Machine learning assisted identification of grey-box hot metal desulfurization model

Abstract

Hot metal desulfurization operates as an extraction process in steel production after the blast furnace. Mathematical modeling of the process provides the basis for control and optimization solutions. Owing to complex dynamics, an exhaustive physico-chemical description of the process is computationally infeasible. Thus, a model with a gray–box structure is considered. In the model, a neural network model for sulfide capacity of the slag is implemented in series with the reaction model. The network is trained using an independent data set. The parameters for the reaction model are identified using genetic algorithm with real-encoded population and hybrid selection and recombination operators. The gray-box model parameter distributions are estimated using Metropolis Hastings Algorithm. This study suggests that a sufficiently low mean absolute error can be achieved for the end sulfur content. The relative contribution of the permanent contact reaction to overall reaction rate under industrial conditions remains uncertain.