SAVE‐AS: Accelerating convergence in network traffic prediction through adaptive optimized variational mode decomposition and an integrated extreme learning machine

Abstract

Nonlinearity and nonstationary data affect the prediction accuracy of network traffic. A practical solution is to use an integrated modeling method based on time-series decomposition and an extreme learning machine (ELM). The original network traffic data series are decomposed in this paper using variational mode decomposition (VMD), and then each subdata series is subjected to phase space reconstruction (PSR). Finally, ELM trains a model for predicting network traffic. We used scalable artificial bee colony (SABC) algorithms to optimize parameter selection throughout the modeling process. However, several processes, such as VMD decomposition, PSR, and ELM training, will be repeated during iterative model optimization, resulting in a significant increase in computational complexity. This article proposes an adaptive selection mechanism for decomposing time series, which named is SABC-VMD-ELM prediction method with an adaptive selection (SAVE-AS) method. From gray correlation degree and approximate entropy, an adaptive selection operator is constructed that can adaptively select the number of data sets after VMD optimization decomposition. The ELM model is trained using several subdata sequences that satisfy the conditions rather than all subdata sequences to minimize computational complexity in modeling. Two publicly available data sets of Mackey-Glass and Lorenz chaotic time series were used; in addition, actual network traffic data from the MAWI Working Group's WIDE Backbone was simulated. It is demonstrated that the prediction model can reduce computational complexity and accelerate convergence speed while maintaining the model's predictive accuracy. In three data sets, training time was reduced by 25.25%, 23.87%, and 41.36%, respectively.