An Explainable Data-Driven Framework for Energy-Efficiency Optimization in Paper Drying

Paper drying consumes >60% of total papermaking energy and remains a major bottleneck to low-carbon production. However, steam-system optimization in full-scale paper mills is hindered by fluctuating operating conditions, nonlinear process coupling, and limited model interpretability. This study develops an explainable plant-scale data-driven framework that integrates operating-regime partitioning, MLP-enhanced T-PLS prediction, four-level energy-efficiency state identification, and SHAP-based root-cause diagnosis. Using real production data from a full-scale paper mill, the drying process was divided into five operating regimes. Across these regimes, the proposed model reduced the prediction RMSE by 70.27% relative to the T-PLS baseline, while the AdaBoost classifier achieved a mean accuracy of 84.14% in identifying energy-efficiency states. SHAP-informed parameter reconfiguration reduced specific steam consumption by an average of 1.37% across four validated regimes (up to 2.90% per regime). Scenario analysis suggests that implementation in technically comparable large-scale Chinese paper mills could save approximately 1.71 million metric tons of steam annually. By extending conventional soft-sensing from prediction to interpretable, mode-specific decision support, the framework provides a scalable pathway for low-carbon optimization in papermaking and other heat-intensive continuous processes.