GA-BP-based Optimized Control Model For End Temperature And Carbon Content In Molten Metal Tapping From LD Converter

It is essential to predict accurately the end temperature and carbon content in molten metal in LD-process and optimize the main technological factors of LD-process for productive planning, improving steel quality and cutting down cost. Artificial neural network BP algorithm and genetic algorithm GA are introduced to LD converter. A new fitness function based on niche is proposed in order to improve the searching of GA and to prevent from premature convergence. GA and BP algorithm are also integrated to put up a new hybrid genetic algorithm for the optimization of training process of BP neural network. The hybrid algorithm based on GA and BP algorithm has both rapid local searching ability derived from BP and better global searching ability derived from GA. Investigations have been carried out for the optimization of multi-peak complicated function with several evaluating criterion such as convergence probability, average convergence time, average convergence value and so on. The results show that the new fitness function based on niche can improve the searching with genetic algorithm and prevent GA from premature convergence in some degree. The hybrid algorithm based on GA and BP algorithm with rapid local researching ability and global researching ability is more efficient and reliable than GA or BP algorithm on training artificial neural networks. A model for predicting the end temperature and carbon content in molten metal in LD-process is established based on the hybrid algorithm integrated GA with BP. The model has been tested by using three kinds of steel at errors of end carbon content ±0.02% and end temperature ±15℃, respectively. It has been shown that, for the continuous casting steel made in a manner of leaving high carbon content and post blowing in the situation of steelmaking process in the Vanadium-extracting and Steelmaking Plant of PISCO, the accuracy of prediction for end temperature, carbon content and the both are respectively 96%, 94% and 92%; for Stb32 steel made in a manner of leaving low carbon content and to adjusting carbon after the LD-process, the accuracy of prediction for end temperature, carbon content and the both are respectively 94%, 95% and 91%, and for PD3 steel made in a manner of leaving low carbon content and adjust carbon after the LD-process the accuracy of prediction of terminal tapping temperature, carbon content and the both are respectively 79%,74% and 71%.At last, a model for optimizing LD-process control with multi-objective genetic algorithm is established based on the predicting model. When given charging quantity of hot metal, temperature, [C], [Si], [P] in hot metal, the model can lead to the optimized value of process parameters such as the quantity needed of limestone, fluorite, scrap and the time of blowing oxygen and so on, in LD-process.