Abstract:

Papermaking wastewater treatment process is susceptible to uncertain factors such as production process conditions switching and raw material heterogeneity. In the context of the coordinated development of pollution reduction and carbon reduction in the industry, how to ensure the discharge of sewage treatment in the water quality standard, and achieve synchronous reduction of treatment costs, energy consumption, and greenhouse gas emissions is an important issue restricting the development of the industry. In this paper, a multi-objective wastewater optimization method based on Kriging method and high dimensional model representation (HDMR) is proposed for the dynamic uncertainty of papermaking wastewater treatment. In this study, benchmark simulation model No. 1 (BSM1) was used to simulate the biochemical and precipitation processes of papermaking wastewater treatment process. Based on biochemical metabolism mechanism and data fusion, a Kriging-HDMR proxy model for real-time solving of greenhouse gas emissions in wastewater treatment process was established. By integrating the agent model into reinforcement learning, a multi-agent system based on“solving-decision-observation” for dynamic optimization of the wastewater treatment process was established, and a coordinated multi-objective optimization model for pollution reduction and carbon reduction was obtained. The study scenario results show that compared with the open-loop system, the dynamic optimization system can reduce operating cost by 4.10%, energy consumption by 22.10%, and greenhouse gas emissions by 10.30%, and can obtain and maintain an effective multi-objective dynamic optimization control strategy.

Key words: wastewater, optimization, multi-objective, greenhouse gases, operational cost, energy consumption, deep reinforcement learning

摘要：

造纸污水处理过程易受生产工艺条件切换、原材料异质性等不确定因素影响，在行业号召减污降碳协同发展背景下，如何保障污水处理水质达标排放，并实现同步降低处理成本、能源消耗以及温室气体排放，是制约行业发展的重要问题。本文面向造纸污水处理动态不确定性提出了一种基于Kriging法以及高维模型表征（HDMR）的多智能体强化学习的污水多目标优化方法，采用基准仿真1号模型（BSM1）模拟造纸污水处理过程的生化和沉淀过程，基于生化代谢机理和数据融合建立面向污水处理过程温室气体排放量实时求解的Kriging-HDMR代理模型，通过将代理模型集成至强化学习建立基于“求解-决策-观察”动态优化污水处理过程的多智能体系统，得到了减污降碳协同多目标优化模型。研究场景结果表明，相较于开环系统，该动态优化系统可降低运行成本4.10%，减少能源消耗22.10%，减少温室气体排放10.30%，能够得到并维持有效的多目标动态优化控制策略。

关键词: 污水, 优化, 多目标, 温室气体, 运行成本, 能源消耗, 深度强化学习