Network flow prediction by AC-GRU based on improved attention mechanism

Network flow prediction is the key to urban network planning as well as network load balancing. However, due to the highly nonlinear nature of network flow, some traditional methods are difficult to extract its complex spatio-temporal features, resulting in low prediction accuracy. To address this problem, this paper proposes a GRU model AC-GRU based on improved attention mechanism for network flow prediction. AC-GRU extracts the spatio-temporal features of the flow using improved attention mechanism GRU, convolutional GRU, respectively, and then adaptively fuses the features using bilateral gating mechanism. In this paper, experiments are conducted on three public network flow datasets GEANT, CERNET and ABILENE, and RMSE and ACC are used as evaluation metrics. The experimental results show that the AC-GRU method can achieve better prediction results.