Author(s): Weihao Chen; Wenchong Tian; Yuting Liu; Kunlun Xin

Linked Author(s):

Keywords: AAO process; Deep learning; Interpretability; Model predictive control; Wastewater treatment plants

Abstract: Anaerobic-anoxic-oxic (AAO) is acknowledged as a fundamental biological treatment process in urban wastewater treatment plants (WWTPs). Presently, data-driven deep learning models are increasingly utilized for constructing AAO process simulation and optimization control methods; nevertheless, the challenge of poor interpretability persists. The stability and security of the data-driven AAO optimization control system remain elusive, resulting in a lack of reliability in practical applications. In this study, we designed a deep learning network structure called TransLSTM-Net and developed a model predictive control method (TransLSTM-Net-MPC) with rationality verification. By optimizing the aeration volume, internal recycle, and sludge recycle processes, safe and effective real-time control of AAO is achieved. Additionally, the SHAP method was utilized to analyze the interpretability of the deep learning model, thereby further enhancing the credibility of TransLSTM-Net-MPC. These methods underwent validation using data from a real-world WWTP in eastern China. The results demonstrate that the TransLSTM-Net-MPC model accurately predicts the water quality variables of the AAO system. Furthermore, the explanations provided by the model have verified its credibility to a certain extent, thereby enhancing the confidence of operators and stakeholders. The integration of TransLSTM-Net-MPC with rationality verification not only reduces the aeration volume by nearly 8% compared to traditional controllers but also establishes a stable control trajectory, thereby proving to be a robust real-time control method for WWTPs.

DOI: https://doi.org/10.64697/HIC2024_P216

Year: 2024