A Traffic Forecasting Method with Function to Control Residual Error Distribution for IP Access Networks

Takeshi KITAHARA  Hiroki FURUYA  Hajime NAKAMURA  

Publication
IEICE TRANSACTIONS on Communications   Vol.E93-B   No.1   pp.47-55
Publication Date: 2010/01/01
Online ISSN: 1745-1345
DOI: 10.1587/transcom.E93.B.47
Print ISSN: 0916-8516
Type of Manuscript: PAPER
Category: Internet
Keyword: 
traffic forecasting,  access network,  capacity planning,  neural network,  

Full Text: PDF>>
Buy this Article




Summary: 
Since traffic in IP access networks is less aggregated than in backbone networks, its variance could be significant and its distribution may be long-tailed rather than Gaussian in nature. Such characteristics make it difficult to forecast traffic volume in IP access networks for appropriate capacity planning. This paper proposes a traffic forecasting method that includes a function to control residual error distribution in IP access networks. The objective of the proposed method is to grasp the statistical characteristics of peak traffic variations, while conventional methods focus on average rather than peak values. In the proposed method, a neural network model is built recursively while weighting residual errors around the peaks. This enables network operators to control the trade-off between underestimation and overestimation errors according to their planning policy. Evaluation with a total of 136 daily traffic volume data sequences measured in actual IP access networks demonstrates the performance of the proposed method.