Machine learning-assisted prediction and identification of key factors affecting nitrogen metabolism for aerobic...

To achieve higher denitrification efficiency with reduced energy consumption in aerobic granular sludge (AGS) system, a systematic evaluation of the carbon and nitrogen metabolism process for AGS under different stage is essential. Herein, this study established the prediction models via interpretable machine learning (ML) for simulating the nitrogen metabolism by using 312 sets of data collected from four reactors with different kinds of AGS. The results indicated Gradient Boosting Decision Trees (GBDT) achieved R² values of 0.729, 0.875, and 0.807, respectively by selecting water temperature, carbon source components and particle size as input factors and NH₄⁺-N, NO₂^--N, and NO₃^--N as prediction targets. Furthermore, Shapley Additive Explanations (SHAP) analysis was used to make global and local interpretations of the GBDT models. The global explanation revealed that particle size of AGS and carbon source of component 4 (C4) with the Ex/Em at 225/335 nm significantly influenced denitrification process. And local analysis results proved that enhanced nitrogen removal performance is attainable when the DX50 and DX10 range between 500-600 and 200-500 μm and the F_max value of C4 exceeds 0.2 R.U. These findings provide an effective tool for evaluating nitrogen removal performance and identifying the key factors influencing nitrogen metabolism in AGS systems.