Network traffic prediction is one of the effective way to improve user QoS. The network topology information is not fully utilized in current network algorithm prediction. A network traffic detection model based on high order graph convolutional network algorithm is proposed，and further predicts network congestion based on traffic information. The traffic prediction model utilizes the graph convolutional to capture the mix-hop effect of traffic. And the gated recurrent unit （GRU） obtains the time correlation information of the traffic in the network. The autoencoder model implements the unsupervised learning and traffic prediction. The simulation experiment is on the real data of the network Abilene. The experimental results show that the mean absolute percentage error（MAPE） value of the method in network traffic detection is 41.56%，which is lower 1.64% than DCRNN methods，at the same time，the prediction accuracy is also optimal.