Multi-target machine learning prediction of constructed wetland effluent quality with model interpretation and experimental validation

In the design and operation of constructed wetlands (CWs), carbon and nitrogen pollutant levels, such as chemical oxygen demand (COD), ammonia nitrogen (NH₄⁺-N), and nitrate nitrogen (NO₃^--N), are key indicators for assessing effluent quality and controlling eutrophication of receiving water bodies. Accurately predicting the effluent concentrations and identifying the key regulatory factors are of great significance for enabling process optimization and precise system control. However, traditional CWs prediction methods often fail to effectively capture the complex nonlinear relationships in these pollutant removal processes, resulting in poor prediction results and a lack of interpretability analysis of feature contributions. Therefore, this study constructed an integrated machine learning and Shapley Additive exPlanations (SHAP) framework to simultaneously predict the effluent concentrations of COD, NH₄⁺-N, and NO₃^--N, as well as quantitatively analyze the contribution of each input feature. Additionally, independent experiments were conducted to further evaluate the applicability and stability of the developed model. By comparing Decision Tree (DT), Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Extreme Gradient Boosting (XGBoost), it was found that GBDT performed best in multi-target predictions, with testing set R² values above 0.87. SHAP analysis further indicated that influent COD, NH₄⁺-N, and aeration rate are the dominant factors affecting effluent quality, while wetland structure parameters have relatively limited influence. Moreover, independent experiments further demonstrated the applicability and stability of the developed model under independently obtained experimental conditions, with R² values above 0.75. The findings of this study can be used to identify factors affecting the treatment performance of CWs and to develop design guidelines for improving treatment processes.