Gaseous fuel mixture parameter prediction in a steel works using minimalist neural networks.

This thesis reports a developmental programme of work undertaken for a number of installations at the Scunthorpe site of Corus plc (formerly British Steel plc). The central theme concerns the optimisation of the combustion of gas mixes produced by onsite processes through parameter prediction, the constituents of the mixtures included blast furnace gas (BFG), coke oven gas (COG), and basic oxygen steelmaking gas (BOS).The main parameters under investigation were calorific value (CV), air/fuel ratio, and specific gravity. Finally, a secondary investigation was conducted into predicting oxygen content in flue gases with a view to reducing recuperator corrosion. Data from three different systems was considered: a power station, a coke oven gas plant and a section mill. All data sets were subject to aliasing both as a result of the slowness of the mass spectrometer measuring devices with respect to gas content fluctuations, and as a result of the relatively long sampling interval employed by the main archiving system. The sets from the section mill were particularly prone to abrupt and extreme variations. Neural network solutions based on function approximation were proposed and developed. There was a specific requirement that any solution be compatible with unsophisticated low-budget hardware. Hence there were major constraints on network size and complexity. A linear time series based network was found to perform more efficiently in the data supplied rather than the more conventional non-linear counterpart. The proposed networks indicated potential gains in accuracy in excess of 50% over a second-order least squares-based method proposed by the collaborating organisation. At the time of writing it is understood that no other similar systems have been investigated in this manner, let alone resulting in a successful minimalist neural network solution. Hence the contribution to knowledge is that it is possible to accurately predict the above parameters with a minimalist linear network, trained with data subjected to varying degrees of aliasing.