Traffic Matrix Prediction in SDN based on Spatial-Temporal Residual Graph Convolutional Network

Traffic prediction is the basis for dynamic network services and resource optimization. Software Defined Networking (SDN) provides a global view for network measurement and management, and a reliable way for network information collection and network traffic prediction. Recently, SDN traffic prediction model enabled by deep learning mainly focuses on mining the temporal correlation of traffic data, while the modeling of its spatial correlation remains to be explored. From the perspective of the problem, a novel Spatial-Temporal Residual Graph Convolutional Network (STRGCN) model for SDN Traffic Matrix (TM) prediction is proposed. Combined with SDN network topology, TM prediction is defined as time series data prediction in non-Euclidean space. Temporal Gated-Conv Blocks are based on Gated Convolutional Layer, and Spatial Graph-Conv Blocks are based on Graph Convolutional Network (GCN). Through identity mapping between Spatial-Temporal Conv Blocks, the deep spatial and temporal features extraction is implemented. Tested by TM data from CERNET and GÉANT datasets, the accuracy and the timely performance of our model are evaluated. Experimental results suggest that the general performance of STRGCN for TM prediction in SDN is superior to the current deep learning models.