Time Series Modeling And Forecasting Of Communication Networks Based On Deep Learning

The rapid development of cellular technology and the rapid deployment of network infrastructure has provided mobile communication users with an ever-improving communications experience and raised their expectations of the quality of service provided by mobile networks at the same time.To meet these expectations,mobile operators need to actively and efficiently allocate network resources during network operations and management,in addition to timely deployment of more advanced physical layer communication technologies.In this context,load forecasting of base stations in mobile networks is an important prerequisite for optimal allocation of network resources.Accurate base station load forecasts,such as data throughput per time unit,number of active user calls,etc.,can effectively support network planning,load balancing and energy saving optimization.However,due to the randomness of mobile users’ communication needs,The group and number of subscribers served by each base station varies over time,and combined with short-term and largescale migrations of subscribers due to modern transport systems and the impact of special events,forecasting mobile base station load has become a challenging task.In this paper,a deep learning modelling method is used to train the communication load prediction model using historical load data from different base stations,and the spatio-temporal correlation of historical load data of base stations is extracted to improve the prediction accuracy of base station load.This paper proposes two new base station load prediction models based on the Gate Recurrent Unit(GRU)model.By optimizing the feature selection and structure of the new models,they can extract the spatial and temporal correlation of the base station load history series more comprehensive and accurate,and improve the base station load prediction precision.In this work,a Dual-Channel GRU(Dual-channel Gate Recurrent Unit)network is proposed to predict the future load of a single base station using its historical load data.The Dual-Channel GRU uses two GRU channels to extract the temporal continuity and periodicity features of the base station load data respectively,and integrates the temporal features of the predicted time periods directly into the features extracted by the Dual-GRU channel to enhance the model’s ability to learn temporal correlations.By verifying base station load data collected from the real-world communication network,the model can learn the characteristics of different types of base station loads and has good adaptability to different types of base stations.Compared with conventional prediction algorithms such as ARIMA,CNN,GRU and CNN-GRU,the experimental results show that the Dual-channel GRU model can provide higher accuracy of base station load prediction.On the basis of single base station load prediction,this paper further proposes a model based on Graph Neural Networks(GNN)and GRU(GNN-GRU)to jointly predict the load of multiple communication base stations.The GNN-GRU model uses the spatial information and historical load data of all base stations as inputs to predict the future load of all relevant base stations at once.In this model,the spatio-temporal correlation of the load data of different base stations is mined by extracting the time-spatial characteristics of the load time series of multiple base stations.The temporal features of the target prediction period are added to the input end of the model by using a method similar to the Dual-channel GRU model,so as to help the model better extract the features of the relevant period and realize the high-precision joint prediction of the multi-base station load.The experiment shows that GNN-GRU has higher base station load prediction accuracy than GRU,DCRNN and STCNet.