Forecasting of traffic variations from their preceding dynamics: Parametric vs non-parametric approaches

Abstract

Internet traffic intensity variations contain significant information on the access pattern dynamics. On the one hand, variability in access patterns is a direct manifestation of the end users' and IoT devices behavior. On the other hand, a better understanding of the access pattern dynamics provides essential information for an early redistribution of traffic, leading to potentially more efficient dynamic routing algorithms. Traffic in large networks is typically governed by a complex interplay of auto-and cross-correlation patterns that largely determine its non-stationary nature. Here we have considered two approaches to the identification of the traffic variation model. The first approach is parametric and focuses on fitting the parameters of Seasonal Auto Regressive Integrated Moving Average with exogenous factors (SARIMAX). The second approach is based on training of a recurrent neural network (RNN). Both approaches have been validated explicitly using traffic data records over several days of monitoring at the uplink of a local campus network.