点击次数：

论文类型：期刊论文

发表时间：2010-01-01

发表刊物：Journal of Information and Computational Science

收录刊物：EI、Scopus

卷号：7

期号：1

页面范围：143-152

ISSN号：15487741

摘要：For purpose of meeting accurate, large-scale and real-time prediction requirements, a GLB-SVR approach is presented. As the theoretical advantage of applying support vector regression (SVR) to prediction highly depends on good parameter selection, simple yet practical formula methods to select the parameters for SVR directly from training sets are discussed and determined. Furthermore, despite some promising results for many SVR application studies, the difficulties in their design and implementation remain unresolved, for which reason a greedy load balancing algorithm (GLB) and a practicable real-time prediction method based on SVR are presented in this work. Experiments on the GLB-SVR-combing presented prediction method with GLB-applying to large-scale and real-time traffic flow prediction with real urban vehicular traffic flow data of Dalian city demonstrates that it outperforms not only the generalized neural network (GNN) and P-GNN (parallel GNN) in satisfying the accurate and real-time demands but also the naive parallel SVR (P-SVR) in meeting the real-time demand of prediction, and can close the gap between research and practical application of the prediction method. Copyright ? 2010 Binary Information Press.