Predicting single-cell protein production from food-processing wastewater in sequencing batch reactors using ensemble learning

Producing single-cell protein (SCP) from food-processing wastewater offers a sustainable approach to resource recovery, animal feed production, and wastewater treatment. Decision-makers need accurate system performance data under variable influent conditions to select operational parameters for efficiency. However, predicting system performance under variable conditions is challenging due to the complexity of unsteady-state bioreactions. This study trained and tested ensemble learning algorithms, including the ensemble of Support Vector Regression, the ensemble of Gaussian Process Regression (GPR), Random Forest, and Extreme Gradient Boosting, to predict outcomes in a continuous-inflow, sequencing-batch-reactor-based SCP system using industrial soybean-processing wastewater. Interpretable analysis and trials validate feature significance for model optimization. Results show that ensemble-learning models, particularly GPR-based ones, outperform linear regression in predicting key effluent and biomass variables essential for operational decision-making. Notably, GPR-based ensembles with influential features predict biomass production (coefficient of determination (R²) = 0.72) against overfitting much better than linear regression (R² = 0.4).