Blast Furnace Thermal State Prediction Based on Multiobjective Evolutionary Ensemble Neural Networks

Blast furnace ironmaking is the largest energy-consuming and greenhouse gas-emitting process in the iron and steel industry. As a key indicator of blast furnace operation status and energy consumption level, silicon content prediction has been very important for blast furnace operators to save energy and ensure stable ironmaking production. Traditional data-driven methods usually use feature selection when building models, which cannot guarantee high accuracy and generalization ability of the model due to the high correlation between features. To solve this problem, we propose an ensemble model based on multiobjective differential evolution, which consists of two stages. In the first stage, a multiobjective differential evolution is employed to evolve the input feature weights of each base-learner, and feature pruning is embedded in the evolution process to achieve sparse selection of all available features for the base-learners. In the second stage, a two-layer ensemble method based on autoencoder and extreme learning machine is proposed to achieve nonlinear ensemble of good candidate base-learners. Experimental results based on actual production data show that the developed model outperforms previous studies and can achieve more accurate values of silicon content, which in turn helps to reduce energy consumption in actual production. Graphical Abstract The modelling process of the proposed silicon content prediction model consists of two stages. In the first stage, a multi-objective algorithm is employed to evolve the input feature weights of each base-learner. In the second stage, a two-layer ensemble method is proposed to achieve nonlinear ensemble the base-learners.