWTP), the recovery of phosphorus (P) has become a recent focus of the wastewater engineering industry. The potential economic savings behind improved sludge management and the control of struvite encrustation in AWWTP are two of the primary driving forces behind this. Process control of phosphorus (struvite) recovery systems has only been partially successful because: (1) key control variables have yet to be identified and (2) there is no adequate performance evaluation model that is applicable to struvite recovery technologies. In process control, two different types of modeling are most commonly seen: mechanistic and "black-box" style models. In recent years, varying models have been developed to try to predict the formation of struvite in both sludge digestion process lines and P-recovery technologies designed for struvite removal. All of these are strictly mechanistic models, based on either the chemical equilibrium of the system or the associated kinetic parameters, with varying degrees of complexity. Artificial neural networks (ANN), as a type of black-box modeling, have seen limited application in wastewater treatment with regards to phosphate recovery. The analysis of several historical daily operational databases evaluated the predictive ability of two mechanistic and one ANN models. It was determined that the newly developed ANN model was not site specific and had the highest predictive ability of the three. This would be beneficial for the development of an automated control system for struvite removal package treatment processes. © 2007 NRC Canada.