Prediction Of Dioxin Emissions In Flue Gas From Municipal Solid Waste Incineration Based On Support Vector Regression

The scale of municipal solid waste incineration in China is increasing rapidly,the regulation of dioxin emissions in flue gas is an urgent problem to be solved.It is difficult to monitor dioxin emissions in real time because of the high cost.However,multiple regressions can be applied to predicting dioxin emissions in flue gas,with the aid of pollutant concentrations and operating parameters,which are collected by on-line monitoring systems.The critical problem is that there are very few samples of dioxin available for training prediction models due to the high cost.This leads to a poor performance of generalization when using linear regression models.A nonlinear method named support vector regression is presented in this paper,in order to improve the generalization performance of prediction.Sampling 10 times of dioxin from flue gas in a waste incineration plant,and simultaneously recorded online monitoring data of the operating conditions and conventional pollutants.A support vector regression(SVR)model for the dioxin concentration and TEQ was respectively established using the 10 samples.The average absolute percentage error(MAPE)was used as the index,Leave-one-out cross validation was used to evaluating the model and came to the following conclusions:(1)This paper compared 3 parameter optimization algorithms including the grid search method,genetic algorithm and particle swarm optimazation for parameter searching ability,it turns out that although the speed of grid search method is slower,but its parameter searching ability is best.Using the grid search method,polynomial kernel function,RBF kernel function and Sigmoid kernel function were used to establish the model respectively,the prediction error of the model with polynomial kernel function is minimum.(2)Multiple linear regression(MLR)and SVR model were established for predicting dioxin concentration and toxicity equivalent,using Variables selected by stepwise regression analysis.MAPE for dioxin concentration and TEQ of MLR was 20.94% and 31.20% respectively,while SVR was 17.41% and 22.77% respectively.Although the latter is smaller than the former,its error was large still.(3)A wrapper feature selection method combined with exhaustion method was used to reselect characteristic variables and established the SVR model.This time,the MAPE for dioxin concentration and TEQ was 10.78% and 6.56% respectively,much lower than the result of SVR based on stepwise regression analysis,which can achieve a satisfactory prediction precision.