Research On Network Traffic Prediction Technology Based On Deep Learning

As the network environment becomes more and more complex,there are more and more types of network devices,resulting in more and more complex network traffic trends.Traffic trends in modern networks are fluctuated by various factors in the real environment anytime and anywhere.Under the influence of increasing network traffic and increasingly complex network traffic trends,network load balancing and even network security are facing increasingly severe challenges.Network traffic forecasting is an important part of the network load system,which provides allocation basis for resource allocators;network traffic forecasting technology learns and generates forecasting models through historical traffic data,which can perceive and predict traffic trends in advance,making the network system There is more time to face possible network fluctuations.The rapid development of the network has made the characteristics of the network more and more complex.The modern network not only shows the characteristics of changing trends and large amount of data on the time scale,but also has an increasing spatial scale and spatial structure.complex.The majority of conventional models primarily focus on the temporal aspects of network traffic and overlook its spatial characteristics.This limitation hampers their ability to deliver accurate and stable traffic prediction outcomes.To address this,the paper introduces a spatiotemporal fusion prediction model that leverages the spatiotemporal traits of network traffic.To enhance prediction accuracy,the model adopts a two-step approach,the self-similarity and burstiness of network traffic are modeled and analyzed.Taking two public data sets in the field of network traffic forecasting,the R/S method is used to find Hurst parameters to quantitatively analyze the self-similar characteristics of network traffic.Next,the alpha-stable distribution function is employed to fit the probability density function of network traffic,enabling the assessment of its burstiness.Aiming at the self-similarity and burstiness of network traffic,a pure convolution multichannel graph convolution spatio-temporal fusion model(ST-MWGCN)is designed.The model consists of multiple spatio-temporal fusion blocks(ST-Fusion Module),and each spatio-temporal fusion block consists of 2 graph convolutional networks with multi-weight input and 3 GLU-based expansion convolution blocks.The multi-weight input based on the graph convolutional network can better capture the connectivity and near correlation between nodes;the extended convolution block based on GLU is used to process the temporal feature part of the input data.ST-MWGCN has conducted multiple experiments on the real Abilene and GEANT data sets,and compared with the traditional ARIMA model: in the short-term prediction task,the MAPE decreased by 8.41% and 8.838% respectively;in the long-term prediction task,the MAPE decreased 14.26%,11.472%.The results show that the model can extract the spatiotemporal features in network traffic,and has accurate and stable long-term prediction ability.