Abstract: Models of oxygen consumption in a 45 t converter were set up based on actual production data with BP neural network, support vector regression machine (SVR) and LGBM(Light Gradient Boosting Machine) algorithms where their parameters were optimized with Bayesian optimization algorithm and the models′ features were selected by data preprocessing and mutual information methods. The actual production data of 1 176 heats were used to train these models, and the data of another 504 heats were used to verify the prediction results of the models. The results showed that the hit rate with LGBM model within ±50 m³, ±40 m³ and ±30 m³ was 94.04 %, 85.91 % and 76.58 %, respectively. Compared with the support vector machine and BP neural network models, the LGBM model had higher prediction accuracy and stability and better generalization ability. 

Key words: converter steelmaking, oxygen consumption, data drive model, hit rate