Research On RBFNN And Application In Building Iron Water Desulfuration Static Model

Sulfur is a harmful element in most kinds of steel, so it is necessary to remove it from the iron water before steel making. The present desulfuration technique is still based on manual control that depends on the experiences of workers, so it lowers the automation level of the factory and the stability of steel quality. Because most of the popular desulfuration control models abroad are based on desulfuration mechanism, which require the very stable production situation, it is difficult to transplant them to the different factories. Analyzing the techniques and factors of dusufuration, this dissertation constructs an iron water desulfuration static intelligent prediction model based on RBF neural network.Firstly, the dissertation carefully introduces the basic theory of RBF neural network, and then conducts a deep research in the network structure learning: the center numbers, spread and weight. The most important factor of RBF neural network is the center selection, which influences the performance of RBF greatly. A kind of rival penalized competitive learning(RPCL) technique is implemented to decide the centers of RBF, by which the center numbers are acquired automatically. In order to optimize the centers of RBF more effectively, this dissertation introduces the input-output clustering method and combines it with RPCL algorithm in clustering. The spread of RBF affects the generalization of the neural network directly. Large spread will lead inaccuracy, on the contrary, small spread will harm the generalization. Therefore, the chaos search algorithm is adopted to optimize the spread of RBF. To solve the problem of the determination of the regularization coefficient in the weight learning of regulariztion RBF, the dissertation provides EM algorithm to learn the weight without determining the value of regularization coefficient in advance.Finally, an iron water desulfuration static prediction model is constructed by the optimized RBF neural network. The input parameters include the weight of iron water, the quantity of sulfur before desulfuration and the quantity of sulfur after desulfuration. The output parameter is the weight of dusulfuration agent. The offline simulation is conducted and the results show the following conclusion: 1. The improved RPCL algorithm can not only determine the center numbers of RBF, but also optimize the location of these centers, which improved the accuracy of the model; 2.The chaos search algorithm can find the suitable spread, which guarantees the good generalizationof RBF neural network; 3. The iron water desulfuration static prediction model based on RBF neural network can be successfully applied to the dusulfuration to predict the weight of desulfuration agents in steel making process.