Multi-factor Time Series Prediction Research Based On SVM And Its Parallelization

As the new algorithm of the machine learning method, Support Vector Machines is widely used for pattern recognition, classification and prediction. In recent years, wide range applications and improved technology of support vector machine development give a strong impetus to the development of time series prediction reasearch.The paper systematically introduce time-series forecasting methods and support vector machine regression theory of knowledge, and propose the use of relevant multi-factors of the support vector machine regression prediction model. In the analysis of factors relationships, we introduce the Bayesian Networks model in order to get the dependencies between the any two factors, and then rely on the basis of relations between the factors to select the factors sets of time series prediction, and finally build multi-factors support vector machine regression prediction model by the sets of the selected factors.In the multi-factors selection process, the discrete Bayesian algorithm is improved to make the Bayesian network model more efficiency and accuracy. In the process of determining the time series prediction dimensions, we use the phase-space reconstruction techniques. During the parameters optimization in the establishment of model, we improved method called parallel-cross-validation (PCV) is proposed in order to make the support vector machine parameters optimization more capacity and faster and more efficient.Parallel technology with MPI algorithm is used to improve the training model. We use the parallel training model of support vector machine with single program with multiple data (SPMD) method, and use parallel-cross-validation method in parameters optimization.Meteorological data is used as the experimental data. Through a group compariment experiments, the method has higher accuracy compared with the single factor method, and it has better generalization ability and higher speed. This method provides effective applications for the complexity time series prediction, and it has good prospects for non-stationary time series studies.