An Experimental Evaluation of Deep Learning for Internet Traffic Prediction

Abstract:
Internet traffic prediction has become increasingly important in our modern society, with the large number of users and devices contributing to increasing traffic flows. Previous studies have explored both linear and machine learning models for these tasks. However, deep learning models have recently been proposed as new solutions. To gain a better traffic prediction performance, gated recurrent unit (GRU), long short-term memory (LSTM), temporal convolutional network (TCN) and time series Transformer (TST) are used, with six public datasets, namely, Abilene, brain-1h, brain-1min, GEANT, germany50, and nobel-germany. A comprehensive evaluation reveals that GRU achieves the best performance with minimal training time consumption.