Research On Traffic Prediction Based On Generative Confrontation Network

The birth of the 5G technology and the arrival of the IoT era have caused a large number of mobile services and consequently more traffic data with multiple sources and heterogeneity has also emerged.Huge traffic data makes the burden of network management even heavier.A large number of traffic detectors are deployed in the urban network,and the difference in their detection capabilities is likely to cause that the detected traffic coverage area is too large,but the accuracy of spatial distribution is too low.In each dense cell,obtaining highly accurate traffic consumption is of great significance for network resource management and call admission control.Traditional forecasting methods require a lot of time to analyze the traffic characteristics.Due to the complex correlation characteristics exhibited by the traffic data in the city,the study on spatial domain traffic prediction has less work.In this study,we forecast the Spatial mobile traffic in a city-wide cellular architecture.The Bicubic interpolation and Generative Adversarial Networks are used to study and analyze the open source traffic data set.In order to solve the problem of city domain traffic forecasting,a RessNetGAN model based on generating anti-network is proposed.It uses the deeper network to extract more features and puts different rough data of the traffic data to be predicted into the research scope.The work done in this study is as follows:First,we preprocess the data set and map the traffic data to the two-dimensional matrix according to the spatial and temporal correspondence.Then used the Pearson correlation coefficient to analyze spatial correlation of traffic.Second,we use Bicubic interpolation and convolutional neural network(CNN)to predict spatially mobile traffic.We present the framework and structure of the model for traffic prediction.And conduct experimental analysis of the two models.It is found that the prediction performance of convolutional neural network is superior to the Bicubic Interpolation,and the prediction result is closer to the actual value.Then,For the high spatial complexity of traffic data,we first give the definition and framework of the RessNetGAN model.Secondly,we give the concrete structure of the generator,and introduce the principle of structural stability training of spatial eigenvalues and composite residuals of urban traffic through 3D convolution operation.Finally,we give the structure of the generator and introduce the training process and parameter update process of the modelFinally,we changed the pooling parameters of the traffic to be predicted for experimentation.The results show that the RessNetGAN model with GAN structure has a standard mean square error of 34.9% compared with the convolutional neural network model.The structural similarity has increased by 46.6%.It has better predictive performance.Even if the accuracy of the input flow data is low,the predicted result section is also authentic.In summary,based on the great complexity of urban traffic,the deep learning method,especially the proposed RessNetGAN model,uses deeper networks and generated adversarial structures to automatically extract complex traffic feature It reduces the enormous workload of analyzing traffic characteristics in traditional methods in order to get closer to reality and can get closer to the actual predicted value.It can help telecommunications operators to conduct network resource management and call admission control more reasonably according to the predicted results,and ensure the quality of its user services.