Emission Characteristics And Correlation Model Of Dioxin And Chlorobenzene In The Stack Gas Of Hazardous Waste Incineration

The increasing attention of the government and the public to the emission of dioxin and China’s increasingly stringent environmental protection requirements make the online monitoring of dioxin particularly critical.In the process of hazardous waste incineration,the emission of dioxin is relatively high and the monitoring and control methods are limited.There are fewer targeted studies.To gain an in-depth understanding of the emission characteristic of dioxin during hazardous waste incineration and realize online monitoring of dioxin,this work focused on the emission characteristics and correlations of polychlorinated dibenzo-p-dioxin and dibenzofuran（PCDD/F）and chlorobenzene（CBz）of a hazardous waste incinerator（HWI）in China,as well as the prediction model of dioxin emission with Bagging-based ensemble learning.The main content and results of the dioxin emission prediction model include:（1）We selected a HWI in China as the research subject to study the emission characteristics of PCDD/F and CBz,and the influence of inhibitors and the chlorine content of raw material on the emission of PCDD/F and CBz.The emission concentrations of low-chlorinated PCDD/F were significantly higher than those of high-chlorinated PCDD/F.23478-Pentachlorinated dibenzofuran contributed the most to the International Toxic Equivalent Quantity（I-TEQ）value.13-Dichlorobenzene contributed the most to the total concentration of CBz.Inhibitors could effectively reduce the emission of PCDD/F and CBz,and the inhibitory effect of thiourea+ammonium dihydrogen phosphate was significantly better than that of Ca（OH）₂+ammonium dihydrogen phosphate of the same concentration.The increase in the chlorine content of raw material led to a significant increase in the emission of PCDD/F and CBz.The emission concentration of PCDD increased more significantly than that of PCDF.（2）The correlations between CBz,conventional pollutants（SO₂,NO_x,dust,CO,HCl）,key incinerator parameters,and PCDD/F were analyzed.There were good correlations between some CBz congeners and PCDD/F.Hexachlorobenzene showed the highest linear correlation with the I-TEQ value.If the nonlinear correlation was also considered,the correlation between 135-Trichlorobenzene and the I-TEQ value was the highest.The correlations between low-chlorinated CBz congeners and low-chlorinated PCDD/F were not as good as those of high-chlorinated CBz congeners and high-chlorinated PCDD/F.There were no obvious correlations between conventional pollutants and PCDD/F.The water spray rate of the quenching tower had a negative correlation with PCDD/F,which meant the quenching tower had an inhibitory effect on the formation of PCDD/F.Among the linear regression models,the binary linear regression model with the water spray rate of quenching tower and pentachlorobenzene as independent variables had the best fitting effect（R²=0.801）.Adding independent variables caused collinearity,which had bad effects on the generalization ability of the model.（3）We applied Artificial Neural Network（ANN）and Classification And Regression Tree（CART）to build PCDD/F emission prediction models.Then ANN and DT were stacked into Bagging-based ensemble Neural Networks（BGNN）and Random Forest（RF）to improve the accuracy and robustness.Finally,the Shapley additive explanation（SHAP）method was used to explain the effect of input variables on the model output,laying a foundation for feedback control.BGNN is the most suitable model for dioxin emission prediction.Compared with RF,BGNN has a stronger generalization ability and the disadvantage of too many parameters can be ignored on small-scale data sets.ANN and CART can cause relatively serious over-fitting problems,and the generalization ability of Bagging-based ensemble learning models is better than that of the corresponding base learner.In the municipal solid waste incinerators（MSWI）data set,machine learning nonlinear models are better than linear regression models.In the smaller-scale HWI data set,the generalization ability of the multiple linear regression model（ML）is second only to BGNN.