Deep Learning for Network Traffic Prediction

Network traffic prediction is an important tool for the South African Research and Education Network (SANREN) in managing network congestion, resources and security. It predicts future network traffic flows based on previous data, using either statistical time-series or machine learning approaches. Efficient prediction can improve the quality of service and lower operating costs for network service providers. Existing literature shows that deep learning models learn network traffic patterns more efficiently and predict future traffic more accurately than traditional prediction models. Three different Long Short Term Memory (LSTM) models were de- veloped for the SANREN: Bidirectional, Simple and Stacked. These models were evaluated both in terms of their predic- tion accuracy, and their computational complexity. Dataset size and prediction accuracy have a strong correlation, at the cost of increasing training time as more data are added. The results demonstrated that a Stacked LSTM was the most accurate prediction model, at the expense of using the most computational resources. Using a Simple LSTM greatly saved on computational resources, and was only slightly less accu- rate at predicting future traffic on the SANREN compared to a Stacked LSTM.