Knowledge embedding and interpretable machine learning optimize comprehensive benefits for water treatment

Abstract

Perikinetic and orthokinetic flocculation are the first steps in drinking water treatment plant (DWTP) and affect all subsequent processes. Leveraging multi-stage water quality parameters, we developed a machine learning (ML) framework for coagulation control that incorporates knowledge embedding (KE) through hyper-parametric constraints on threshold water quality, energy consumption, and economic costs. Random forest (RF) has the best performance among the eight methods with a percentage error of 2.53% and a coefficient of determination of 0.9922. The results of the interpretability analysis show that the model can accurately identify the coagulation demand and balance the removal effect with the energy consumption and economic cost. Through real experimental validation and simulation extrapolation, the RF-KE model can reduce turbidity by 16.36% and dosing cost by 9.64%. This framework reduces economic costs while optimizing water quality through KE and interpretability analyses, providing evidence for the safe and reliable application of future models.