Unveiling Hydrogen Fluoride Emission Mechanisms in Municipal Solid Waste Incineration Using a Machine Learning Approach.

Hydrogen fluoride (HF) emissions from municipal solid waste incineration (MSWI) pose significant environmental and health risks. However, their complex formation mechanisms remain poorly understood. This study presents an integrated machine learning framework combining XGBoost for HF prediction, SHAP for feature interpretation, structural equation modeling (SEM) for mechanistic analysis, generalized additive models (GAMs) for threshold identification, and self-adaptive nondominated sorting genetic algorithm II (SA-NSGA-II) for multiparameter optimization. Using over 150,000 high-frequency (5 s interval) sensor records from a waste-to-energy plant in Hainan Province, China (June 1–10, 2024), the XGBoost model showed the best performance among the evaluated models (R² = 0.755, RMSE = 0.041 mg/m³, MAE = 0.031 mg/m³) via 5-fold cross-validation. SHAP analysis identified flue gas temperatures- especially the second flue right side (10.97%) and first flue top (10.19%)- as dominant factors. SEM confirmed the grate incineration zone as the primary HF source (path coefficient = 1.058, p < 0.001). GAM identified location-specific critical temperature thresholds for HF emission control, specifically 767 °C at the upper second flue gas pass, 875 °C at the first flue top, and 212 °C at the low-temperature economizer inlet. SA-NSGA-II optimization, validated with June 11 data, reduced HF emissions in 89.74% of cases, achieving a 17.61% average reduction (0.1176 mg/m³). This framework advances mechanistic understanding and provides data-driven strategies for sustainable MSWI operation and pollution mitigation.

The Supporting Information is available free of charge at es6c00686_si_001.pdf (2.11 MB)

• Tables of feature variables, descriptive statistics, Boruta algorithm-selected features, machine learning model hyperparameters, boiler operating parameter stability, and co-optimization control parameter ranges; figures of data collection number, integrated analytical framework flowchart, core boiler parameter I-MR control charts, XGBoost model validation, SHAP-based feature importance ranking, and top parameter impact on HF emissions; and textual descriptions of machine learning models and Bayesian optimization, delay time determination via mutual information, optimization performance evaluation metrics, and framework applicability to other MSWI plants (PDF)