Abstract: As a key indicator in blast furnace production, the stability of the blast furnace gas utilization rate directly affects the economic benefits of ironmaking enterprises. To this end, this study proposes a blast furnace gas utilization rate prediction method that integrates mechanism modeling with data-driven approaches. First, the operational data from the blast furnace process were collected. Abnormal and missing data were preprocessed by combining big data techniques with process theory. Mechanism-derived features were constructed from multiple mechanistic models to capture the time-lag effects and state-accumulation characteristics of the blast furnace operation, including the burden composition, regional ore-to-coke ratio, and state transitions. After integrating these mechanism-derived features with the time-series features, a light gradient boosting machine combined with time-series cross-validation was employed to select the 18 features most relevant to the blast furnace gas utilization rate. These selected features were further analyzed in detail from a process perspective and were confirmed to be suitable as inputs to the model. Finally, a blast furnace gas utilization rate prediction model was developed by integrating a gradient boosting decision tree, Mamba, and a temporal convolutional network. By leveraging the strengths of these algorithms, the model effectively captures the time-lag effects, state-accumulation effects, and short-term fluctuations in the blast furnace operation. The superiority of the proposed model and the effectiveness of the constructed features were verified through ablation experiments. The results show that the model achieves a prediction hit rate of approximately 92% within an error margin of ±0.3%. After industrial deployment, the model delivered an increase of approximately 1 percentage point in the blast furnace gas utilization rate, which translated into enhanced economic returns for the company.