Nonlinear prediction of the hot metal silicon content in the blast furnace

The processes in metallurgical industry are often extremely complex and measurements from their interior are scarce due to hostile conditions. Today's constraints on high productivity and minor impact on the environment still require that the processes be strictly controlled. Mathematical models can play a central role in achieving these goals. In cases where it is not possible, or economically feasible, to develop a mechanistic model of a process, an alternative is to use a data-driven approach, where a black-box model is built on historical process data. Feedforward neural networks have become popular nonlinear modeling tools for this purpose, but the selection of relevant inputs and appropriate network structure are still challenging tasks. The work presented in this paper tackles these problems in the development of a model of the blast furnace hot metal silicon content. A pruning algorithm is applied to find relevant inputs and their time lags, as well as an appropriate network connectivity, for solving the given time-series problem. In applying the model, an on-line learning of the upper-layer weights is proposed to adapt the model to changes in the input-output relations. The analysis shows results in good agreement with findings by other investigators and practical metallurgical knowledge.