Prediction Method And Process Optimization Of Dioxin Precursor HCL Based On Integrated Model

Waste incineration power generation was the main method to realize the reduction,recycling and harmless treatment of municipal solid waste.However,unstable waste combustion conditions will produce more serious secondary pollution,the most serious of which is the production of highly carcinogenic dioxins.Dioxin was difficult to measure online,and discrete samples obtained by analytical instruments had the disadvantages of long time,high cost,and small number of samples.The soft measurement method was also restricted by various objective conditions,and the research progress was very small.In this paper,by analyzing the formation mechanism of dioxin,it was found that the formation of dioxin was inseparable from the CL element,and it was mainly converted from HCL in the main combustion zone.Therefore,this paper analyzed and optimized the operating parameters of process variables by constructing an online prediction model of HCL concentration,as a basis for evaluating the change of dioxin concentration,and provided a new research direction for the emission reduction of dioxin flue gas.1.Statistical analysis of operating condition data sets and screening of high-dimensional variables.First,the process data of the waste incineration plant was sorted out using the HOLLi AS MACS software.Then,89 operating variables and 5 flue gas variables were statistically analyzed,and it was found that the data presented discrete distribution,high coupling and nonlinear characteristics.Next,outlier removal was performed.Under the SW method test,it was found that the data in this paper did not conform to the normal distribution.Therefore,44 process variables were screened out by Spearman’s correlation coefficient method.After that,the XGBoost method was used to sort the importance of variables twice,and 38variables were screened out.To prevent the model-biased learning caused by the high correlation between data,an independence test was carried out.On the premise that the p-value between the variables was greater than 0.05 and the correlation was less than 0.8,27 processes were finally determined.Variables are used for model building.2.Construction of flue gas HCL prediction model based on GA-multilayer BPNN.Based on the discrete distribution,high coupling and nonlinearity of the data,this paper used the GA-multilayer BPNN model with strong fitting ability to establish the HCL prediction model.To determine the optimal structural parameters of BPNN,we optimized the parameters of the number of hidden layers and the number of hidden layer nodes in the network respectively,and evaluated the model’s optimality with four evaluation indicators:MAE,MSE,RMSE,and R².The structure of the multi-layer BPNN model was 27-12-6-2-1,and the regression accuracy was 83.98%.Finally,using the genetic algorithm GA to optimize the parameters of the model,the accuracy of the optimized HCL prediction model was 84.28%.3.Improvement of hybrid flue gas HCL prediction model based on deep forest and ensemble learning.Since the GA-multilayer BPNN model had the shortcomings of complex structure determination,many parameters,poor stability,and the accuracy needs to be improved,this paper used deep forest to automatically build the best model.The regression accuracy on the test set was 90.2%,the maximum error of HCL concentration was 0.59 mg/m³,and the average regression error was 0.28 mg/m³.Compared with the GA-multilayer BPNN model,the deep forest had more advantages in determining the best parameters,accuracy and stability.To further improve the accuracy,effectively use the STACKING ensemble algorithm to combine the advantages of multiple algorithms,and avoid the defects brought by the inferior model,this paper proposed the stacking integrated algorithm based on dynamic parameters（MK-STACKING）.Compared with the STACKING model,the accuracy was improved from 90.08%to 92%.Finally,by comparing the prediction effects of GA-multi BPNN,deep forest and MK-STACKING,it was found that MK-STACKING had the best prediction effect for HCl concentration.4.Design of human-computer interaction simulation system for HCL concentration prediction.The system was to facilitate the central control system to obtain the prediction result of HCL concentration in real time.The system mainly included the login interface,the main interface of HCL concentration prediction and the evaluation and analysis interface of HCL concentration prediction effect,including such as user login,operation instructions,and result display,to making users quickly understand and use.5.Optimization of process parameters.In order to provide a treatment method for HCL concentration control to indirectly reduce dioxin emissions,this paper used genetic algorithm to optimize 26 operating variables except SO₂ concentration,and determined the optimal range of operating parameters.Taking the minimum value of HCL concentration predicted by MK-STACKING model as the objective function,the optimal values of 26 process variables were obtained.And based on the optimal adjustment value,the process parameters were adjusted within a certain range up and down,which verified the rationality of the optimal process variable selection value.