Emissions prediction and mechanistic interpretation for sludge gasification via multi-source temporal modeling

Sludge gasification technology has emerged as a promising method for sludge treatment, offering significant advancements in resource recovery, pollution reduction and carbon mitigation. However, the generated air pollutants limit the widespread adoption of this technology. We construct a timeline-based framework using actual production data with minute-level resolution. This framework integrates historical emission concentrations, operational parameters (e.g., temperature, pH), input materials (e.g., gasifying agents, steam), as well as predetermined operational settings and inputs (e.g., steam flow rate, gasification agent pressure and fan frequency) to construct multi-source features, while constantly rolling the lookback and forecasting window to dynamically predict the SO₂ concentrations, oxygen content and particulate matter. On this basis, we systematically evaluate four machine learning models and four deep learning models using the Python-based Darts time series library, and use interpretability including SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDP) in conjunction with process knowledge to analyze the underlying causes of SO₂ fluctuations. The results suggest that model performance is closely related to the generation mechanisms of pollutants: TiDE performs optimally in predicting SO₂, which has strong nonlinearity and hysteresis characteristics, while the timeline-based Light Gradient Boosting Machine (LightGBM) model demonstrates robust performance across multi-tasks. Additionally, the gasifier outlet temperature, pressure and steam flow rate are key factors for SO₂ fluctuation, with their effects exhibiting a lag window of 5–8 min. Notably, the synergistic effect of low steam flow rate (<300 kg/h) and high outlet pressure (>4.0 kPa) contribute to a steep increase in SO₂. By integrating data science with process mechanisms, this study aims to provide accurate and reliable decision support for proactive control and parameter optimization, thus enhance the environmental friendliness and economic feasibility of sludge gasification.