Development of Data-Driven Models for Performance Prediction and Chemical Dosing of a Full-Scale Controlled Phosphorus Precipitation Reactor

Controlled struvite precipitation of postanaerobic digestion streams enables both phosphorus removal and fertilizer generation. Optimizing process control for these reactors remains difficult due to the lack of real-time measurements for key variables, including influent orthophosphate (OP) concentration, effluent OP concentration, and OP removal. This study applied data-driven modeling to predict these variables for a struvite precipitation reactor (MagPrex) at a water resource recovery facility in Denver, CO, USA. The models were used to recommend a magnesium chloride (MgCl 2 ) dose to meet a target effluent OP concentration. Ridge regression provided the most accurate influent OP (test normalized root mean squared error, RMSE, of 9.3%) and OP removal (8.2%) predictions, while XGBoost performed best for effluent OP (9.1%). Effluent OP predictions were used to determine whether the system was underdosed, ideal dose, or overdosed, which supported directional MgCl 2 set point changes. In simulation, the model correctly recommended a directional set point change 97.7% of the time, corresponding to 3.0% MgCl 2 savings. Full-scale deployment yielded lower accuracy (88.5%), which was attributed to nonsteady state conditions such as reactor startup and limited historical data at high OP concentrations. These findings demonstrate the technology and utility potential and limitations of data-driven decision support for struvite reactor control.