Machine learning and genetic algorithm for effluent quality optimization in wastewater treatment

This paper integrates the genetic algorithm (GA) with a fine tree machine learning model to simultaneously tune controller and setpoint for the effluent quality optimization while keeping the constraint violation time within an acceptable level for wastewater treatment Benchmark Simulation Model no.1 (BSM1). The weather and load uncertainties complicate the proportional-integral (PI) controller tuning in the BSM1 simulator. Conventional tuning approaches based on step test cannot take all possible weather scenarios into account, struggle with pollutant constraints, and fail to capture complex dynamics of the wastewater treatment process. Our approach resolves these issues by first generating 30 rain and storm scenarios with varying influent flow rates and duration for BSM1 to expand this testbed. Then, a fine tree model is developed to quantify the impact of integral constant, gain, anti-windup constant, dissolved oxygen and nitrate setpoints, on the effluent quality index and constraint violation. The GA is applied to yield a pool of potential solutions based on the machine learning model to optimize the effluent quality index subject to violation time constraint. We select representative designs from the solution pool by balancing the exploitation and exploration, then evaluate them on the BSM1 for surrogate model updating. This iterative process is repeated until the computational budget is reached. We compare the proposed approach with conventional Bayesian optimization to show its superiority in identifying the high-quality solution for BSM1 control system and improve the effluent quality by 6–43 kg pollution unit per day. • PI controller parameters and setpoint for BSM1 are optimized simultaneously. • Thirty rain and storm scenarios are generated to evaluate system robustness. • Fine tree models effluent quality and constraint violation with R 2 more than 90 %. • Genetic algorithm achieves 6–43 kg less pollutants than Bayesian optimization.