Abstract:Considering the characteristics of multivariables and large sample in BOF endpoint neural network forecasting model, a method of outliers identification and judgment is introduced. Using this method, the data set of neural network model is selected, and the result of selection is primarily validated. After data pretreatment with this method, the mean and maximum residual absolute value of training set and testing set are decreased by 267%、41% and 10%、45% respectively. The result shows that the method is effective to improve the neural network model for BOF endpoint forecasting.