基于机器学习的短程硝化/短程反硝化-厌氧氨氧化工艺脱氮性能预测与关键参数识别  

PREDICTION OF NITROGEN REMOVAL PERFORMANCE AND IDENTIFICATION OF KEY PARAMETERS OF PARTIAL NITRIFICATION/PARTIAL DENITRIFICATION-ANAMMOX PROCESS BASED ON MACHINE LEARNING

作　　者：吴宇伦 李泽敏 成晓倩 邱光磊 韦朝海 WU Yulun;LI Zemin;CHENG Xiaoqian;QIU Guanglei;WEI Chaohai(School of Environment and Energy,South China University of Technology,Guangzhou 510006,China;School of Environment,South China Normal University,Guangzhou 510006,China)

机构地区：[1]华南理工大学环境与能源学院,广州510006 [2]华南师范大学环境学院,广州510006

出　　处：《环境工程》2024年第9期180-190,共11页Environmental Engineering

基　　金：国家自然科学基金项目“工业废水的形成、资源化利用与污染控制——以钢铁、冶金、矿山废水为例”(U1901218)。

摘　　要：短程硝化-厌氧氨氧化(PNA)与短程反硝化-厌氧氨氧化(PDA)工艺的脱氮性能会受到许多参数的影响。在综合考虑各种参数的基础上,对2种工艺的脱氮性能进行预测,并识别关键参数,能够为其实际工程应用提供优化目标。解决上述问题时,实验方法耗时耗力,而传统数学模型难以处理非线性关系。因此采用机器学习技术,构建的随机森林(RF)机器学习模型对2个工艺的出水总氮(TN)浓度进行了高精度预测,对PNA和PDA工艺出水TN浓度预测结果的决定系数(R^(2))分别为0.728、0.812。SHAP方法能够较好地解释模型的预测过程,并对各参数进行了重要性排序。在PNA工艺中,出水TN浓度主要受到进水TN浓度及COD浓度的影响。在PDA工艺中,出水TN浓度首先受进水TN浓度及氮负荷的约束。在此基础上,进水COD浓度作为另一重要因素影响着工艺的出水TN浓度。进水COD浓度在2个工艺中的共同重要性表明,2种工艺在实际应用时需要预先做好污废水中碳源的管理与分配,预分离与应用策略非常重要。该研究采用机器学习模型为PNA与PDA工艺脱氮性能的预测提供了方法指导,并基于SHAP的模型解释为2种工艺在实际应用时的关键参数识别与优化提供了选择依据。The nitrogen removal performance of the partial nitrification-Anammox(PNA)and partial denitrification-Anammox(PDA)processes are affected by many parameters.Predicting the performance of the two processes and identifying the key parameters based on a comprehensive consideration of various parameters can provide an optimization target for their practical engineering applications.When solving the above problems,experimental methods are usually time-consuming and labor-intensive,while traditional mathematical models are difficult to deal with non-linear relationships.Therefore,in this study,machine learning techniques were used.The constructed Random Forest(RF)machine learning model predicted the effluent nitrogen(TN)concentration of the two processes with high accuracy,and the coefficient of determination(R^(2))of the PNA and the PDA processes were 0.728 and 0.812,respectively.The SHAP method explained the prediction process of the model well and ranked the importance of each parameter.In the PNA process,the effluent TN concentration was mainly influenced by the influent TN concentration and COD concentration;in the PDA process,the effluent TN concentration was firstly constrained by the influent TN concentration and nitrogen load.On this basis,influent COD concentration is another important factor that affects the effluent TN concentration of the PDA process.The common importance of the influent COD concentration in both processes indicated that both processes should be managed and allocated to the carbon source in the wastewater well in advance of practical application.It is of significant importance to consider the pre-separation and application strategies.The machine learning model used in this study can provide methodological guidance for the prediction of the nitrogen removal performance of the PNA and PDA process.The SHAP-based model interpretation can provide a foundation for the identification and optimization of key parameters for the two processes in practical application.

关 键 词：短程硝化-厌氧氨氧化 短程反硝化-厌氧氨氧化 机器学习 SHAP分析 工艺参数 

分 类 号：X703[环境科学与工程—环境工程]