Modeling And Application Of Spatio-temporal Traffic Based On Deep Learning

With the development of 5G technology,the increasingly dense net-work architecture and the emergence of various emerging mobile applica-tions,meeting user demand for network traffic has become an important issue for carrier companies to address.The accurate prediction of network traffic faces many challenges due to the randomness of user mobile behav-ior,the complexity of traffic requests,and the spatial correlation and con-straint of traffic distribution.This thesis addresses this problem,and the main work and innovations done include:Firstly,this thesis proposes a traffic prediction model based on the spatio-temporal characteristics of network traffic in the scenario of ultra-dense network,and proposes a method of spatial fusion division,firstly dividing the spatial region and capturing spatial features based on the 3DCNN algorithm;secondly capturing temporal features in the traffic se-quence based on the LSTM algorithm.In order to verify the effectiveness of the proposed model,the thesis conducted simulations based on the actual data provided by the operators in the school-enterprise cooperation project.The experimental results show that the proposed model can effectively cap-ture the spatio-temporal characteristics of the traffic and significantly re-duce the prediction error,compared with the traditional algorithms ARIMA and LSTM,the standard root mean square error is reduced by 20%and 2%respectively.Secondly,this thesis proposes an improved LSTM algorithm,which is based on the idea of multi-task learning in deep learning,and achieves parameter sharing in the hidden layer of the network,so as to complete the multi-dimensional prediction of network traffic.After the model is applied to the actual data provided by the operator companies,the simulation re-sults show that compared with the traditional algorithms ARIMA and LSTM,the standard root mean square error of the network traffic predic-tion results of the LSTM model proposed in this paper is reduced by 11%and 3%respectively,and the generalization ability of the model is en-hanced.Again,this thesis implements further optimization improvements on the basis of Conv-LSTM.The improvement idea combines the classifica-tion algorithm in data mining,constructs a binary classification model of whether the network traffic is abnormal based on the spatio-temporal pre-diction results of the network traffic,and designs the value function of this binary classification model as a metric for model performance evaluation.After the model is applied to the actual data provided by the operator com-panies,the simulation results show that the overall correctness of the binary classification prediction of the Conv-LSTM model is 68.5%,and the com-putational complexity of the model is reduced.Finally,this thesis proposes a load balancing scheme based on the spa-tio-temporal traffic prediction results,using particle swarm algorithm to optimize the historical traffic data,a distributed real-time scheduling of the prediction results,and a regression algorithm to propose a final load bal-ancing scheme for each cell based on the optimization results.The experi-mental results show that in the simulation scenario given in this thesis,the network resource utilization using this scheme is improved by 6.54%,which has certain practical significance.In summary,this thesis investigates the prediction of spatial and tem-poral characteristics of network traffic and the application of load balanc-ing,which is of great significance for the improvement of user experience and the progress of communication network technology.