Interpretable Spatiotemporal Coupling Network for Robust Emission Prediction and Implicit Mechanism Mining in Sludge Gasification

Abstract Thermochemical gasification is a vital green technology for sewage sludge reduction and resource recovery, yet real-time emission forecasting remains exceptionally challenging due to chaotic operational fluctuations, pronounced time lags, and noisy sensor data. Conventional artificial intelligence models often rely on purely empirical mappings that lack structural and physical consistency. To overcome these limitations, we develop a process-informed spatiotemporal coupling network (PISTC-Net) that integrates directed industrial topologies with feature-level optimal time lags. Validated on an industrial scale, the framework achieves high forecasting fidelity, yielding R2 values of 91.18, 80.37, and 78.09% for SO2, NOx, and particulate matter (PM), respectively. Under extreme stress testing simulating abrupt sensor failures, PISTC-Net demonstrates remarkable robustness by leveraging learned process inertia to accurately track emission trajectories. Crucially, without prior mechanistic labels, the embedded hierarchical attention mechanism spontaneously learns spatiotemporal attention patterns that are consistent with process mechanisms. The model adaptively shifts focus from upstream thermochemical zones during SO2 forecasting to full-process collaborative analysis for PM, capturing the dynamic spatiotemporal propagation of operational disturbances 5–8 min prior to emission peaks. This work offers practical insights for the optimization and environmental compliance of complex engineering systems in sludge gasification and analogous industrial processes.