| 引用本文: | 梁诗琪,李坦,王越,余华荣,瞿芳术.基于机理模型与机器学习模型的污水厂运行策略优化[J].哈尔滨工业大学学报,2026,58(6):130.DOI:10.11918/202508009 |
| LIANG Shiqi,LI Tan,WANG Yue,YU Huarong,QU Fangshu.Operational strategy optimization for wastewater plants based on mechanistic and machine learning models[J].Journal of Harbin Institute of Technology,2026,58(6):130.DOI:10.11918/202508009 |
|
| |
|
|
| 本文已被：浏览 1150次 下载 78次 |
码上扫一扫！ |
|
|
| 基于机理模型与机器学习模型的污水厂运行策略优化 |
|
梁诗琪,李坦,王越,余华荣,瞿芳术
|
|
(1．广州大学 土木与交通工程学院,广州 510006; 2．珠江三角洲水质安全与保护教育部重点实验室(广州大学),广州 510006)
|
|
| 摘要: |
| 为提升污水处理厂运行效能,以污水处理厂出水总氮为主要控制指标,比较机理模型与机器学习模型的预测能力,在模型构建的基础上考察参数优化对运行效能的提升（出水总氮、能耗）。采用响应面法耦合活性污泥模型,构建响应面二次多项式优化好氧区溶解氧、污泥内回流比及污泥外回流比3个关键工艺参数,以获得最优的出水总氮。同时,考察并优选随机森林模型、极端梯度提升模型、轻量级梯度提升机模型等6种经典机器学习模型,进一步构建双延迟深度确定性策略梯度算法耦合随机森林模型的多目标优化架构,协同优化出水总氮与能耗。结果表明:响应面法耦合活性污泥模型拟合度高（R2=0.944 5）,在最优条件下出水总氮去除率可达89.14%,最佳出水总氮质量浓度为3.533 mg/L,比实际水厂出水降低45.48%；机器学习模型预测中随机森林表现最佳（R2=0.79,平均相对误差EMR=7.5%）,采用双延迟深度确定性策略梯度算法强化学习多目标优化控制策略后,平均出水总氮质量浓度为6.20 mg/L,比实际水厂出水降低4.32%,曝气与泵送能耗降低33.12%,运行成本指数降低25.16%；在短期内单一的总氮去除目标上,活性污泥模型的处理效果更好,但双延迟深度确定性策略梯度算法优化的优势在于多目标优化,在保证总氮稳定达标的基础上,在控制能耗、成本和稳定性方面取得了显著提升,其在长期操作中综合处理效果更为优越。综上,双延迟深度确定性策略梯度算法耦合随机森林模型实现多目标优化控制,为污水处理厂智能运行提供理论与方法支撑。 |
| 关键词: 污水处理 活性污泥模型 机器学习 双延迟深度确定性策略梯度算法 多目标优化控制 |
| DOI：10.11918/202508009 |
| 分类号:TU992.3 |
| 文献标识码:A |
| 基金项目:国家重点研发计划(2023YFC3207903) |
|
| Operational strategy optimization for wastewater plants based on mechanistic and machine learning models |
|
LIANG Shiqi,LI Tan,WANG Yue,YU Huarong,QU Fangshu
|
|
(1.School of Civil Engineering and Transportation, Guangzhou University, Guangzhou 510006, China; 2.Key Laboratory of Water Quality Safety and Protection of the Pearl River Delta (Guangzhou University), Ministry of Education, Guangzhou 510006, China)
|
| Abstract: |
| To enhance the operational efficiency of wastewater treatment plants, this paper took effluent total nitrogen as the primary control indicator, comparing the predictive capabilities of mechanistic and machine learning models. Based on model construction, the improvement effects of parameter optimization on operational efficiency (effluent total nitrogen and energy consumption) were further investigated. First, the response surface methodology was coupled with the activated sludge model to establish a quadratic polynomial response surface for optimizing three key process parameters: dissolved oxygen in the aerobic zone, internal sludge recirculation ratio, and external sludge recirculation ratio. This optimization aimed to achieve the optimal effluent total nitrogen. In parallel, six classical machine learning models such as random forest model, extreme gradient boosting model, and light gradient boosting machine model were evaluated and selected. Furthermore, a multi-objective optimization framework was developed by coupling the Twin Delayed Deep Deterministic Policy Gradient Algorithm with a random forest model, synergistically optimizing effluent total nitrogen and energy consumption. The results demonstrate that the response surface methodology coupled with the activated sludge model achieves a high goodness of fit (R2=0.944 5). Under optimal conditions, the removal efficiency of effluent total nitrogen can reach 89.14%, with the optimal effluent total nitrogen concentration of 3.533 mg/L, representing a 45.48% reduction compared with that of actual wastewater treatment plants. Among the machine learning prediction models, the random forest model exhibits the best predictive performance (R2=0.79,EMR=7.5%). After applying the multi-objective optimization control strategy processed by a Twin Delayed Deep Deterministic Policy Gradient Algorithm based on reinforcement learning, the average effluent total nitrogen concentration is 6.20 mg/L, 4.32% lower than that of actual wastewater treatment plants, while aeration and pumping energy consumption decrease by 33.12%, and the operating cost index is reduced by 25.16%. In the short term, the activated sludge model achieves superior performance for the single objective of total nitrogen removal. However, the advantage of the Twin Delayed Deep Deterministic Policy Gradient Algorithm optimization lies in its multi-objective optimization, delivering significant improvements in controlling energy consumption, cost, and stability while maintaining effluent total nitrogen compliance. In the long run, this approach provides superior overall treatment results. In conclusion, the Twin Delayed Deep Deterministic Policy Gradient Algorithm coupled with random forest model enables effective multi-objective optimization control and offers both theoretical and methodological support for the intelligent operation of wastewater treatment plants. |
| Key words: wastewater treatment Activated Sludge Model machine learning Twin Delayed Deep Deterministic Policy Gradient Algorithm multi-objective optimization control |
|
|
|
|
