Optimisation of sinter plant operating conditions and BF burden material resources using advanced multivariate statistics

The purpose of the project was to develop an ironmaking production strategy, which was able to define quality and production targets, whilst minimising the costs of sinter material and coke blends. This was to be achieved by developing relationships to differentiate between the effect of blend and plant operations and the effects on quality and productivity. Thus an improved operating window for process operations and blend optimisation would be achieved. The classification of coke and sinter ores would improve confidence in assessment of new ores, and seek to minimise the need for undertaking plant assessments which would be costly and which could have a detrimental effect on quality and productivity and therefore lead to furnace upsets. The project has improved the ability to differentiate between the effect of blend and plant operations and their relative effects on quality and productivity. This has enabled improved windows for process operation and blend optimisation to be defined. These have been related to blast furnace requirements having established the effects that sinters of different physical, mineralogy and chemistry and coke have on blast furnace operation. By defining the sinter and coke requirements for the blast furnace operating regime, sinter/coke quality and productivity have been optimised as a function of cost and energy usage. Extensive database analysis was employed to investigate methods to identify problems in operating conditions of the sinter plant, especially due to quality, the accuracy of property measurements, and capacity problems. Suggested solutions include the restriction of materials and additions etc, (raw material preparation, fuel, additions, capacity changes). The methods investigated have included multivariate statistical techniques, Principal Component Analysis, Linear Programming, Genetic Programming, Neural Networks in their differing forms. Principal Component Analysis has been used for a multivariate statistical approach to classifying sintering ores. It was shown that it was possible to separate the ores into different clusters, highlighting those which were interchangeable since resident in the same cluster. The classification of sintering ores by use of data clustering techniques gave the ability and confidence to assess new ores without undertaking pilot or plant based assessments. In practice it enables more focused testing to be achieved. Cluster techniques including K means clustering are demonstrated, showing it is possible to classify similar cokes, thus giving a means of automatic development of a model which will be able to identify the coals that will behave in a similar manner in the production process. Linear Projection to Latent Structure models are developed that give good predictions of sinter productivity and quality. The weightings of the model when examined in detail showed which inputs had the greatest effect on productivity and sinter quality parameters. In many cases the main factors influencing productivity and quality were as expected by the plant technologists and thus more acceptable as replacements to existing models. Genetic programming methods are compared to linear PLS models generated from the same data sets. The technique is shown to give better predictions for a number of the sinter quality variables. It was shown that strong non-linear relationships, in particular ISO tumble, existed between the process and the quality parameters.