Traffic Matrix Prediction for Optical Networks

Abstract

The rapidly increasing data demand from current Internet services, such as, cloud computing and high-quality video streaming is raising the pressure on network operators to provide reliable high-speed connections while keeping costs low. Without having to rely on the bandwidth expansion of optical waveguides or modulation level efficiency, one of the ways to increase spectral efficiency of a network is by the optimal utilization of already existing resources. Since the resource management can be improved when the future traffic is known beforehand, in this work, we investigate the ability of recurrent neural networks (RNN) with long short-term memory (LSTM) to realize traffic matrix prediction based on previous traffic history of the computer network. Real traffic data has been used to train the RNN, from an anonymized dataset containing the traffic history of the Abilene and GEANT optical networks from the years 2004-2005 and test the viability of LSTM RNN models for proper traffic prediction. The proposed models achieved a mean square error of 0.0026 for the Abilene and 0.00058 for the GEANT network.