Handover Management Research On Mobile Terminal Side Based On Machine Learning

In order to satisfy the increasing demand for mobile communication system traffic and access,the ultra dense network could guarantee high system throughput and access amount by increasing deployment density of base station.However,dense deployment brings the problems of frequent handover,highly overlapping coverage and strong interference,etc.Meanwhile,the new generation communication system also has the characteristics of high communication rate and high reliability.All of the above characteristics have higher requirements on communication handover management to ensure the mobile users’ timely and reliable handover and reduce the burden on the system via cutting down ping-pong handover.This paper aims to utilize reinforcement learning methods to dynamically optimize the handover threshold and delay parameters of the typical handover process to improve the system handover management performance.Firstly,this paper reviews the typical handover process and methods,introduces key evaluation metrics,and then points out the importance of the optimization of handover thresholds and delay parameters on handover performance.In order to solve the problems of inflexible parameter configuration and poor compatibility of general reinforcement learning methods in ultra-dense scenarios,this paper proposes a handover event prediction model based on Long-short-term-memory(LSTM)network,which uses the RSRP time series of mobile terminals and each base station to predict handover failure and handover ping pong.The prediction method applies a federation learning architecture in order to solve the user privacy protection problem and reduce the data interaction overhead of distributed training.The simulation results demonstrate that the LSTM prediction in this paper can predict both types of handover events with high accuracy,and the federation learning architecture brings improved accuracy,privacy protection,reduced overhead and more flexible configuration.Thirdly,this paper proposes a Deep Q Network(DQN)method based on signal fading,which dynamically selects handover parameters according to the signal fading condition.The proposed method does not change the existing handover process and is an independent module with good backward compatibility.In order to solve the problems of inefficient data set collection and unstable convergence,a DQN method based on digital twin enhancement is proposed,which takes LSTM handover event prediction as a core component.Simulation results verify that the DQN design of this paper effectively optimizes the handover performance,while the proposed enhancement method outperforms the classical DQN method in terms of convergence stability,convergence speed and final performance.