Research And Application Of Fixed Size Least Squares Support Vector Machine In Temperature Prediction Of Molten Steel

The accurate forecasting of molten steel in LF is an important precondition to organizing production, improving steel quality, reducing cost and controlling temperature of molten steel. But in ladle furnace refining process, it is difficult to establish an accurate mechanism model because of the nonlinear system. Therefore, a method based on the data is considered for model building. When training kinds of sample data sets, the fixed size least squares support vector machine (FS-LSSVM) method effectively avoids the sparse problem in least squares support vector machine(LS-SVM) method, at the same time, the training speed is quick and the precision is good. In this paper, a model based on fixed size least squares support vector machine for molten steel temperature prediction is proposed.Based on the analysis of the support vector machine (S VM) and the least squares support vector machine (LS-SVM), the principle of the FS-LSSVM is elaborated in detail to solve the sparse problem in LS-SVM method. Due to the problem of complexity of calculating the nuclear matrix eigenvalues, the Nystrom method is used to estimate the feature space mapping; and, the quadratic Renyi entropy criterion realizes optimal selection of the data in working set. The size has a direct effect on modeling time and precision through the analysis of the FS-LSSVM. At present there is no one method can effectively determine the size. In response to this issue, this paper proposes a method based on subtractive clustering to determine the size. The class based on subtractive clustering does not contain unbalanced data points, but can include all potential support vectors. The clustering number determined by subtractive clustering is representative and can be used as the size of FS-LSSVM. Therefore, the problem of the size determination is effectively solved.Through the analysis of energy balance and the mechanism model of ladle furnace, the mathematical model based on FS-LSSVM for molten steel temperature prediction is established. The inputs of the model are determined to be:power consumption, initial temperature, smelting time, addition of alloy and slag, steel weight. Then, a comparative analysis is done between the result of the LS-SVM model of and the result of the FS-LSSVM model. The analysis indicates the validity of the application of FS-LSSVM for molten steel temperature prediction. Finally, the results for the situations of different size illustrate the validity and feasibility of the subtractive clustering method.