Traffic Prediction for Research and Education Networks using an Ensemble GRU-LSTM with Varying Lead Times

Research and Education Networks (RENs) are critical infrastructures that facilitate high-bandwidth data transfers for scientific collaborations and educational activities. While numerous traffic prediction techniques exist for public Wide Area Networks (WANs), they often fail to capture the unique characteristics of REN traffic. This paper presents a novel approach to traffic forecasting tailored for RENs, using real-world traffic traces from Internet2, a major REN in the US. We leveraged these traces to develop a high-performance forecasting model and provide key insights into REN traffic behavior. Our main contributions are as follows: (i) We performed an in-depth analysis of core router traffic from the Internet2 backbone, revealing distinct cyclical patterns that are highly amenable to prediction. Specifically, we examined traffic from two core routers, identifying unique temporal dynamics. (ii) We designed an innovative hybrid Recurrent Neural Network (RNN) model, combining Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) layers (GRU-LSTM), to effectively capture both long-term dependencies and short-term fluctuations inherent in REN traffic. (iii) We demonstrate that the prediction lead time is a critical hyperparameter for REN traffic forecasting, significantly impacting model accuracy. This contrasts with typical WAN traffic patterns and highlights the unique predictability of REN traffic due to consistent, experiment-driven flows. (iv) We propose to enhance our traffic prediction model by integrating traffic anomaly detection, thereby improving its robustness against unexpected traffic surges.