Predicting cyanobacteria removal efficiency in flocculation-DAF: Improving interpretable automated machine learning...

Flocculation-dissolved air flotation (DAF) is an efficient and widely adopted technique for cyanobacteria separation. However, optimizing its removal efficiency remains challenging due to complex interdependencies among water quality, cyanobacterial characteristics, and operational parameters. To enable accurate prediction and identify key control parameters, we developed an interpretable machine learning framework integrating Conditional Variational Auto-Encoder (CVAE) with H₂O AutoML to accurately predict cyanobacteria removal efficiency and identify critical control parameters. Our approach leverages CVAE-based augmentation to expand the experimental dataset with high-quality synthetic samples, effectively mitigating data scarcity issues. The CVAE-AutoML model achieved significantly superior prediction accuracy (R² = 0.98), outperforming traditional models like Random Forest and even AutoML models trained solely on experimental data (an R² improvement of 4.42 %). Model interpretability analyses revealed flotation time, flocculant dosage, and cyanobacteria density as the most influential variables. Specifically, flotation times exceeding 60 s were detrimental to efficient cyanobacteria removal (response value reached 76.29 %); a flocculant dosage around 25 mg/L resulted in an immediate 33.5 % increase in the response value; and when cyanobacteria density surpassed 5.34 × 10⁶ cells/mL, the response value initially improved by 8.87 % and continued to rise thereafter. This methodology not only elucidates the underlying dependencies between the cyanobacteria separation process and its key influencing factors but also offers valuable guidance for optimizing DAF system operational strategies.