Research On Multi-model Modeling Method Based On Information Entropy

Modeling method of data driven is widely used to set up the model of complex industrial processes, but the characteristics of high dimension, redundancy and noisy impact the model’s property heavily. In this thesis, study on the problem of redundancy that impact the generalization of model deeply.In this thesis, the selection method is used to wipe out the redundant data. The strength of redundancy has a relationship with probability, when the redundancy is high the information entropy is small. So we can consider that when the number of samples is established the sample set redundancy lower when the information entropy is maximal. In this thesis, a swap way is used to select samples. The result indicates that this method can get the redundancy data simply and efficiently then get the lower redundancy sample data.In order to set an efficient model based on subset (selected sample data). Modeling use the Least Squares Support Vector Machines (LS-SVM) method was used. In the method of LS-SVM based on subset data, subset data as support vectors and solve the original LS-SVM problem. According to the simulation results, the model has a good property of generalization, but the property impact deeply by the number of samples.In this thesis, times of select is used to determine the best number of samples and then many different subsets generated and many models set up. By the analysis, the models have strong diversity and then fell the models together.In this thesis, a weighting sum structure is used to fuse the outputs together. The fuse method based on information entropy was proposed to mix the outputs of sub-models together. Finally, we get the weight values by means of solve a multi-objective optimization problem and set the multi-model up in the end.Finally, applied the multi-model modeling method based on information entropy in the example of electric arc furnace steelmaking and then build prediction model on end-point temperature in EAF. The simulation result indicates that this method is efficiently and the generalization is stronger than single model.