Graduate Studies
Embargoed Master's Theses
Traffic Prediction for Research and Education Networks: Anomaly-aware Deep Learning and Benchmarking
First Advisor
Byrav Ramamurthy
Committee Members
Lisong Xu, Sasitharan Balasubramaniam
Date of this Version
8-2025
Document Type
Thesis
Citation
A thesis presented to the faculty of the Graduate College at the University of Nebraska in partial fulfillment of requirements for the degree of Master of Science
Major: Computer Science
Under the supervision of Professor Byrav Ramamurthy
Lincoln, Nebraska, August 2025
Abstract
Research and Education Networks (RENs) and High-Performance Computing (HPC) environments are critical infrastructures for modern scientific discovery, demanding sustained high-throughput and low-latency data transfers. Unlike commercial networks, RENs exhibit unique traffic characteristics, including predominant “elephant flows,” inherent burstiness, and complex temporal-spatial dynamics often decoupled from human-driven cycles. Traditional traffic forecasting methods, tailored for commercial Wide Area Networks (WANs), consistently fail to capture these distinct REN dynamics, leading to inefficient resource management and potential impediments to scientific progress.
This thesis addresses this critical gap by developing and validating a robust, scalable, and anomaly-aware traffic forecasting framework specifically tailored for REN/HPC networks. Our contributions are threefold: (1) We developed and analyzed a two-month, multi-billion packet traffic corpus from ten Internet2 core routers, overcoming data scarcity challenges and providing a unique foundation for large-scale empirical analysis. (2) We designed and empirically validated a novel hybrid GRU-LSTM model that effectively captures both long-term dependencies and short-term fluctuations in REN traffic, demonstrating the critical impact of prediction lead time on operational utility. (3) We integrated Isolation Forest anomaly detection into forecasting models, significantly enhancing robustness against unexpected traffic surges, and conducted comprehensive benchmarking of various state-of-the-art deep learning models (N-BEATS, TiDE, PatchTST) to assess their performance with anomaly awareness.
Our findings demonstrate that tailored hybrid deep learning models, augmented with anomaly detection and optimized for lead time, achieve superior forecasting accuracy and robustness in REN environments. This work provides valuable tools for proactive resource management, congestion prevention, and optimized network operations, thereby accelerating scientific discovery and ensuring the efficient utilization of critical digital infrastructures.
Advisor: Byrav Ramamurthy
Comments
Copyright 2025, Mohammad Arafath Uddin Shariff. Used by permission