Application of a genetic algorithm based model selection algorithm for identification of carbide-based hot metal desulfurization

Abstract

Sulfur is considered as one of the main impurities in hot metal. Hot metal desulfurization is often carried out with pneumatic injection of a fine-grade desulfurization reagent using a submerged lance. The aim of this study was to develop a data-driven model for the process. The model selection algorithm carries out a simultaneous variable selection and optimization of number of hidden neurons with a combination of binary and integer coded Genetic Algorithm. The objective function applied in the search is repeated Leave-Multiple-Out cross-validation. The model considered is a feedforward neural network with a single hidden layer. In the inner loop of the algorithm, the computational load is reduced by making use of Extreme Learning Machine (ELM) architecture. The final model is trained using the Bayesian regularization. The results show that a well-generalizing data-driven model with good prediction performance can be repeatedly selected based on noisy industrial data with the help of a Genetic Algorithm, provided that the model is validated comprehensively with internal and external data sets.