Network Traffic Prediction Based on Spatio-Temporal Graph Convolutional Network

Abstract

Network traffic prediction has a good application prospect in network operation and maintenance. It is to predict the characteristics of future network traffic from the past traffic. In order to solve the problems of insufficient utilization of temporal features and spatial features in traditional network traffic prediction, a network traffic forecasting method based on temporal representation and spatial convolution is proposed. Concretely, we first utilize graph convolution network to explore the topological characteristics of network nodes and then employ LSTM to characterize the temporal features of the networks. Moreover, a representation vector of time is generated based on timestamp data to help the network to learn daily and weekly long-term temporal features. In this paper, the public dataset Abilene is used for testing. The results show on different scales show that the MSE, RMSE and MAPE value of our method reduced by 22.2%, 30.0% and 18.5%, respectively, compared with the original LSTM.