Internet Traffic Prediction Using Recurrent Neural Networks

Abstract

Network traffic prediction (NTP) represents an essential component in planning large-scale networks which are in general unpredictable and must adapt to unforeseen circumstances. In small to medium-size networks, the administrator can anticipate the fluctuations in traffic without the need of using forecasting tools, but in the scenario of large-scale networks where hundreds of new users can be added in a matter of weeks, more efficient forecasting tools are required to avoid congestion and over provisioning. Network and hardware resources are however limited; and hence resource allocation is critical for the NTP with scalable solutions. To this end, in this paper, we propose an efficient NTP by optimizing recurrent neural networks (RNNs) to analyse the traffic patterns that occur inside flow time series, and predict future samples based on the history of the traffic that was used for training. The predicted traffic with the proposed RNNs is compared with the real values that are stored in the database in terms of mean squared error, mean absolute error and categorical cross entropy. Furthermore, the real traffic samples for NTP training are compared with those from other techniques such as auto-regressive moving average (ARIMA) and AdaBoost regressor to validate the effectiveness of the proposed method. It is shown that the proposed RNN achieves a better performance than both the ARIMA and AdaBoost regressor when more samples are employed.