Full-process monitoring indispensability for two-stage PN/A effluent prediction: Interpretable machine learning from...

Multi-stage partial nitritation-anammox (PN/A) processes offer transformative advantages for sustainable nitrogen removal, yet their engineering application remains constrained by operational instability and a critical translational gap: existing machine learning (ML) models are predominantly built on laboratory-scale data, leaving their predictive reliability in full-scale scenarios largely unverified. Moreover, systematic quantification of full-process parameter synergy and simultaneous accurate prediction of multi-dimensional effluent quality are critically lacking. To address these gaps, this study develops an interpretable ML framework grounded in a unique four-year engineering-scale dataset from a full-scale two-stage PN/A facility. Three input feature sets are designed to evaluate the necessity of full-process monitoring. Key findings reveal that (i) the full-process feature set (incorporating both PN and Anammox stage parameters) delivers superior predictive accuracy for effluent chemical oxygen demand (COD), NO₃^--N, and total inorganic nitrogen (TIN) (test R² = 0.941, 0.944, 0.837), establishing that holistic monitoring is indispensable for multi-stage prediction; (ii) no universal optimal model exists-XGBoost excels for COD, LightGBM for NO₃^--N, and Gradient Boosting for TIN-demonstrating target-aware model selection is essential; (iii) PDP quantification yields objectively determined, flexible operating windows for key parameters, transforming black-box predictions into evidence-based engineering decisions. This study provides the first data-driven demonstration, grounded in long-term real-world engineering data, that full-process monitoring is not merely beneficial but indispensable for accurate multi-stage PN/A prediction. The transferable interpretable ML workflow establishes a robust methodological paradigm for intelligent control of complex environmental bioprocesses.