Abstract
A machine learning model was developed to predict the desulfurization process during the Kanbara reactor hot metal treatment. Compared with other algorithms, the LR algorithm model exhibited the smallest error in current calculations, which was used to predict the final S content with various operation parameters. The final S content in the hot metal obviously rose from 0.001% to higher than 0.003% with the increase of the initial S content from 0.03% to 0.06%, while it decreased from 0.003% to below 0.001% with the increase from desulfurizer addition from 4 kg/ton to 7 kg/ton. The final S content changed little with the increase of C content, Mn content, and rotation speed. The feature selection using RReliefF algorithm was conducted to evaluate the correlation between inputted parameters and outputted final S content. The addition of desulfurizers was beneficial to improve the desulfurization efficiency, while it obviously increased desulfurization costs. The longer desulfurization time lowered the S content, while it resulted in higher desulfurization costs due to the refractory erosion and electric power consumption.