Modelling noisy blast furnace data using genetic algorithms and neural networks

Noisy blast furnace data from a Finnish steel plant was modelled by artificial neural networks, which relied upon a novel Genetic Algorithm for training. It allowed the neural networks the flexibility of evolving their optimum architectures both in terms of their weights and the utilized neurons and neuron connections. The important alloying elements in the hot metal, C, S and Si, were monitored as a function of five input variables related to the two reducing agents: coke and injected oil. The analysis indicated an intricate interaction between the variables and also highlighted the importance of lagged data in describing the complex relations. Despite these complexities the models developed were able to quantify relationships that have been generally observed and reported in the literature.