Long-term prediction is one of the most difficult problems in the area of VBR video traffic prediction.As to the time variation

non-linearity and long range dependence in VBR video traffic trace

a novel method of feature based on multi-scale decomposition was proposed.On the analysis of the time-frequency distribution characteristics of the video trace

the wavelet packets which have the trait of arbitrary distinction and decomposition are selected.After space partition of wavelet packets

the best wavelet packet basis for feature extraction is picked out.Based on the best basis

it can do fast arbitrary multi-scale WPT（wavelet packet transform）

and obtain each higher dimension wavelet coefficients matrix.And then wavelet coefficients prediction is proposed based on LS-SVM and LMS algorithms.The long-term pre-diction of VBR video traffic is obtained through reverse wavelet transforms on the predicted wavelet coefficients.Nu-merical and simulation results are provided to validate the claims.