Handover Management Based On Machine Learning In Ultra-dense Network

Due to the explosive growth of various smart devices and Internet applications as well as people's high dependence on mobile communications in daily life,wireless data traffic is increasing dramatically.As one of the key technologies of 5G,the ultra dense network(UDN)is regarded as an effective solution.However,such a large-scale deployment of base stations(BSs)has caused several problems in handover management,including frequent handovers,sharply increased signaling costs,and huge energy consumption by BSs.These issues are closely related to the network architecture.In the traditional cellular network architecture,control coverage and traffic coverage are tightly coupled.BSs are not only responsible for the wireless signal transmission and reception,but also for signal processing,resource scheduling,handover and synchronization control services.This makes handover management in UDN more complicated.The network architecture with a separation between control plane and user plane(CP/UP)can overcome the limitations of the traditional cellular network architecture.Nonetheless,the passive triggering strategy commonly used in traditional handover algorithms prevents the mobile network from reserving resources,which has a negative impact on both the user side and the network side.The traditional handover based on prediction has the problems of high computational complexity and low prediction accuracy.Therefore,traditional handover algorithms are no longer suitable for UDN scenarios,and more efficient handover algorithms are needed.With the continuous development of artificial intelligence,deep learning is expected to be a data-driven method to achieve the optimized handover decision in UDN.Therefore,noting the above-mentioned challenges of handover management in UDN,we study the handover management algorithm based on deep learning in UDN.The main reaserch focuses and key contributions are listed as follows.This thesis proposes a handover prediction scheme based on deep learning in the CP/UP separation network architecture,which can predict the target base station with higher accuracy and reduce the signaling costs.This thesis designs a three-layer Long and Short Term Memory(LSTM)structure.Based on this structure,the number of the target BS is predicted,this allows the target BSs to make resource preparations in advance to reduce the signaling costs related to handover.Simulation results show that compared with the traditional non-predictive solution,the prediction handover management strategy based on LSTM proposed in this thesis can significantly reduce the signaling costs.This thesis proposes handover-aware energy-efficient BS on/off framework and method based on deep learning.Under the network architecture with CP/UP separation,the proposed framework is based on LSTM to predict the user location at the next moment,and then predicts the handover and uses it to estimate the base station load at the next moment.It is further based on the Feedforward Neural Network(FFNN)to predict the best base station on/off configuration through the user's location information and data rate requirements,thereby improving the energy efficient of the network(EE).Numerical results show that the proposed framework significantly improves the energy efficient of the network compared to the traditional scheme without BS switching and reactive solution.