In the basic oxygen furnace (BOF) steelmaking process, accurate determination of the molten steel composition is a critical step in determining the tapping point. Currently, this decision relies primarily on operator experience, supplemented by manual sampling and laboratory analysis. However, such an approach not only limits production efficiency but is also subject to human error. To reduce the influence of subjective judgment, an improved random forest (RF) model optimized by the grey wolf optimization (GWO) algorithm was proposed. Using a 120-ton converter at a steel plant as the research object, multiple process parameters were selected as input features, including the hot metal weight, scrap ratio, blowing time, Si, Mn and P contents of the hot metal, hot metal temperature, the converter operation parameters, and the consumption of oxygen, argon, and nitrogen. The model enables real-time prediction of the endpoint concentrations of C, Si, Mn, P and S in molten steel. The model was trained and dynamically updated via 1 783 sets of actual industrial data. Through hyperparameter tuning, the prediction time is reduced to 0.1~0.3 s, with a prediction accuracy exceeding 90%. While improving generalizability and stability, the model achieves fast and reliable prediction of steel composition and significantly reduces dependence on manual decision-making.