Research Of Time Series Prediction Method And Application Based On LS-SVM

Time series prediction has attracted great attention of scentists in time series analysis area. Due to the increasing demand of prediction efficiency in many application fields, fast prediction method is to be developed. As an improved method of Support Vector Machine (SVM), Least Square Support Vector Machine (LS-SVM) has many advantages, such as high training efficiency, good learning ability and simple modelling, exhibiting a good application prospect in fast time series prediction area. The present study mainly focused on the development and application of time series prediction based on LS-SVM.Firstly, the methods of time series forecasting and the basic principle of LS-SVM were introduced. Due to the difficulty in parameter selection, the parameter was optimized using multi-grid searching method and genetic algorithm method. After that, the input vector construction method of LS-SVM was investigated. Taking time series with certain periodicity as research target, we proposed a method of constructing input vector based on power spectrum analysis, which can achieve more accurate single-step prediction. In order to prevent the declining precision in the multi-step prediction with continuous sample points, an input vector construction method based on correlation analysis was used in the present study which maximized the preservation of historic information, reduced the sample dimension, and achieved multi-step forecasting with high precision. At last, we appled the above methods in mobile communication traffic prediction field and achieved real-time, online and multi-step traffic forecasting. Due to the particularity of holiday communication traffic, a combined forecasting method was employed based on LS-SVM and historical value similarity. The forecasting system software developed in the present study was applied in the network management system of China Mobile Communications Corporation (CMCC) Heilongjiang Co.Ltd, which is valuable for improving communication quality.A large number of experiment results indicated that LS-SVM can improve training and predction efficiency with higher precision. Compared with other methods, for the time sequence with certain periodicity, the input vector construction method proposed in this study can achieve accurate single-step or multi-step prediction. The stable operation of the forecasting system proved that it was an effective way to use LS-SVM in mobile communication traffic forecasting .