Elucidating drivers of full-scale struvite recovery system performance with thermodynamic and data-driven modeling

Sidestream struvite recovery systems can significantly enhance the performance of mainstream biological phosphorus (P) removal processes by reducing intraplant P cycling from solids handling return flows. While adoption of struvite recovery systems has increased rapidly over the last decade, system performance continues to be highly variable. Currently available modeling tools are not yet capable of diagnosing and predicting upsets. This study presents novel data-driven modeling approaches to evaluate and predict the performance of struvite recovery reactors. More than 5 years of historical data from a full-scale fluidized bed struvite crystallization reactor at a municipal water resource recovery facility were analyzed to identify the factors controlling P removal from solution (Conversion) and the fraction of P recovered as struvite (Yield). Reactor performance analysis indicates that struvite loss in the crystallizer effluent can be significant, accounting for up to 25% of mainline P mass flow, highlighting the need to understand and improve struvite Yield. Data driven models that account for fluidized bed recycle flow and composition were far more accurate for both Conversion (R² = 0.83) and Yield (R² = 0.95) than models based on wastewater feed and chemical dosing alone. Feature importance analysis indicated that feed phosphorus concentration and chemical dosing rates were most influential for Conversion prediction, and the concentration of particulate P in the crystallizer recycle stream was critical to predicting struvite Yield. Data driven models were then used to illustrate how operational changes can drive improvements in P Conversion and struvite Yield.