融合历史过程与未来工况的污泥热解气化废气排放动态预测

The pyrolysis–gasification process has emerged as a cutting-edge technology for sludge treatment and disposal because of its resource-recovery potential and high efficiency. However, the emissions of harmful gases such as SO₂ during operation limit the widespread adoption of this technology. Achieving accurate emission prediction and optimizing process parameters to improve both economic and environmental performance are therefore crucial. In this study, we used a high-resolution industrial dataset of 106 variables and 64,801 minute-level records collected continuously over a 45-day operational period at a full-scale plant. We developed a comprehensive time-series prediction framework that integrates historical process records with future operating conditions. The predictive performance of representative algorithms—including XGBoost, CatBoost, NLinear, and the Temporal Fusion Transformer (TFT)—was systematically evaluated and validated. Experimental results show that the proposed multi-source time-series prediction framework, which accounts for process dynamics and lag effects, is essential for modeling complex industrial gasification processes. Among the tested models, CatBoost performed best, achieving a mean absolute error (MAE) of 269.17 and a coefficient of determination (R²) of 76.53%. To assess the reliability of these results for production guidance, we compared the framework with a traditional non-temporal cross-sectional baseline model. The baseline attained an R² of 22.51% and an MAE of 542.20. Thus, the proposed framework improved the R² by 54.02 percentage points and reduced the MAE by 50.36%, indicating that traditional models fail to capture critical temporal correlations and the delayed response of pollutant generation to control inputs. In contrast, the proposed framework effectively leverages historical inertia and future setpoints to provide robust, actionable insights for industrial regulation. By combining interpretability tools such as SHAP and ALE with process knowledge, we identified the complex nonlinear factors affecting SO₂ concentration fluctuations. The interpretability analysis reveals a high sensitivity of emissions to temperature gradients, suggesting that coordinated control of the gasification and combustion stages is key to emission suppression. Specifically, the results indicate that optimizing steam pressure to approximately 0.28 – 0.30 MPa, gasifier outlet temperature to about 100 – 160 °C, and combustion furnace temperature to about 800 – 900 °C can maximize resource recovery while effectively reducing SO₂ emissions. In conclusion, by integrating process mechanisms with advanced data-driven analysis, this study achieves precise emission prediction and operational optimization for sludge gasification and provides a generalizable methodology for intelligent modeling of other dynamic industrial systems.