Calibration for Wastewater Treatment Modeling: Hybrid Approaches Combining Fuzzy Logic, Reinforcement Learning, and Particle Swarm Optimization

In this paper, novel calibration approaches are proposed for modeling wastewater treatment processes. The benchmark simulation model has been widely adopted to represent the integrated activated sludge and sedimentation processes. However, this model remains challenging due to its inherent mathematical nonlinearities and the complex interdependencies among its parameters. To address these challenges, fuzzy logic and reinforcement learning techniques are combined in the hybrid algorithms, searching for optimal parameter configurations within defined domains. Two kinds of hybrid algorithms referring to reinforcement learning are proposed here, which are, respectively, based on Q-map and deep Q networks. Their convergence speed, accuracy, and robustness in parameter calibration are verified and compared. Furthermore, a modified particle swarm optimization method is incorporated into both hybrid algorithms to enhance their performance. Modeling real-world systems presents several challenges, including incomplete parameter information, the low dimensionality of the observed data, and limited adaptability of existing models. The resulting combined approaches demonstrate effectiveness in addressing these issues. The application of the present model will facilitate automated operations, process optimization, and energy optimization in the future.