Research Into Short-term Forecasting Models Of Manufacturing System Product Damand Based On Support Vector Machine

The short-term prediction of manufacturing system product demand is actually a prediction of multi-dimensional nonlinear sales time series with small noised samples.The support vector machine(SVM)method has been widely used in forecasting nonlinear time series,and many achievements have been made.Therefore,to make up for the deficiencies of the basic SVM algorithm,a few new methods are proposed aiming at the three cases of multi-factor,noise and small sample based on the SVM method.First,an expanded RBF kernel support vector machine is proposed aiming at the characteristics of multi-dimension and nonlinearity existing in the product sale series through expending the RBF kernel function.Then,an iterative support vector machine and a SVM with adaptive segmented loss function are proposed aiming at data characteristics of noise existing in the product sale series.Finally,aiming at the small sample of the product sale series,a SVM based on the interval of samples is proposed.Meanwhile,an improved immune optimization algorithm is designed to optimize the parameters of these models.The experiment results indicate that the models proposed are effective and feasible.The main contents of this dissertation are introduced in detail as follows:(1)Aiming at the characteristics of multi-dimension and nonlinearity existing in the product sale series,an expanded RBF kernel function is designed and applied to the support vector machine to get an expanded RBF kernel support vector machine(ERBF-SVM)in this dissertation,and an improved immune optimization algorithm is designed to optimize the parameters of ERBF-SVM.This method proposed is applied to the automobile sales forecasting in contrast with the BP neural network(BPNN),the standard support vector machine using the RBF kernel(v-SVM)and multi-scale support vector machine(MS-SVM).The experiment results indicate that ERBF-SVM is effective and feasible,by which more accurate forecasting results are obtained over the three other methods.(2)Aiming at data characteristics of noise and nonlinearity existing in the product sale series,an iterative support vector machine(I?-SVM)is proposed in this dissertation.During the gradually reducing process of I?-SVM's parameter ?,the samples that may include big noises are iteratively amended to reduce their influence on the final forecasting model generated.And it is proved that the training support vector machine with the updated sample set can obtain larger sample interval than the original training set in this dissertation.Finally,I?-SVM is applied to a numerical value example and the automobile sales forecasting in contrast with the ? support vector machine(?-SVM).The experiment results indicate that I?-SVM is effective and feasible,by which more accurate forecasting results are obtained over the ?-SVM.(3)A support vector machine based on adaptive segmented loss function,by name AS?-SVM,is proposed aiming at data characteristics of noise and nonlinearity existing in the product sale series.In AS?-SVM,a separate insensitive loss value is assigned to each sample point adaptively,especially the larger insensitive loss values to the samples with big noises to reduce the influence of inaccurate samples on the final model,and it is proved that it can enhance partial generalization performance of the model.Finally,AS?-SVM is applied to a numerical value example and the automobile sales forecasting in contrast with the ? support vector machine(?-SVM).The experiment results indicate that AS?-SVM is effective and feasible,by which more accurate forecasting results are obtained over the ?-SVM.(4)A support vector machine based on interval of samples,by name IoS-SVM,is proposed aiming at data characteristics of small sample and nonlinearity existing in the product sale series.In the case of small samples,there is usually insufficient information to train the support vector machine model.By introducing the interval between two samples,more information between samples can be obtained through training support vector machines.Then,IoS-SVM can be obtained.by deriving its dual optimization problem.And it is proved that the solution of IoS-SVM has similar properties to that of SVM.Finally,IoS-SVM is applied to a numerical value example and the automobile sales forecasting in contrast with the ? support vector machine(?-SVM).The experiment results indicate that IoS-SVM is effective and feasible,it can obtain more accurate forecasting results over the ?-SVM in most cases,it shows that IoS –SVM is more robust.