Modeling Silicon Content In Molten Iron Of Blast Furnace Based On Neural Network

As our country’s basic raw material industries, iron metallurgy is one of the pillar industries of Chinese national economy. As the main process of iron and steel the blast furance iron making is an important part of the industry. Moreover, the blast furance iron making plays a vital role in influencing the development of the steel industry status and the enery saving. Whether the furnace condition is smooth influences the enery saving and eruption reduction of the blast furance ironm making process, directly. In most cases, the temperture of the blast furance is used to determine the condition of the blast furance. Therefore, the temperture of the blast furance is an very inportant indext for condition of the blast furance. The silicon content in hot metal can be taken as the characteristic index of the temperature of the blast furance. Therefore, the reliable predicition model for the furance temperature not only has an important theoretical value, but also has important practical value. However, the complexity of the blast furance caused by nonlineaity, big noise, distributed parameters and so on brings gret difficulties to the prediction of silicon content. As a kind of self learning network, neural network can solve the nonlineaity and big noise problem. Until now, neural network has been applied in pattern recognition, prediction control, functional approximation and so on widely. In this dissertation the key problem on modeling of the silicon content is considered by employing the neural network. The main research works are summarized as follows:Firstly, the related data of the blast furnace ironmaking are analyzed and processed. The expert experience and data statistical analysis are combined to determine the input variables for the silicon content in hot metal. Furthermore, the multivariate linear regression method is applied to determine the related parameters of the delayed sequence. All the data have been processed two times. The normalization processing is also conducted for these data.Secondly, consider the shortage of BP neural network that the BP neural traps into local optima easily, an improved method is proposed. The simulation date analyzed and processed in advance. The comparisons with the existing BP neural network are conducted. The simulation results demonstrate the superior performance of the propsed improved BP neural network.Finally, a new extreme learning machine algorithm is presented. The shortage existing in general extreme learning machine can be overcomed with the propsed new algorithm because of the employment of regularization method.Then this new algorithm is used in modeling of blast furnace temperature. The effectiveness of the proposed algorithm is proved with the simulation results. Moreover, the superior performance of the improved extreme learning machine.