Optimizing membrane bioreactor performance in wastewater treatment using machine learning and meta-heuristic techniques

Sustainable water management increasingly relies on reclaimed wastewater, and membrane bioreactor (MBR) technology offers an advanced treatment approach that yields superior effluent quality compared to conventional systems. However, predictive models for full-scale MBR performance-especially for contaminant and nutrient removal-remain poorly developed. Addressing this gap, this study applied machine learning to predict chemical oxygen demand (COD) removal efficiency in a full-scale MBR wastewater treatment plant. The plant achieved consistently high removal efficiencies for COD, ammonia nitrogen, total suspended solids, and fats, oils, and grease, whereas phosphorus removal was comparatively low. Initially, 23 input variables capturing influent characteristics, aeration conditions, and operational settings were considered. A correlation analysis distilled these into seven key parameters for model training. Among multiple algorithms tested, random forest regression provided the most accurate predictions. This model's performance was further improved via hyperparameter tuning with three meta-heuristic optimization techniques: particle swarm optimization, a genetic algorithm, and simulated annealing. The genetic algorithm yielded the greatest performance enhancement, boosting the model's coefficient of determination by 16 %, increasing the Nash-Sutcliffe efficiency by 8.3 %, and reducing the root mean square error by 22 %. These results demonstrate a novel integration of machine learning and meta- heuristic optimization for wastewater treatment modeling that significantly improves predictive accuracy. This approach underscores the potential for data-driven optimization of wastewater treatment operations, contributing to more efficient and sustainable water resource management.