Periodicity Impacts on the Accuracy in GreyModel Based Internet Tra±c Prediction
-
Graphical Abstract
-
Abstract
The Grey model (GM) is used to predict
Internet tra±c. The Mean relative error (MRE) of pre-
diction varies regularly when one of the parameters for
GM(1,1), the Modeling length, increases. Moreover, the
prediction error becomes unacceptable in some scenarios.
The reason lies in such facts: (1) The Internet backbone
tra±c exhibits multi-scale properties in temporal domain,
which results in periodical variation of the tra±c sequence;
(2) GM (1, 1) requires that the accumulated generating
sequence of the data should be the form of exponential
function. However, the periodicity of the tra±c sequence
violates the condition. In order to keep MRE acceptable,
the Modeling length should be far shorter than the Pe-
riod length. What's more, the accuracy of four models,
ARIMA, ENN, GM (1, 1) and Residual GM (1, 1), was
compared and we found that the Residual GM (1, 1) con-
tributes little to the prediction accuracy while it doubles
the computational complexity.
-
-
DownLoad: