Generalized prediction and optimal operating parameters of PCDD/F emissions by explainable Bayesian support vector regression
The current derived models for predicting polychlorinated dibenzo-p-dioxins and -furans (PCDD/F) emissions from incineration can only be applied to a specific incinerator due to high deviation or systematic errors. And the models fail to provide quantized guidance for the operation of full-scale municipal solid waste incinerators. To address the problem, explainable Bayesian support vector regression (E-BSVR) has been established to generalized predict and maximumly reduce the PCDD/F emissions. First, forty-two PCDD/F samples were determined from a whole year experiment in a full-scale incinerator. Meanwhile, 1,2,4-trichlorobenzene(1,2,4-TrCBz), carbon monoxide, sulfur dioxide, oxynitride, particulate matter, fluoride, and hydrogen chloride were measured, as input features. Second, after box-cox transformation normalization, and hyperparameters tuning, the R-Squared and root mean square error (RMSE) of the proposed method are 0.983 and 0.044, exhibiting high accuracy. The high accuracy (R-Squared = 0.992) and generalization are also proven on the dataset with high PCDD/F emissions. Then, the performances of BSVR are compared with kernel ridge regression, multiple linear regression, and unary linear regression, indicating afar smaller RMSE of BSVR. Finally, the optimal operating parameters are calculated through local interpretable model-agnostic explanations and the partial dependence plot. Results indicate that reducing the content of organic chlorine in municipal solid waste and inhibiting the deacon reaction are important methods for reducing PCDD/F emissions. The optimal operating parameters for the maximal reduction of PCDD/F emissions are 1,2,4-TrCBz < 0.098 ug/m3, fluoride > 0.452 mg/m3. As a whole, the E-BSVR method can be used as a reliable and accurate approach for the prediction and reduction of PCDD/F emissions.