The Research Of Endpoint Carbon And Temperature Control For Converter Steelmaking On The Basis Of Neural Network

Steelmaking is an important process in iron and steel industry, which plays a decisive influence to improve the quality of products, reduce the production cost and expand product range. Converter steelmaking as modern mainstream production mode, its main task is to get the qualified temperature and composition of molten steel when smelting process reaches to the end, namely, endpoint control. Generally speaking, the endpoint control of converter steelmaking is endpoint carbon content and endpoint temperature control. The steel industry in our country is under restrictions like small and medium converter being in the majority, fund and equipment shortage etc. So the static model endpoint control is given priority to be adopted in our country, except a few large-scale steel enterprises which have a conditong to carry out dynamic or full automatic control. However, regardless of the dynamic endpoint control or the full automatic control mode, they are all on the basis of static model control. Therefore, the research of endpoint static control model for converter steelmaking is still has a far-reaching influence whether in our country or within the scope of the world. The core and key of static control is:accurate static model. Accuracy and practicability of the model is proportional.In this article, we aim at to improve endpoint control when smelting process reaches to the end for the first time. A static statistical model is built first, then we built a optimization control model based on BP network forecasting model by using the neural network predictive control theory. Data is the foundation of model. So firstly we adopt a series of data preprocessing operation to the collected actual production data, in order to build smelting data sample library. The operation of characteristics variable selection based on multidimensional mutual information theory provides a universal framework which does not depend on the specific problem object to identification model of the key input selection problem for all types of multiple input and multiple output complex system. Based on static statistical modeling principle and the effect of combined model preceding a single model theory, we established a combined model of non-equidistant grey model and GRNN (Generalized Regression Neural Network) as static control model. The endpoint carbon content and the endpoint temperature respectively be taken as interval sequence to set up the non-equidistant grey model, and with series of GRNN carries out nonlinear optimization combination for the grey model simulation predicted values. Then the rationing relationship of converter main controllable charging material and other information can be obtained to guide smelting operation later, in order to improve the endpoint control effectively. The combined model in this article has more advantages like high precision, less modeling data, easy to be realized etc. if the smelting environment is changed, we can use a few qualified historical data to update work timely.We use the neural network predictive control theory on the basis of combined static control model to realize the optimal control of converter steelmaking endpoint. The neural network predictive control is a kind of control mode which can predict the future behavior of the system through the neural network identification model, and choose input which can make the system performance reach optimization stage by the optimization algorithms at the same time. BP neural network as a identification model of converter steelmaking, which has the most mature theory, and is generally used at present, but alse has disadvantages like easily to fall into local optimum and a slow convergence. Here we not only use the PSO(particle swarm algorithm) which is improved by the compression factor and genetic variation to optimize the structure of BP network, and the improved PSO is also used as the optimization algorithm of neural network predictive controller. We use grey correlation degree function to get a algorithm fitness function which is enough to reflect the expect carbon content and temperature and prediction carbon content and temperature relationship to realize the optimization of endpoint carbon contend and temperature control throgh optimizing the main controlled factors-blow oxygen and steel scrap. The experimental result from actual product data indicate that the combined static control model combined with neural network predictive control can realize the control of the carbon content and temperature effectively.