In the basic oxygen furnace (BOFBasic Oxygen Furnace (BOF)) steelmaking process, recovering waste heatWaste heat from high-temperature off-gases to generate steam plays a crucial role in enhancing energyEnergy efficiency and reducing production costs. This study proposes a hybrid data–mechanism approach for dynamic predictionDynamic prediction of BOFBasic Oxygen Furnace (BOF) waste-heat steamWaste-heat steam. Starting from historical steam data, a framework of key influencing factors is established by incorporating the mechanism of off-gas generation. Correlation analysis is conducted to identify relevant factors, enabling variable selection and dimensionality reduction. An adaptive, structurally improved deep neural networkNeural network (DNN) model is developed, in which a tolerance-based error function replaces the conventional loss function. Considering the similarity of steam recovery curves across heats, clustering is employed to transform the predicted total steam recovery of each heat into continuous flow curves, which are further aligned with production schedules to generate BOFBasic Oxygen Furnace (BOF) steam flow forecasts. The model performs real-time rolling predictions by continuously updating with production data, thereby improving accuracy. SimulationSimulation results demonstrate that the improved DNN achieves a mean absolute error (MAE) of 0.95 t, significantly outperforming a pure data-driven modelData-driven model (MAE 2.21 t), the XGBOOST model (MAE 1.80 t), and the backpropagation neural networkNeural network (BPNN) model (MAE 2.43 t). For the predicted BOFBasic Oxygen Furnace (BOF) steam flow curve, the MAE is 15.06 t/h with an R2 of 0.94. These results provide a solid theoretical and technical basis for precise dynamic regulation and efficient utilization of BOFBasic Oxygen Furnace (BOF) waste-heat steamWaste-heat steam.