Research On Mobile Network Traffic Prediction And Base Station Sleep Control Based On Deep Learning

In recent years,the large-scale construction of 5G has led to the explosive growth of mobile traffic demand.The intensive deployment of base stations directly leads to a sharp increase in system energy consumption.Base station sleep technology has become a key research direction to improve the energy efficiency of communication networks and achieve sustainable development because of its advantages of easy implementation without hardware replacement,effective reduction of carbon emissions,and significant energy saving results.In the wireless communication network architecture,the design and planning of the network follow the principle of maximum load.During the low load period,keeping all base stations active causes unnecessary waste.Therefore,the energy consumption can be saved by accurately predicting the traffic demand and dynamically turning off the low-load base stations.However,when some base stations are turned off,the service delay may deteriorate due to the redirection of traffic,and the relationship between energy saving and service quality degradation should be carefully balanced.In addition,base station sleep control is also subject to several practical constraints,such as carbon emission allowance,service migration cost,service delay,hardware wear and tear,etc.To solve the above problems,this paper combines the traffic prediction technology with the base station sleep control technology,and the goal is to provide users with a stable service experience under the condition that the system energy consumption does not exceed the given long-term energy consumption target.Firstly,the historical data and network traffic characteristics were used to predict the traffic load of the base station in the future.Specifically,a new graph neural network layer with location attention mechanism is proposed to aggregate dynamic information from nearby base stations.We also combine recurrent neural networks and Transformer to capture local and global temporal dependencies.Secondly,a base station sleep control method based on deep reinforcement learning was proposed to optimize resource allocation based on the prediction results.Specifically,the actor-critic architecture and the dual-delay deep deterministic policy gradient algorithm are used to accelerate the training.A new benchmark transformation strategy is designed to reduce the variance of the cost estimation caused by the change of the underlying business requirements.Finally,extensive experiments are carried out using public real datasets,and the results verify the accuracy of the traffic prediction method and the effectiveness of the base station sleep control method in improving the energy efficiency of the system.