面向流量预测的时间相关图卷积网络构建方法

Construction method of temporal correlation graph convolution network for traffic prediction

现有数据中心虚拟网络中流量预测方法难以表征链路之间相关性,导致数据中心网络流量预测精度难以提升。基于此,提出了一种时间相关图卷积神经网络(TC-GCN),使能数据中心网络链路流量的时间和空间相关性表征,提升了流量预测精度。首先,构建具有时间属性的图卷积神经网络邻接矩阵,解决虚拟网络链路间流量异步性导致的预测偏差问题,实现了链路相关性的精准表征;其次,设计基于长/短窗口图卷积神经网络加权的流量预测机制,利用有限长度长/短窗口适配流量序列的平滑段与波动段,有效避免了神经网络梯度消失问题,提升了虚拟网络的流量预测精度;最后,设计了一个误差加权单元对长短窗口图卷积神经网络的预测结果进行加权求和,该网络的输出即为链路流量的预测值。为保障结果的实用性,基于真实的数据中心网络数据对所提时间相关图卷积网络进行了仿真实验。实验结果表明,所提预测方法相比于传统的图卷积神经网络流量预测方法具有更高的预测精度。

The existing traffic prediction methods in the virtual network of data centers characterize the correlation between links with difficulty,which leads to the difficulty in improving the accuracy of traffic prediction.Based on this,this paper proposes a Temporal Correlation Graph Convolutional neural Network(TC-GCN),which enables the representation of Temporal and spatial Correlation of the data center Network link traffic and improves the accuracy of traffic prediction.First,the graph convolutional neural network adjacency matrix with the time attribute is constructed to solve the problem of prediction deviation caused by traffic asynchronism between virtual network links,and to achieve accurate representation of link correlation.Second,a traffic prediction mechanism based on long/short window graph convolutional neural network weighting is designed,which adapts the smooth and fluctuating segments of the traffic sequence with a finite length long/short window,effectively avoids the vanishing gradient problem of the neural network,and improves the traffic prediction accuracy of the virtual network.Finally,an error weighting unit is designed to sum the prediction results of the long/short window graph convolutional neural network.The output of the network is the predicted value of link traffic.In order to ensure the practicability of the results,the simulation experiments of the proposed temporal correlation graph convolutional network are carried out based on the real data center network data.Experimental results show that the proposed method has a higher prediction accuracy than the traditional graph convolutional neural network traffic prediction method.