Predictive Models Of BF Hot Metal Silicon Content Using Wavelet-networks

As the main upper procedure of metallurgical industry, Blast Furnace(BF) ironmaking is an important component of the pillar industry in national economy, which plays a significant role in energy saving and technical development of the whole industry. The process of ironmaking is highly complicated, whose operating mechanism is characteristic of nonlinearity, time lag, high dimension, big noise and distribution parameters etc.To maintain the appropriate temperature in BF is crucial for a smooth ironmaking process. Due to the high complexity, it is almost impossible to measure the exact temperature of hot metal in BF. Thus the difficulty of ironmaking process automation lies in the construction of an effective prediction control model to make forecast of temperature of hot metal in BF. Silicon content in hot metal is not only an important index to evaluate the status of ironmaking process and quality of hot metal, but also the most important parameter to reflect temperature of hot metal in BF. To effectively control the BF ironmaking process, prediction of the silicon content becomes an important question in discussion.The method of wavelet-networks is relatively new in time series analysis, which inherits both the virtue of wavelet in regionalizing of time-frequency field and self-learning ability of neural networks and has strong capability in function approximation and error tolerance. There are two combinational methods for wavelet and neural networks. The first one-called wavelet networks with relaxed structure is to deform the input series into different levels of resolution scales by the wavelet transform, and for each level of corresponding input parameters, we use neural networks to get the output values. A reconstructive transformation is applied to the output values of neural networks and the forecasts are generated. The other-called wavelet networks with compact structure is to substitute the base function in neural networks with wavelet function, and the wavelet function is adjusted adaptively to do wavelet transform and new networks are enabled.This paper uses data collected from Intelligent Control Expert System in Handan Steel Corporation as sample data to implement the algorithms. The sample contains 500 data pairs. We apply the three models based on the theory of wavelet analysis and neural networks to do forecasts and the results are analyzed in detail. It is shown that the method of neural networks with compact structure can give better hit rates of prediction for silicon content. The result is helpful for operating of BF.