Predicting effluent total nitrogen removal in full-scale WWTPs using a novel optimized machine learning model: SVR enhanced by SHAP and LIME explainability

In sustainable cities, forecasting effluent properties within wastewater treatment plants (WWTPs) offers significant benefits by decreasing sampling frequency, energy consumption, and operational expenses. Overcoming these problems necessitates establishing a suitable advanced nitrogen removal framework for municipal WWTPs. Therefore, this research primarily aims to create precise prediction models for nitrogen elimination in municipal WWTPs through sophisticated machine learning methodologies. The investigation proposes ten powerful metaheuristic optimization algorithms to develop enhanced support vector regression (SVR) frameworks. The research employed a structured workflow where models first underwent development using an extensive database, then faced rigorous testing to determine their accuracy and prediction quality. To ensure dependable total nitrogen forecasting, each approach was evaluated under eight different combinations of input parameters. SHAP and LIME interpretability frameworks were also employed in this investigation, with their primary function being to reveal which variables drive nitrogen concentration forecasts. Through these analytical tools, the study gained insights into feature hierarchies and discovered the fundamental relationships connecting input data to model predictions. Moreover, to assess model reliability, evaluation was conducted through four key statistical metrics, which facilitated comprehensive assessment of predictive performance for every model. According to RMSE analysis, the three highest-performing models were as follows: SVR-HFA exhibited the best results for total nitrogen effluent prediction (TNeff) with an RMSE of 1.285 mg/L and correlation coefficient of 0.905, making it the leading approach. SVR-GWO achieved second place, delivering an RMSE of 1.293 mg/L and R-value of 0.902, followed by SVR-TSA in third position with an RMSE of 1.316 mg/L and R-value of 0.898. These results offer practical guidance for WWTPs, highlighting reliable models for forecasting TNeff. By adopting these predictive tools, facilities can optimize operational performance, streamline monitoring tasks, and reduce reliance on laboratory testing. This investigation created ten integrated metaheuristic SVR frameworks for predicting TNeff within WWTPs. TNeff estimation employed eight distinct input variable combinations. SHAP and LIME were deployed to identify the most impactful and significant input variables. The SVR-HFA approach achieved the highest prediction accuracy for TNeff among the evaluated methods, with an RMSE of 1.285 mg/L. SHAP analysis demonstrated that TN7, COD6, TNinf, and COD6 were the most influential parameters for modeling TNeff.