Theory And Application Research On LS-SVM In Time Series Prediction

Time series prediction method has been applied to almost all forecasting anddecision-making fields, and is widely used in practice. The study of this method notonly has important theoretical research significance, which is a research hotspot anddifficult problem for the scholars both at home and abroad. In support vector machinemodel, structural risk minimization, kernel function mapping and convex quadraticprogramming techniques are successfully applied to solve the curse of dimensionalityand local minimum problems during traditional machine learning. Least squares supportvector machine (LS-SVM) is an improved model of support vector machine (SVM).The prediction accuracy of LS-SVM doesn’t weaken, and LS-SVM has the advantagethat its operation is more simplified than SVM. This paper has studied some problemsabout LS-SVM, and it mainly concentrates on the following contents:①A forecasting method which combines empirical mode decomposition and theLS-SVM is proposed. Combined with the practical application of the prediction ofbuilding energy consumption, the main idea of the method is that the EMD method isbroken down energy consumption data into several intrinsic mode component, and thenLS-SVM models are established for each eigenmode separate forecast, the sum of theprediction of all eigenmodes component corresponding LS-SVM model is finalpredicted results. Better predict building energy consumption prediction experiments,the method of non-stationary time series prediction accuracy is superior to thetraditional LS-SVM、SVM and BP neural network.②In order to solve the parameter optimization problem of LS-SVM model intime series prediction, this paper puts forward a new method of parameter selection. Themain idea is that the training data is divided into two groups, respectively, as thetraining data and test data in the process of selection, and culture of immune geneticalgorithm which combines global optimization using immune clone selection algorithmand local optimization based on Baldwin learning mechanism which is adopted toenhance the excellent individual fitness. This way speeds up the convergence speedduring parameter optimization. This method is tested by Lorenz chaotic time sequence,and the experimental results prove the feasibility and superiority of this optimal method.③In some high real-time demand forecast of application, the traditional off-lineprediction model can’t meet the requirements, but the online prediction model better meets the requirements. Then this paper proposes an improved LS-SVM onlineprediction method which is based on incremental learning and pruning algorithm. Whennew sample arrives, traditional online prediction method adopts direct matrix inversionmethod to update support vectors, but this way increases the computation complexityand can’t meet the requirements of real-time prediction. In order to avoid the directinverse and greatly reduce the time needed for prediction, the incremental learningmethod is used to recursively update support vectors. For the sake of reducing thenumber of incremental learning, this paper uses incremental selected learning waywhich based on the prediction error of the new coming sample before fast pruningstrategy prunes the support vector which first joins the support vector machines.Because this method not only ensures the prediction accuracy, but also has quickprediction speed, this method better meets the requirements of practical application.