A hybrid machine learning–based multi-objective supervisory control strategy for cost-effective and sustainable wastewater treatment under varying influent conditions

Abstract This study develops a multi-objective supervisory control (MOSC) strategy for wastewater treatment based on hybrid machine-learning algorithms that search optimal setpoints of multiple controllers under varying influent conditions. A wastewater treatment plant (WWTP) operation was modeled by Benchmark Simulation Model No. 2 (BSM2), and influent conditions were generated in consideration of a H-WWTP in South Korea. Two proportional-integral (PI) controllers for dissolved oxygen and biogas, and one cascade-PI controller for nitrate were used as local control loops. The MOSC strategy identified five influent scenarios using fuzzy c-means algorithms and nitrogen-to-carbon ratios. Then, the control performance according to influent changes was gauged employing a deep-learning-based approximation model, and optimal setpoints for the controllers were determined by a non-dominated sorting genetic algorithm. The results demonstrate that an intelligent MOSC strategy can identify optimal setpoints to improve WWTP performance and outperform a reference control across a range of possible ratios of total Kjeldahl nitrogen to chemical oxygen demand (TKN/COD) in influent disturbances. The MOSC strategy was also able to accommodate extreme influent conditions, reduce operational costs by 8%, maintain effluent quality, and produce biogas for sustainable WWTP operation.