Feature Selection and Long-term Modeling for the Blast Furnace Ironmaking Process Based on Random Forests

Abstract

So far, the accurate modeling and control of blast furnace iron-making process (BFIP) is still an open problem due to its excessive complexity. Aiming at the issue of long time-delay and strong cross-coupling characteristics of BFIP, the random forests (RF) algorithm is introduced for predicting the silicon content in hot metal, which is the most key indi- cator of inner state of blast furnace. In the proposed model, both short and long-term BFIP features are adopted as inputs, without variable pre-selecting, to modeling the long-term dynamics of BFIP. Simulation results show that the RF algo- rithm can successfully identify the importance of different features (the latest silicon content in hot metal obtains the larg- est value of importance), can effectively decrease the effect of the redundancy and cross-coupling among variables. The RF model also can achieve similar or better prediction performance compared with support vector machines (SVM), which indicates that it is potential to modeling such BFIP-type complex industrial process using RF algorithm.