Effect of Feature Selection on the Prediction Model of FeO Content in Sinter

The prediction of FeO content in sinter is important for operators to adjust the raw material ratio and process parameters. Feature selection is used to choose the original features that are strongly correlated with the target parameters, which is considered as one of the essential steps in the predictive model. This work uses the three feature selection methods of Pearson (P), SVM-RFE (SR), and Random Forest (RF) to explore the impacts of feature selection on the generalization performance and predictive capability of the Tent-SSA-BP model. The results of experiments show that the RF feature selection is more suitable for the sinter FeO content prediction. The range of prediction error is stable at − 1.67 %~ 2.37%, and 99.0% of the predicted value is within the error of ± 0.2 of the actual value. The model r2 reaches 0.921, and MSE (0.009), MAE (0.082), and RMSE (0.094) perform well, while the training time of the model (13.97) is short. This prediction model can be used as the forewarning system tool in giving FeO content information to operators to formulate improvement strategies in a timely manner.