Spatial-temporal graph attention networks for dynamic traffic prediction of SDN

For the problem of static space and time dependencies based on traffic prediction in SDN traffic engineering, this paper proposes a dynamic network traffic prediction method, Attention mechanism for GCNGRU model (AGCNGRU), which integrates graph convolutional neural networks (GCN) with gated recurrent units (GRU) and incorporates an attention mechanism. By leveraging GCN, it captures the spatial dependency of traffic between nodes in the network, while GRU captures the temporal dependency of traffic passing through various nodes. The time attention mechanism is designed to assign weights to each hidden state, adjusting the importance of traffic information at different time points. Simultaneously, a data-driven spatial attention mechanism dynamically and adaptively adjusts the Laplace matrix, enabling dynamic extraction of spatial-temporal correlation in traffic data. This ultimately leads to accurate prediction of dynamic traffic. Experimental results on the GEANT datasets demonstrate that the proposed method significantly outperforms other approaches.