Optimal set point generation based on deep reinforcement learning and transfer knowledge for Wastewater Treatment Plants

In this work, a model-free Reinforcement Learning method is implemented to control the aeration in the nitrification process in wastewater treatment plants (WWTPs). Controlling real complex non-linear processes is challenging due to model mismatch between the model used in the controller and the real process. Model-based control strategies can fail in a real implementation due to modeling errors. WWTPs include multivariable, nonlinear, and complex biological processes, which makes it difficult to obtain adequate models to represent the dynamics of the process under all possible conditions. For this reason, a model-free reinforcement learning agent that takes advantage of Transfer Learning is selected. This data-driven procedure is trained to deal with the trade-off between effluent quality and operating costs. The results demonstrate that model-free RL can reach sub-optimal solutions through a data-driven procedure.