Research On On-line Time Series Prediction Based On Online SVR

Recently, on-line time series analysis and prediction is becoming a research hotspot with remarkable academic and practical values, as well, is being attached more and more importance by scholars in China and abroad. Time series prediction methods based on Support Vector Regression(SVR) have become the main stream in the research field of time series analysis and is being popularized in various application fields gradually for their solid statistical theory basis and excellent application effect.As an on-line modeling method, Online SVR can realize dynamic on-line modeling. However, the high complexity of Online SVR insufficiently fulfills the efficiency requirement of many real on-line applications. Accordingly, considering the influence of kernel function types and sample scales on the Online SVR algorithm, this paper focuses on improving the precision and efficiency of on-line time series prediction with kernel combination and sample reduction. The main research work includes the following aspects:1. To guarantee the prediction precision, an Online SVR prediction approach with the combination of global and local kernel functions is proposed. The approach combines the excellent trend fitting characteristic of global kernel function and powerful neighborhood nonlinear approximation ability of local kernel functions. The experiments demonstrate that this approach can achieve higher prediction accuracy compared with the single kernel function methods.2. A local Online SVR algorithm based on the compensation of residual errors is proposed to resolve the low efficiency problem caused by the kernel function combined approach mentioned above. This algorithm first trains models with off-line SVR algorithm and then models residual error by Online SVR to correct the off-line SVR prediction outputs. The complexity is reduced through a combination of off-line and on-line methods. Simulation experiments show that this method can get more accurate and faster results compared with the single kernel function methods and kernel function combination Online SVR methods.3. The validity of support vector samples directly determines the efficiency of forecasting. An improved accelerated decremental Online SVR algorithm is proposed in this paper. The algorithm employs the accelerated selective forgetting strategy which leads to the decrement of the on-line data samples and the decrease of algorithm complexity. Experiments illustrate that the efficiency of the algorithm can be improved while maintaining the prediction accuracy.4. The decrement of on-line data samples weak the generalization ability and reduce the forecasting precision of Online SVR algorithm. A segmental Online SVR time series prediction algorithm is suggested to solve this problem. In this algorithm, the Online SVR models are trained and memorized as segmentation. Then the prediction result is output with the most suitable segmental Online SVR model. This algorithm decreases the scale of samples while remaining most of the historical knowledge to enhance the generalization ability of the model. Experiments show that the algorithm can realize effective prediction without decrease of precision.5. A multi-scale parallel forecasting algorithm is proposed for solving the prediction whose on-line time series contains multi-scale characteristic. Firstly, sub-series of different scale are gained and different parallel models of Online SVR are trained. Then the prediction result is output with the most appropriate model. Finally, the multi-scale time series reconstruction is achieved. With this method the size of on-line data can be effectively reduced. Experimental results indicate that the algorithm can effectively improve the on-line prediction accuracy while keeping efficiency.6. To evaluate the validity and practicability of the proposed algorithms, two schemes are designed to apply the algorithms to the on-line fault prognostic and mobile traffic forecasting. The experiments and tests confirm the algorithms can be effectively applied to the on-line time series prediction with excellent performance in both efficiency and precision.