Integrated deep learning-driven multi-stage steam forecasting and scheduling optimization for converter energy systems

Converter steam is a valuable secondary energy in steelmaking, and improving its recovery and utilization is essential for reducing energy cost. However, converter steam data are inherently time-segmented: before steelmaking starts, only limited planning information is available, while during production, high-frequency process signals drive rapid steam fluctuations. Existing studies mainly rely on single-stage time series forecasting, and rarely combine planning forecasts with sequential production forecasts to support steam scheduling. To address this gap, this study proposes an integrated two-stage, multi-time-scale forecasting framework for converter steam. First, the planning-to-production nature of converter steam generation is explicitly formulated, and a time-segmented forecasting workflow is established to link long-horizon planning prediction with short-horizon rolling prediction. Then, in the planning stage (approximately 2 hours ahead), a hyperparameter-optimized SVM model (IBKA-SVM) is employed to generate interval steam output forecasts, providing a reliable baseline for upper-level scheduling tasks such as steam network target setting and buffer allocation. Finally, in the production stage, a DSC-Transformer model performs fine-grained rolling steam-flow prediction based on time-series production data, and its outputs are used to dynamically update and correct the planning-stage baseline for operational control. Experiments on industrial converter data demonstrate that the proposed models achieve strong predictive performance (IBKA-SVM: 92.3%, DSC-Transformer: 98.7%) and consistently outperform conventional baselines. The proposed framework provides practical, multi-time-scale predictive support for converter steam scheduling and energy-efficient operation.