Towards traffic matrix prediction with LSTM recurrent neural networks

Abstract

This Letter investigates traffic matrix (TM) prediction that is widely used in various network management tasks. To fastly and accurately attain timely TM estimation in large-scale networks, the authors propose a deep architecture based on LSTM recurrent neural networks (RNNs) to model the spatio-temporal features of network traffic and then propose a novel TM prediction approach based on deep LSTM RNNs and a linear regression model. By training and validating it on real-world data from Abilene network, the authors show that the proposed TM prediction approach can achieve state-of-the-art TM prediction performance.