Research On Intelligent Management Of RAN Slicing Based On Federated Load Awareness

Network Slicing(NS)is a vital technology that enables flexible orchestration and on-demand scheduling of communication network resources in the fifth-generation mobile communication system(5G).However,managing NS on the radio access network(RAN)side poses significant challenges due to the dynamic and time-varying wireless environment and diverse user requirements.Meeting these challenges requires a highly intelligent wireless network.To achieve efficient and customized RAN slice resource management,it is crucial to accurately perceive and intelligently predict network load patterns.Unfortunately,the existing centralized resource management model presents several issues,such as high data transmission overhead,high latency,and sensitive user data privacy,making it difficult to achieve slice-level load awareness.To address these issues,federated learning(FL)is considered a potential technology for realizing distributed intelligence management in future wireless networks.This thesis focuses on developing an intelligent management scheme for RAN slicing based on federated learning load awareness.The proposed scheme operates under the FL wireless distributed intelligence framework and performs distributed perception on the network load generated by RAN slice user traffic and mobility.Through distributed prediction,the scheme obtains the globally optimal slice resource reservation and access selection strategy.Moreover,federated transfer learning(FTL)is used to improve the generalization of the prediction model and enable efficient management of RAN slice resources.To achieve this goal,this thesis proposes an overall framework for intelligent management of RAN slices based on federated load awareness and conducts research in the following three aspects within this framework:Firstly,the current wireless environment is dynamic and time-varying,and multiple heterogeneous services coexist.There are problems such as coarse-grained resource orchestration,centralized management,high data transmission overhead,and high delay.It is difficult to achieve slice-level traffic prediction.Therefore,this thesis proposes an efficient slice resource reservation scheme based on federated slice-level traffic prediction.This solution achieves accurate perception of slice service traffic load through multi-base station federation collaboration and reduces slice Service Level Agreement(SLA)violation rate by designing differentiated prediction loss function.The simulation results show that the proposed prediction scheme has higher accuracy and can reduce the SLA violation rate by 41%.Furthermore,the proposed scheme improves resource utilization by 15%through federated traffic prediction and fine-grained resource orchestration.Secondly,the thesis proposes a RAN slice access selection scheme based on federated mobility prediction to address frequent changes in wireless network status caused by user mobility,high mobile handover overhead,and increased slice management complexity.The scheme achieves this by improving the prediction accuracy of user trajectory through multi-base station cooperative perception of user spatial mobility without continuous sharing of user data.Additionally,a user location-based RAN slice access selection scheme is proposed to achieve optimal matching of user-base station-RAN slices and scheduling of slice resources to specific users on a small time scale.The proposed scheme reduces the handover overhead and ensures the continuity of slice services during user mobility.Compared with most traditional access selection schemes,the proposed scheme effectively improves system throughput and reduces handover overhead by 35%～42%.Finally,the thesis proposes an inter-slice cross-domain load sensing based on federated transfer learning to address the problems of unreusable inter-slice models,poor model generalization ability,high training data requirements,high data transmission,and model training overhead in the above two research schemes.The scheme utilizes the similarity between loads to realize the reuse of the load prediction model between slices,improve the generalization of the prediction model,and further reduce the overhead of data transmission and model training while ensuring a certain prediction accuracy.The simulation results show that the proposed scheme can improve the learning performance of the predictive model and reduce the prediction mean square error of the intra-cluster transfer learning by 33%～47%.Additionally,the proposed scheme reduces the overhead of data transmission and model training by about 64%compared to the federated traffic prediction scheme in the first research scheme.