HCl emission prediction in municipal solid waste incineration using a Transformer–XGBoost hybrid model with deep feature extraction and SHAP-based interpretation

Accurate prediction of hydrogen chloride (HCl) emissions plays an important role in efficient control of deacidification systems in municipal solid waste incineration (MSWI) plants. However, accurate prediction remains challenging due to the dynamic nonlinearity, multivariable coupling and time series characteristics of MSWI process. This study proposes a Transformer–XGBoost hybrid model to predict HCl emissions in MSWI process. Firstly, a feature extractor based on Transformer model is designed to capture target-relevant temporal dependencies and complex interactions among process variables, extracting features and generating auxiliary prediction. Subsequently, a weighted feature fusion strategy is introduced to integrate the extracted features with auxiliary prediction result, thereby enhancing the representational capacity of features. Finally, the fused features are input into the XGBoost predictor to predict the HCl emissions at time t + 3. The proposed model is trained, validated, and tested using real industrial data from the MSWI plant. Additionally, the SHapley Additive exPlanations (SHAP) analysis reveals model behavior and the importance of combustion air distribution and deacidification operating variables to HCl predictions. Comparative experimental results demonstrate that the proposed model achieves optimal prediction performance, with an R 2 of 0.974, a root mean squared error (RMSE) of 0.388 mg/m 3 , a mean absolute error (MAE) of 0.262 mg/m 3 , and a symmetric mean absolute percentage error (SMAPE) of 8.14 % on the testing set. These results indicate that the proposed method achieves accurate and robust HCl emission prediction in MSWI process, thereby providing critical data support for the precise, stable and cost-effective control of HCl emissions in MSWI plants.