Automated multi-chemical dosing control based on machine learning for magnetic coagulation in aquaculture wastewater

Magnetic coagulation using magnetic powder with polyaluminum chloride and polyacrylamide is a promising option for aquaculture wastewater treatment, yet practical deployment is constrained by the need for real-time multi-chemical dosing under highly variable influent conditions. This study develops and validates an automated, control-oriented framework for optimizing combined magnetic coagulant and coagulant-aid dosages during dynamic operation. A fast data-driven surrogate based on an Extreme Learning Machine (ELM) was constructed and tuned with Grey Wolf Optimizer (GWO), and dosing decisions were formulated as a desirability-based, chance-constrained optimization problem to account for multi-objective trade-offs and reliability. The optimized setpoints were implemented through a feedforward-feedback controller, while Shapley Additive Explanations and partial dependence plots were used to interpret dose-response patterns and chemical interactions. Using lab-scale dynamic datasets, GWO tuning improved geometric-mean desirability by 0.12-0.29 and reduced prediction errors by 13-38%. Under abrupt disturbances, the GWO-ELM-PID strategy achieved 99.7% on-specification time and reduced total chemical consumption by 9.98% and 7.82% compared with open-loop and conventional proportional-integral-derivative strategies, respectively. These results support a reliable, interpretable, and deployable framework for real-time multi-chemical coagulation control.