Research On Network Traffic Prediction Based On Spatio-Temporal Characteristics

In recent years,with the rapid development and popularization of smart devices,the traffic data accessing the network has surged,and the rapidly increasing user traffic demand has brought great challenges to the current network infrastructure.In this case,predicting network traffic and understanding traffic trends in advance can help network operators improve communication network management,network performance and resource utilization.Traditional linear models predict network traffic by modeling temporal features,but it is difficult to describe the randomness and nonlinearity of network traffic,and ignores spatial correlation,so the prediction results have great limitations.With the development of deep learning algorithms,more and more prediction models are applied to network traffic prediction problems.However,the prediction accuracy of network traffic is not high enough due to the complex dynamic spatiotemporal correlation of traffic and the easy accumulation of errors in long-term prediction.Therefore,how to use the spatiotemporal characteristics of network traffic to perform single-step and multi-step accurate prediction is the focus of this paper.For single-step prediction of network traffic,the existing prediction models ignore the spatial correlation between different regions and only consider the temporal features of traffic sequences;or use convolutional neural networks to extract spatial features,but require traffic data to be divided into grids,can only be applied in the Euclidean distance space scene.In order to fully extract the temporal and spatial features of network traffic and enable the model to be widely used,this paper proposes a multi-attention spatiotemporal network model MASTN(Mutil-attention Spatial-Temporal Network)for single-step prediction of network traffic.The model contains a graph convolution module and an autoencoder module.A graph convolution module is used to extract spatial features of network traffic.The autoencoder module is used to extract the temporal features of network traffic,and its encoder and decoder contain multiple spatiotemporal attention modules to model the impact of spatiotemporal factors on the state of network traffic.Experiments on real mobile network traffic datasets show that MASTN can effectively reduce the error of single-step prediction of network traffic.At the same time,since the model does not require the flow data to be divided into grids,MASTN can be applied more widely.Different from single-step forecasting,multi-step forecasting performs autoregressive iterative updates based on the single-step forecast value,which can predict long-term traffic.However,as the prediction step size increases,problems such as accumulation of prediction errors,disappearance of model gradients,and explosions will occur,resulting in nonconvergence of predictions.In addition,it is difficult for existing models to memorize information for a longer period of time,and the multi-step prediction accuracy is poor.Based on this,this paper proposes a multi-step prediction model of network traffic based on spatiotemporal convolution ASTCN(Attention Based Spatial-Temporal Convolution Network).First,the model models the network traffic sequence on the time scales of hours,days and weeks to capture the temporal characteristics and global correlations at different scales in the traffic sequence;then based on graph convolutional network(GCN)and temporal convolutional network(TCN)jointly builds a spatiotemporal convolution module to capture the complex spatiotemporal characteristics of network traffic;the model also considers the impact of external factors(such as holidays,etc.)on changes in network traffic,and finally fuses the three time components and external factors,in order to achieve accurate multi-step prediction of network traffic.In addition,when training the ASTCN model,the training method of Scheduled Sampling is used to reduce the accumulated error.Experiments on real mobile network traffic datasets show that compared with the existing models,the ASTCN model has higher prediction accuracy in multi-step prediction,and with the increase of the number of prediction steps,the advantages of this model are more obvious.