Prediction of Average Iron Silicon Content Over the Blast Furnace Smelting Cycle Based on a Model Combination Framework

Abstract

Aiming to address the issues of missing monitoring indexes, unstable prediction models, and unreliable single-point predictions of average molten iron silicon content during the blast furnace smelting cycle, this paper proposes a multi-objective optimization combination strategy based on an absolute prediction error-variance framework for the prediction model of average molten iron silicon content. First, the midpoint sequence is extracted according to the smelting cycle, and the sequence is reconstructed by ensemble empirical mode decomposition (EEMD) and the Hurst method. Then, the Elman recurrent neural network (Elman), support vector regression (SVR), and long short-term memory (LSTM) models are used as sub-models, and the non-dominated sorting genetic algorithm II (NSGA2) algorithm is employed to optimize the combined model weights. Finally, the prediction intervals are constructed indirectly based on prediction error. It is found that the established combined prediction model achieves accurate and stable prediction results on the iron silicon content, and the kernel density estimation method can provide a better quality prediction interval. The prediction intervals at appropriate confidence levels can be selected based on the furnace manager’s prediction of the average chemical heat state in the next smelting cycle. Based on the best-predicted results: On the point prediction, the hit rate in the range of plus or minus 0.3% is 96.4286%; On the interval prediction, the coverage of the prediction interval is 92.8571%, with an average width of 0.0631% and an average width deviation of 0.0027. The average silicon content within the controllable range can realize extensive cost-effectiveness and carbon emission reduction benefits.