Interpretation of machine learning models to identify key factors for sludge dewatering

ABSTRACT Graphical abstract showing the scheme of this study. Sludge dewatering is an essential process in sludge disposal to reduce the volume and improve the efficiency of final disposal. Machine learning models for sludge dewatering were developed in this study, and all models showed good prediction accuracy on the test dataset with a correlation coefficient of more than 0.9. The interpretation of developed sludge dewatering models was carried out to identify the important factors affecting water content in dewatered sludge (WCDS) and sludge dewatering speed (SPD) by different approaches, including the mean decrease in impurity (MDI), permutation feature importance (PFI), sensitivity analysis (SA), SHapley Additive exPlanations (SHAP), individual conditional expectation (ICE), and partial dependence (PD). The analyses of feature importance by MDI, PFI, SA, and SHAP showed that fiber-like content (FC) and volatile suspended solids (VSS) in raw sludge as well as viscosity in polymer solution (0.2%) were important factors affecting WCDS and SPD. All approaches demonstrated that FC in raw sludge was an important factor affecting both WCDS and SPD, and the higher FC contributed to the lower WCDS and the higher SPD. PD and ICE plots showed that WCDS decreased sharply, and SPD increased sharply when FC was more than 100 g·kg-TS−1.