Comparison of interpretable machine learning models and mechanistic model for predicting effluent nitrogen in WWTP

Abstract

Accurate prediction of effluent nitrogen is crucial for optimizing operations of wastewater treatment plants (WWTPs). This study systematically compared the performance of a mechanistic model, i.e., the Activated Sludge Model (ASM), with six machine learning (ML) models in predicting effluent total nitrogen (TN), by using one year of high-resolution full-scale operational data obtained from a municipal wastewater treatment plant (WWTP). Notably, the integration of Shapley Additive Explanations (SHAP) into the ML models enabled transparent interpretation of model predictions. ASM was dynamically calibrated through sensitivity analysis, which identified key parameters such as μ_AOB and η_OHO,anox related to nitrification and denitrification. Despite capturing TN trends, the ASM model showed limited accuracy (R² = 0.26 for training and 0.06 for validation). In contrast, ML models, particularly Random Forest, XGBoost, and LightGBM, demonstrated superior predictive performance (highest R² = 0.79, lowest MRE = 7.5 %). The ML can directly learn complex relationships from a large amount of running data, while ASM relies on simplified mechanism equations and has difficulty reflecting the dynamic changes in actual operation. SHAP analysis further revealed that return sludge rate, MLSS, influent ammonia, and nitrate concentrations were the most influential features determining TN removal. These findings were consistent with the ASM sensitivity analysis, verifying the ML model's capacity to uncover biologically meaningful insights. This study demonstrated that interpretable ML models not only outperformed traditional ASM in prediction accuracy but also provide transparent and actionable explanations, marking a significant advancement in the application of AI for wastewater process modeling.