Research On Network Traffic Prediction Based On Deep Learning Method

In recent years,the Internet of Things,big data and other fields have developed rapidly,and with the popularity of computers,mobile phones and other intelligent devices,the network traffic has increased dramatically.The stability of the network has an important impact on people's access to resources through the network.Therefore,network traffic prediction is an important prerequisite for improving network security and the user quality of service.The research of network traffic prediction has gradually become a research hotspot,and it is also one of the important research directions in the field of network research.At present,the prediction accuracy of the model based on deep learning needs to be improved.The model needs to make full use of the topology between real network nodes,and it also needs to fully extract information in terms of the internal traffic matrix data's the temporal characteristics.These two effects will lead to insufficient accuracy of spatio-temporal prediction results,and the overall universality and predictability of the model need to be improved.Therefore,starting from the spatial and temporal characteristics of traffic,this thesis further elevates the performance of network traffic's spatio-temporal prediction,so as to reduce the prediction error and improve the prediction accuracy of the model.This thesis proposes a network traffic prediction model based on GNNLSTM(Graph Neural Networks and Long Short-Term Memory).The purpose is to solve the spatio-temporal prediction of network traffic.The model is based on GNN(Graph Neural Network)to obtain the spatial topology of real network routing nodes,and LSTM(Long Short-Term Memory)neural network to obtain the time series correlation traffic matrix(TM)information of network traffic.Finally,the effective prediction of network traffic is realized.On the real network traffic data sets Abilene and G(?)ANT,the experimental results show that MSE error loss value and Mae error loss value of the model in network traffic prediction are lower than those of the traditional prediction model based on deep learning.The overall prediction effect is better than the current method,which can effectively improve the accuracy of traffic prediction and reduce the error value.In order to further improve the capture of spatio-temporal information and improve the accuracy of prediction,this thesis introduces the mechanism of spatiotemporal attention.Furthermore,the prediction model of GNN STAM-LSTM(Graph Neural Networks and Long Short-Term Memory of Spatio Temporal Attention)combined with Graph Neural network,Long short-term Memory and Spatio Temporal Attention mechanism is proposed.It can capture the spatiotemporal characteristics of network traffic spatial information and time-domain information.In the LSTM unit,the historical state is directly integrated into the current cell state.In this way,the memory of historical flows is enhanced and the network traffic is fully predicted.In the end,it is verified and analyzed on Abilene and G(?)ANT network traffic data sets with complex topology to achieve accurate and effective prediction.The evaluation index shows that the model proposed in this thesis can effectively improve the prediction effect.To sum up,for the study of network traffic prediction,this thesis starts from the characteristics of network traffic and conducts in-depth studies from the two parts of spatial topology and time series respectively.The proposed model has greatly improved the prediction effect and accuracy.