Data-driven water quality prediction for wastewater treatment plants

Area of Science:

• Environmental Engineering
• Artificial Intelligence
• Water Quality Management

Background:

• Effective monitoring and management of wastewater treatment plants (WWTPs) are vital for environmental protection.
• Accurate prediction of treated water quality is essential for optimizing energy efficiency in WWTP operations.

Purpose of the Study:

• To compare the performance of four machine learning models (MLP, GFFR, MLP-PCA, RBF) for predicting water quality parameters in WWTPs.
• To evaluate two distinct modeling scenarios: a straightforward approach using WWTP inputs/outputs and a multi-step approach incorporating intermediate settler outputs.
• To assess the models' capability in handling high-dimensional data generated by multi-step modeling.

Main Methods:

• Implementation of Multilayer Perceptron (MLP), Generalized Feedforward Regression (GFFR), MLP with Principal Component Analysis (MLP-PCA), and Radial Basis Function (RBF) models.
• Application of two modeling scenarios: direct input-output prediction and multi-step prediction utilizing intermediate process data from primary and secondary settlers.
• Evaluation of model performance based on correlation accuracy (R) and prediction deviations (Normalized Root Mean Square Error - NRMSE, Normalized Mean Absolute Error - NMAE).

Main Results:

• The GFFR model demonstrated superior performance across both scenarios, particularly in the multi-step scenario for predicting conductivity (R=0.893, NRMSE=0.091, NMAE=0.071).
• All models struggled to accurately predict other water quality parameters, exhibiting lower correlations and higher deviations.
• The multi-step modeling technique significantly enhanced prediction accuracy for all models, with improvements ranging from 0.2% to 157% (average 60%).

Conclusions:

• Machine learning models, especially GFFR, show significant potential for accurate water quality parameter prediction in WWTPs.
• The multi-step modeling approach, leveraging intermediate process data, substantially boosts the predictive capabilities of AI models.
• Further research is needed to improve the prediction accuracy for a wider range of water quality parameters beyond conductivity.