Blast furnace smelting is a dynamic process with high complexity, chaos and time delay. In industry, molten iron silicon content is often used to feed back the fluctuation of blast furnace temperature and thermal state. The skew projection depth can reflect the outliers of the data well when the data is biased, and it is very robust in the classification calculation of high-dimensional data. Firstly, 11 influencing factors are determined as input variables through differential processing and correlation analysis in this paper, which are used to study the relationship between changes of various variables and changes of silicon content. Then, the data whose projection depth value is outside 90% confidence interval are regarded as outliers and classified into stable and outliers. Finally, the stable data are predicted by Elman neural network prediction model, and the outliers are classified and predicted by Logistic model under different fluctuation directions of furnace temperature. The simulation results show that the prediction accuracy of stable class 157 furnace is up to 85.3%, and that of outlier class is up to 82.6%.