A model of the thickness of burden layers in the ironmaking blast furnace is presented. Local layer thickness estimates are calculated on the basis of signals from stockrods that measure the burden (stock) level in the furnace. These estimates are used in developing a model for the relation between the layer thickness and variables such as stock level and movable armor settings. Because of the nonlinear dependence of the variables, the models are based on feedforward or recurrent neural networks. The network size is carefully selected based on a cross-validation procedure. The resulting neural model is first studied by analyzing its predictions for different inputs. By further introducing a simplified scheme for considering the practical constraints of the charging process, an autonomous model, where the neural network plays an important role, is formed. This hybrid model is applied to yield insight into the dynamics of the layer formation process; the effect of movable armor settings, stock level and burden descent rate are analyzed and compared with practical experience.KEY WORDS: blast furnace; burden distribution; neural network; hybrid model. signals from the stockrods (see Fig. 1), which measure the stock level close to the wall. Stockrod signals have been logged at high frequency and processed to provide burden layer thickness estimates. The relation between the layer thickness and variables affecting the burden distribution has been modeled by feedforward and recurrent neural networks. The study shows that the variation in the burden layer thickness can be described by the networks, given information about the stock level at the instant of charging and the movable armor position. However, the neural model requires information about the stock level, which is a variable that is affected by the layer thickness of the dumps, i.e., the output of the model. In order to resolve this dilemma, the network has been implemented in a hybrid model, which considers the effects of the descent of burden and practical constraints of the charging process. By this procedure, the dependence between the stock level prior to charging and the resulting layer thickness can be considered. The hybrid model can be used to simulate the effect of stock level set-point, movable armor positions and burden descent rate on the layer thicknesses. The results obtained conform well with observations from the Finnish blast furnace studied in this work and are also in general agreement with practical experience. MeasurementsThe stockrods in the blast furnace are sounding devices (as illustrated schematically in Fig. 1) that sense the burden level after each dump and are elevated before a new dump of burden is charged into the furnace. The main function of the rods is to measure the vertical position of the bed surface (stock level) in order to trigger the charging of the burden at an appropriate moment to maintain a stable stock level. In practice, charging is triggered when the rods have descended below a given set-point. From the stockrod ...

A model of burden layer formation in the blast furnace is developed on the basis of layer thicknesses estimated from radar measurements of the burden (stock) level in the furnace. The dependence between the layer thickness and charging variables is modeled by neural networks. Parsimonious networks are determined by an evolutionary algorithm, which simultaneously trains weights and network connectivity. The efficiency of the training procedure is enhanced by tackling part of the numerical optimization by linear least squares. The resulting network models are utilized in a hybrid model, which considers practical constraints of the charging process in the furnace. The hybrid model is used to evaluate the impact of altered boundary conditions in novel charging programs.

The burden distribution in the blast furnace is estimated on the basis of measurements provided by the stockrods in combination with information about changes in the gas temperatures measured by an above‐burden probe after each dump of burden. The former measurements yield an estimate of the local layer thickness close to the wall while the latter ones are used to evaluate the layer thickness in the center of the furnace. The layer thickness values estimated by the method and values computed on the basis of geometrical considerations under simplifying assumptions were found to show good agreement. The results hold promise for a successful on‐line estimation of the burden distribution in an operating blast furnace.

A neural network-based model of the burden layer thickness in the blast furnace is presented. The model is based on layer thicknesses estimates from a single radar measurement of the burden (stock) level in the furnace and describes the dependence between the layer thickness and key charging variables. An evolutionary algorithm is applied to train the network weights and connectivity by optimizing the model structure and parameters simultaneously, tackling part of the parameter estimation by linear least squares. This enhances convergence and results in parsimonious and transparent network models with actions that can be explained. Finally, the networks are used in a hybrid model for analyzing novel charging programs and for studying the limits of the charging process.

No abstract