Interpretable modeling for effluent organic load control and optimization in wastewater treatment

Area of Science:

• Environmental Engineering
• Data Science
• Wastewater Treatment

Background:

• Effluent organic load (ECOD) is a critical parameter for optimizing wastewater treatment plant (WWTP) operations.
• Effective ECOD prediction is vital for controlling organic pollution and enhancing treatment efficiency.

Purpose of the Study:

• To develop and validate a machine learning framework for accurate ECOD prediction in a full-scale WWTP.
• To interpret the factors influencing ECOD using explainable AI (XAI) techniques.

Main Methods:

• Implemented an Extreme Gradient Boosting (XGBoost) model for ECOD prediction.
• Utilized Shapley Additive Explanations (SHAP) for model interpretability and identification of key predictors.
• Analyzed partial dependence plots to understand variable interactions and nonlinear effects.

Main Results:

• The XGBoost model achieved high accuracy, with an R² of 0.919 and an RMSE of 0.791 mg/L.
• SHAP analysis identified aeration rate, influent flow rate, and suspended solids removal rate as significant predictors of ECOD.
• Nonlinear relationships and interactions among variables were revealed, providing insights into operational thresholds.

Conclusions:

• The proposed XGBoost-SHAP framework offers a scalable and practical solution for WWTPs.
• Actionable strategies for aeration control, process sustainability, and nutrient removal can be derived from the model's insights.
• The framework facilitates adaptive operational adjustments, leading to energy savings and emission reductions.