A federated learning approach for inter-cell handover optimisation in open 6G ran

Abstract

Fifth-generation (5G) and beyond mobile communication systems need to support diverse service types. These are enhanced mobile broadband (eMBB), ultra-reliable low-latency communications (URLLC), and massive machine-type communications (mMTC), while operating in unpredictable radio and traffic environments. Handover decisions directly impact data protection, service continuity, and traffic distribution between radio access network (RAN) nodes, creating fundamental challenges in network mobility management. The 3GPP-defined traditional handover systems operate through real-time signal strength measurements, which include reference signals received power (RSRP) and reference signal received quality (RSRQ) values, together with fixed threshold values and time-to-trigger parameters. These methods provide strong performance, but they operate as reactive systems that do not utilise time-based traffic patterns, user movement data, and cell traffic levels for optimisation. This thesis develops a mobility-based handover optimisation system which uses federated learning (FL) within the open radio access network (O-RAN) infrastructure. The system proposal uses near-real-time RAN intelligent controller (near-RT RIC) programmability to run xApps, which work together for closed-loop handover control. The system contains four essential xApps, which include a Metrics Collector for immediate KPI data collection, an FL-based traffic predictor for load prediction, an FL coordinator for global model unification and a handover handler, which performs E2 interface-based mobility execution. The system employs federated learning to train models across gNBs while maintaining user measurement data locally, which reduces backhaul traffic and preserves data storage at its original location. The training process of a recurrent neural network model uses multi-dimensional feature vectors which unite radio measurement data with traffic load indicators and user velocity and direction information. The federated averaging (FedAvg) algorithm combines model updates to create a global predictor which operates throughout the network. The proposed framework is implemented and evaluated on an end-to-end open-source testbed combining OpenAirInterface (OAI) gNBs, an Open5GS 5G Core, and a FlexRIC-based near-RT RIC. The system operates under actual Gauss–Markov mobility and Pareto traffic models, which include various UEs and working gNBs for performance evaluation. The experimental data show that the proposed FL-based system achieves higher handover success rates, enhanced inter-cell load distribution, and lower packet loss rates than A2/A3-based and centralised machine learning approaches while keeping decision times at real-time levels.