Towards better process management in wastewater treatment plants: Process analytics based on SHAP values for...

Area of Science:

• Environmental Engineering
• Water Treatment Technologies
• Machine Learning Applications

Background:

• Understanding pollutant removal in Wastewater Treatment Plants (WWTPs) is vital for effluent quality.
• Complex interactions within WWTPs often obscure process comprehension.
• Previous machine learning (ML) efforts lacked granular interpretation.

Purpose of the Study:

• To upgrade an ML framework for uncovering cause-and-effect relationships in WWTPs.
• To compare the performance of three interpretable tree-based models (RF, XGBoost, LightGBM) using multiple metrics.
• To apply an advanced interpretation system (SHAP) for detailed process insights.

Main Methods:

• Implemented an upgraded ML framework with RF, XGBoost, and LightGBM models.
• Utilized R², RMSE, and MAE metrics for model evaluation.
• Employed SHapley Additive exPlanations (SHAP) for granular interpretation of operational factors and effluent parameters.
• Demonstrated the framework using a case study from Umeå WWTP, Sweden.

Main Results:

• XGBoost models demonstrated optimal performance for predicting Total Suspended Solids (TSS_e) and Phosphate (PO4_e) in effluent.
• Random Forest models showed suboptimal performance due to overfitting and polarized fitting.
• SHAP analysis provided granular insights into cause-and-effect relationships for individual instances.

Conclusions:

• The upgraded ML framework with XGBoost and SHAP offers superior interpretability and predictive power for WWTPs.
• Model comparison requires multiple perspectives and granular interpretation methods like SHAP for comprehensive understanding.
• Findings offer significant insights for controlling TSS_e and PO4_e in WWTPs and inform broader ML applications in process industries.