Research On Traffic Prediction Technology Considering Wireless Service Features

With the integration of communication technologies,big data,and artificial intelligence,the next-generation wireless network will be an intelligent network,which is self-managed,self-configured,and self-optimizedunder the dynamic enviroment,and provide users with intelligent services.High-precision traffic prediction can effectively reflect the real needs of users and help intelligent wireless networks to dynamically manage resources.However,traditional traffic prediction mostly use centralized prediction models,that is,the traffic data of each subnet is uploaded to the central server for unified modeling,which brings a serious problem of excessive communication overhead.As the computing power of wireless network edge nodes increases,people begin to use federated learning models for traffic prediction,that is,the edge server first uses local data for training,uploads model parameters to the central server for parameter fusion and distribution,and iteratively obtains a distributed prediction model.Although the problem of high communication overhead of the centralized prediction model is solved,the data features are not further analyzed,resulting in feature redundancy,which leads to excessive model parameters,and the prediction speed and accuracy are not effectively improved.In order to improve the accuracy and speed of prediction,we first conducts a detailed analysis from the perspective of features,innovatively proposes the concepts of common and unique features,and finds that common features are suitable for centralized modeling,and unique features are suitable for local modeling.Based on the above observation,we propose a common-unique feature based feteraged learning(CUFL)model,which includes three modules: a common feature prediction module,a unique feature prediction module,and a merge module.Specifically,the common feature prediction module of each cell obtains the overall distribution trend of the service of each cell through federated learning,and the unique feature prediction module of each cell learns the specificity of the service distribution through local training.The merge module integrates the above two sub-modules.The prediction results of the fusion are combined to effectively combine the advantages of the two to ensure the effectiveness of the model.We verify the effectiveness of the proposed CUFL model through the actual measurement data on the live network.Compared with the current typical STCNet centralized prediction model,the accuracy is increased by 10%,and the convergence speed is increased by more than 50%,which shows the effectiveness of the proposed model.Furthermore,in order to enable our proposed CUFL model to be stably implemented in a wireless network environment and to ensure the prediction performance of the model,we designed a CUFL model parameter transmission protocol.First,we analyze the prediction performance of the model from the three aspects of cell selectivity,link quality and clock size.It is found that,on one hand,selecting highly correlated cells can obtain more effective global parameters,thereby improving the prediction accuracy of the model,on the other hand,reliable transmission channels and appropriate clock settings can improve the prediction speed while ensuring the accuracy of the model prediction.In view of this,we design the CUFL model parameter transmission protocol based on the wireless channel.By selecting the cell with high correlation and good link quality,and setting a reasonable clock,the prediction performance of the CUFL model under the wireless channel is guaranteed.At the same time,in order to solve the problem of model update instability caused by unreliable transmission,a missing parameter learning algorithm and model update mechanism are proposed to ensure the robustness of model training and the flexibility of update.Finally,the CUFL model parameter transmission protocol under the wireless channel is implemented in the semi-physical simulation communication system based on Opnet and USRP to verify the predictive performance of the CUFL model under this protocol.