Reconciling SHAP attribution with wrapper-based subset selection for efficient effluent NH3-N prediction in full-scale industrial food wastewater treatment

Accurate prediction and interpretable analysis of effluent ammonia‑nitrogen (Eff_NH 3 -N) are crucial for wastewater treatment plant optimization; however, food wastewater—characterized by high chemical oxygen demand (COD) (up to 4320 mg/L) and severe fluctuations—presents significant challenges. This study developed an interpretable machine learning (ML) framework for Eff_NH 3 -N prediction in a full-scale continuous-flow system consisting of two microaerobic reactors (MA1 and MA2) followed by a membrane bioreactor (MBR). Using 920 daily records and eight candidate parameters, eight regression algorithms were evaluated. Random Forest (RF) and LightGBM showed the highest independent test performance among the evaluated models. SHapley Additive exPlanations (SHAP) analysis identified influent NH 3 -N, COD, mixed liquor suspend solids (MLSS), and temperature (T) as the dominant contributors to model predictions. Exhaustive evaluation of 255 feature subsets demonstrated that reducing the number of input parameters could retain predictive capability. The re-optimized four-parameter configuration (COD, NH 3 -N, MLSS, and T) yielded test R 2 values of 0.583 and 0.631 for Random Forest and LightGBM, respectively, compared with 0.613 and 0.602 for the corresponding full eight-parameter models. The re-optimized three-parameter configuration (COD, NH 3 -N, and T) yielded test R 2 values of 0.495 and 0.510 for RF and LightGBM, respectively, under a reduced monitoring configuration. These results highlight that global feature importance alone cannot determine optimal parameter reduction, and systematic subset evaluation is necessary for identifying parsimonious configurations. This framework provides an interpretable approach for reducing model complexity and may serve as a reference for input selection in similar full-scale wastewater treatment scenarios.