Network Traffic Prediction Method Based on Multi-Channel Spatial-Temporal Graph Convolutional Networks

Recent years, the use of network traffic is increasing year by year, so network traffic forecasting has ushered in new challenges. Network traffic forecasting enables network operators to adjust usage demands in real-time, thereby rationally allocating network resources. Based on the characteristics of high nonlinearity and dynamic spatial-temporal correlation of network traffic data, a graph neural network and temporal extracting modules need to be used to extract temporal and spatial features separately. Existing works usually have some common problems: (1) the use of single-channel input data makes the data feature extraction relatively single. (2) construct graphs only through spatial relationships, lacking consideration of non-spatial relationships. Therefore, we propose a new graph convolutional neural network approach: Multi-Channel Spatial-Temporal Graph Convolutional Networks. Specifically, we do time slicing in the data pre-processing stage, to slice the data with different time dimensions to construct the adjacency matrix. To construct the graph from different angles, functional adjacency matrix and temporal adjacency matrix are introduced. The model demonstrates good performance on the real-world network traffic dataset Telecom Italia.