Research On Network Traffic Prediction Model Based On Graph Neural Network

With the acceleration of the smart city process and the development of the Internet of Everything,the concept of network is expressed in various application scenarios,different application scenarios having different forms of expression,such as SDN networks,traffic networks,etc.SDN networks and traffic networks differ in the construction of the expression of the network topology and in the diversity of network properties.These networks can be presented in the form of graph structured data,thus laying the foundation for research on network-related aspects.In recent years,network traffic forecasting has become an important research topic due to changes in network structure and increased service demands.The implementation of accurate network traffic prediction can greatly improve people’s daily lives,saving time costs and energy consumption,etc.It also helps to promote the improvement of intelligent network governance and the development of smart cities.The intricate spatial and temporal dependencies in the network become the key research object of the network traffic prediction problem.In order to achieve the fusion of spatio-temporal features extraction,this thesis carries out the research of network traffic prediction models based on graph neural network.In the software-defined network(SDN)scenario,to address the problem that artificial neural networks cannot fit network traffic data well,a deep mutual information graph convolution model DI-GCN(Deep Information-GCN)based on mutual information is proposed.The DI-GCN model could capture more detailed spatial structure feature information by defining and constructing a mutual information relationship matrix,thus solving the problem that network models are difficult to fit the network traffic data.The real dataset G(?)ANT was selected and the prediction performance was compared with the benchmark models respectively.The RMSE,MAE,MAPE and R~2 values of the proposed model in the 15min prediction were 4.7561,2.8963,6.628%and 0.9012 respectively,which were all better than other models.The experiments showed that the DI-GCN model not only ensures the ability to represent the actual data,but also reduces the prediction error,and thus achieves better prediction results.In the traffic network traffic prediction application scenario,the Dynamic Topology Man-GCN(DTM-GCN)based on dynamic graph convolutional network was proposed to address the problems of unsatisfactory adaptability of predefined topology graphs and ambiguity of graph structure representation.The problem of unsatisfactory adaptive capability was effectively solved by fusing the adaptive dynamic topology graph module and MK temporal prediction module to complete the processing of traffic network traffic self-similarity,spatial heterogeneity and other features.The proposed model was evaluated on two datasets,including Los Angeles and Pe MS07.The RMSE values were 4.9651 and 4.8861,the MAE values were 3.4906 and 3.2754,the MAPE values were 6.642%and 6.548%,and the R~2 values were 0.9034 and 0.8905respectively in the 15min prediction of two datasets.The experimental results showed that the DTM-GCN model was more widely applied and had a better ability to deal with network mutations.