Analysis And Prediction Of Base Station Traffic Based On Spatial-temporal Characteristics

With the continuous improvement of mobile communication technology and the advent of 5G era,people's demand for mobile data flow is increasing.Mobile Internet is also greatly meeting the needs of people's fast and convenient access to the Internet at anytime and anywhere.The number of users using smart devices and mobile services is also growing.At present,China has the largest mobile communication market in the world.With the increasing number of mobile users,mobile communication operators need to manage wireless networks more intelligently to provide better services.Accurate prediction of the mobile traffic of cellular network base stations can effectively promote the intelligent construction of wireless networks.With the development of artificial intelligence and Internet big data,on the one hand,more and more researchers study using neural networks to predict the base station traffic of cellular networks.In theory,the neural network model with multiple hidden layers can achieve the approximation of any nonlinear function;on the other hand,more and more analysis work also found that the base station traffic of cellular networks not only has strong autocorrelation in the time dimension but also has strong cross-correlation with other base station traffic in the spatial dimension.In this thesis,by mining the real cellular network data,the spatial-temporal correlation of base station traffic and the potential correlation between base station traffic features and non-traffic features are analyzed.Based on the neural network model,a spatial-temporal prediction model for cellular network base station traffic is proposed.On this basis,the simplification of the model in one-dimensional highway or railway scenarios and the introduction of non-traffic features are studied.The main work of this thesis includes:1.Data mining and spatial modeling based on cellular network base station traffic data.Firstly,based on the real network data provided by one communication operator,the temporal autocorrelation and spatial cross-correlation of the base station traffic of cellular networks are analyzed.The correlation between the non-traffic features such as downlink utilization PDCCH,and uplink or downlink traffic is studied by statistical analysis methods and analysis methods based on ensemble learning.It is found that the base station traffic of cellular networks has obvious spatial-temporal correlation and potential correlation with other non-traffic features.At the same time,it is found that the contribution of the base station data at different times to the base station traffic prediction is different.In addition,based on the spatial characteristics of cellular networks,a spatial modeling method is proposed to model the regional level mobile data of base stations.2.The DI-DenseNet model which is the spatial-temporal prediction model of base station traffic is proposed.The input data are divided into closeness dependent data and period dependent data,which are respectively input into two parallel CNN structures,and the output results are fused to achieve high-precision spatial-temporal multi-step prediction.The DI-DenseNet model can get the traffic prediction results of all base stations in the prediction area in one prediction.The experimental results show that in two-dimensional urban and rural areas,the average single base station training cost of the DI-DenseNet model is about 59.3%less than the LSTM model.The average accuracy of five-step prediction of uplink and downlink traffic is similar to LSTM model,which is 17.9%and 5.0%higher than SVR model,and 3.2%and 6.9%higher than CNN model.At the same time,the simplification of the DI-DenseNet model in one-dimensional highway or railway scenarios is studied,which can reduce the computational complexity of the model by about 82.0%without affecting the prediction accuracy when predicting the cellular network base station traffic of highway or railway scenarios.3.The influence of non-traffic features on the prediction results of the spatial-temporal model is studied.Two-level standardization operation is proposed to eliminate the potential impact on prediction model from different orders of magnitude of adjacent base station data.The two-level standardization keeps the trend of the time sequence change of base station data.The experiment shows that the two-level standardization effectively improves the accuracy of traffic prediction after introducing non-traffic features.Then,the single-step mobile traffic prediction experiment based on non-traffic features is carried out.It is found that the introduction of uplink utilization PUSCH and other non-traffic features can improve the accuracy of traffic prediction.At the same time,a multi-step prediction strategy based on auxiliary feature prediction model is proposed to carry out multi-step prediction after introducing non-traffic features.The experiment shows that after introducing non-traffic features,the average error of five-step prediction of uplink traffic decreases by about 4.3%,the average error of the five-step prediction of the downlink traffic decreases by about 4.6%.