Deep Learning Models For Aggregated Network Traffic Prediction

The ability to generate network traffic predictions at short time scales is crucial for many network management tasks such as traffic engineering, anomaly detection, and traffic matrix estimation. However, building models that are able to predict the traffic from modern networks at short time scales is not a trivial task due to the diversity of the network traffic sources. In this paper, we present a framework for network-wide link-level traffic prediction using Long Short-Term Memory (LSTM) neural networks. Our proposed framework leverages link statistics that can be easily collected either by the controller of a Software Defined Network (SDN), or by SNMP measurements in a legacy network, in order to predict future link throughputs. We implement several variations of LSTMs and compare their performance with traditional baseline models. Our evaluation study using real network traces from a Tier-1 ISP illustrates that LSTMs can predict link throughputs with very high accuracy outperforming the baselines for various traffic aggregation levels and time scales.