Ensemble Pattern Trees for Predicting Hot Metal Temperature in Blast Furnace

Xinmin Zhang,Manabu Kano,Shinroku Matsuzaki

DOI: https://doi.org/10.1016/j.compchemeng.2018.10.022

2017-01-01

Abstract:A blast furnace is a huge counter-current reactor, which is used for reducing iron ores to hot metal in the ironmaking industry. Its operation is difficult due to the internal harsh operating environment and the complex chemical-physical reactions. In the blast furnace, hot metal temperature (HMT) is an important product quality and also an important thermal state indicator, therefore HMT needs to be controlled at a desired level so that stable and efficient operation is achieved. To realize this goal, a pattern trees-based virtual sensing technology is proposed in this study. Pattern trees regression is a nonlinear modeling technology, which constructs a hierarchical and tree-structured functional relationship between the process variables and HMT in an interpretable way. Pattern trees uses triangular norms, triangular conorms, and arithmetic operators while the conventional decision trees uses only triangular norms operators. Thus, pattern trees is more flexible and efficient for data-mining and constructing the prediction models. We applied pattern trees to build a soft-sensor, or virtual sensor, of HMT by using real operation data of a blust furnace. The results have shown that pattern trees outperforms the conventional methods such as PLS, decision trees, and random forest both for the current time prediction and for multi-step-ahead prediction.