The software-defined network (SDN) is a new network architecture system that achieves the separation of the data plane and the control plane, making SDN networks more relevant to research. Real-time accurate network traffic prediction plays a crucial role in SDN networks, and the spatio-temporal correlation and autocorrelation of SDN make traditional methods unable to meet the requirements of the prediction tasks. In this article, a SDN network traffic prediction model DI-GCN (deep information-GCN) is proposed, which firstly fuses graph convolution with gated convolutional units; secondly, the matrix of mutual information relation is defined and constructed to obtain the relational weight representation of traffic data. The proposed model was compared with GCN, GRU, and T-GCN on the real dataset GÉANT, respectively. Experiments show that the DI-GCN model not only ensures the ability to represent the actual data but also reduces the prediction error as well as achieved better prediction results.