Abstract: Element yield prediction is a key and difficult problem of ladle furnace (LF) alloying. For improving its precision, main factors of element yield were determined by the mechanism analysis method. Because these factors cannot be readily available in production, factors were obtained indirectly by the feature construction method based on grammatical evolution (GE). Finally, the element yield prediction model was established using constructed features as inputs. Original GE algorithm was improved on the basis of the background issue. The improved method made full use of the existing features, and solved the problem that important data cannot be detected in the smelting process. The proposed method had been applied to element yield prediction of Mn and Si element of Q345B. Experimental results showed that the prediction accuracy could be improved effectively by utilizing the proposed method.