Improving the performance of the blast furnace ironmaking process

Sammanfattning: In order to assist the operators' control activities and to further improve the performance of the blast furnace, efforts have been made to develop models for control of blast furnace operation and improvements of burden quality. The main objectives of the work were to design models for diagnosing and forecasting irregular furnace statuses and for predicting the silicon content of the hot metal, and to study the influence of the fluxes on the melting properties of fluxed pellets. A model for diagnosing the process status, consisting of six sub-models - recognition of the top gas profiles, classification of the heat flux distributions, prediction of the slips, comprehensive evaluation of the furnace process, diagnosis of channeling and cool furnace thermal state, has been designed. The off-line test results indicate that these models can detect and predict some upcoming and existing irregular process statuses, e.g. detecting the irregular top gas distribution, predicting slips three hours in advance and cool furnace thermal state two hours in advance. The most important operational parameters for predicting the upcoming slips are also extracted. A hybrid model for predicting the silicon content, consisting of a knowledge-based system and perceptron networks, has also been developed. The knowledge-based sub-system evaluates the process conditions and determines the applicability of a sub-model for forecasting. When the furnace operation is judged as normal, neural network models will make the predictions. When some irregular process statuses occur in the process, the knowledge-based system will perform the forecasting tasks. Test results show that the hybrid model for predicting the silicon content can make the forecasts about two hours in advance under various conditions, except the occurrence of serious irregular process statuses, e.g. more than 2 meters slip. The hit rate - a ratio of correct predictions - reached about 75% and 86% on acceptable prediction errors ±0.05% and 0.08% Si, respectively. Concerning the networks used for designing the model, test results have shown that a three-layer perceptron with two middle nodes trained with the algorithm - back-propagation with momentum, could give the 'best' prediction capability of the model. Algorithms - Quick Propagation and Resilient Propagation, can accelerate the training but cannot enhance the prediction ability of the model. Training using moving 'windowed data' can generate similar prediction results and enable online automatic update of the model. However, this approach demands greater computer resources. Radial basis Function network did not produce a better model. Softening and melting experimental studies show that the addition of basic fluxes, especially BOF-slag to self-fluxed pellets can considerably worsen the melting properties of pellets, even entirely block the separation of metallic iron from slag. The main reason is the high basicity and low FeO content of the slag, formed by the primary slag of pellets and fluxes in the course of melting. Therefore, it can be concluded that when using high-Fe self-fluxed pellets as the main iron-bearing burden, top-charging fluxes is not favorable to blast furnace operation.