Dynamic Restoration of Collapsed Anammox Biofilm Systems: Integrating Process Optimization, Microbial Community Succession, and Machine Learning-Based Prediction

The majority of extant studies concentrate on the reactivation of dormant Anammox biomass or the recovery of activity under specific storage conditions. Research on rehabilitation strategies for anaerobic ammonium oxidation (Anammox) systems is limited, with the exception of research on inhibitory factors. The recovery characteristics of biofilm systems after collapse induced by varying degrees of ammonia-nitrogen and small-molecular organic compound composite shocks have not been thoroughly elucidated. This study addresses the collapse of Anammox biofilm systems caused by sodium acetate inhibition through multi-phase rehabilitation strategies, stoichiometric analysis, and microbial community succession dynamics. Two regression algorithms—Support Vector Regression (SVR) and eXtreme Gradient Boosting (XGBoost)—were employed to construct predictive models for Total Nitrogen Removal Efficiency (TNRE) and Total Nitrogen Removal Rate (TNRR) in the CANON system, with model performance evaluated via coefficient of determination (R2) and root mean square error (RMSE). Results demonstrated that after terminating moderate-to-high sodium acetate dosing (300 mg/L and 500 mg/L), reactors R300 and R500 achieved TNRE recovery to 57.98% and 58.86%, respectively, and TNRR of 0.281 and 0.275 kgN/m3·d within 60–100 days, indicating the reversibility of high-concentration sodium acetate inhibition but a positive correlation between recovery duration and inhibition intensity. Microbial community analysis revealed that Planctomycetota (including Candidatus_Kuenenia) rebounded to 46–49% relative abundance in R100, synchronized with TNRE improvement. In contrast, R300 and R500 exhibited ecological niche replacement of denitrifiers (Denitratisoma) and partial TNRE restoration despite enhanced performance. Model comparisons showed SVR outperformed XGBoost in TNRE prediction, whereas XGBoost demonstrated superior TNRR prediction accuracy with R2 approaching 1 and RMSE nearing 0, significantly surpassing SVR. This work provides critical insights into recovery mechanisms under organic inhibition stress and establishes a robust predictive framework for optimizing nitrogen removal performance in CANON systems.