Research On Billet Temperature Modeling Method Of Reheating Furnace Based On SVM And RVM

As the research on billet temperature forecast model of reheating furnace becomes a hot topic, establishing a reasonable model is of great significance for improving the heating quality of billets and reducing energy consumption. Usually the prediction model can be established by analysing the parameters, such as heat radiation, heat transfer mechanism in the furnace and so on, but this method is difficultly achieved, because it is based on many assumed conditions and parameters, complex algorithm. By contrast, the information about furnace temperature, billet resident time and other related information can be sampled in the heating process. Then the forecast model is established by using these information, which is based on data-driven approach.In this paper, billet temperature forecast model is established by using data-driven approach, the main contents can be generalized as follows:1) The technology background of the walking beam furnace is analyzed. For the relevance and non-linear relationship between variables may exist, the methods of feature extraction are studied, mainly including PCA and KPCA, extracting features of the process data as the inputs of the trained model.2) The regression modeling method based on SVM is applied to predict billet temperature of reheating furnace, through analysing the three modeling methods of LSSVM, PCA-LSSVM and KPCA-LSSVM, the simulation results have proved that the model is more accurate after using KPCA for feature extraction on the input variables. However, many shortcomings are found by using this method:the parameters need more, the kernel function must satisfy Mercer’s condition, predictions are not probabilistic and so on.3) For the shortcomings in SVM, a modeling method based on RVM is tried. The simulation results show that this approach has higher prediction accuracy, sparser solutions and better model generalization ability, while the uncertainty of model forecast can be given. The prediction accuracy of the model combined with the feature extraction can be further mproved.