Abstract: By means of the alloying data of 100 t converter at tapping, a model for ferrosilicon addition was established by SVR (Support Vector Regression) algorithm through data preprocessing and the feature selection with Pearson correlation coefficient in this study, where steel grade, a non-numerical feature, was used as a feature variable. The hit rates of ferrosilicon addition prediction from the SVR model with steel grade as a feature variable are 94.84 %, 87.58 % and 75.77 % in the errors range of ±40 kg, ±30 kg and ±20 kg, respectively, while those of the model without steel grade as a feature variable are 88.4 %, 80.61 % and 65.85 % under the same errors range, which proves that the model with steel grade as a feature variable has higher prediction accuracy, and has better reference value for actual alloying process.

Key words: converter steelmaking, alloying at tapping, ferrosilicon addition, prediction model, nonnumerical feature, hit rate