Personalized federated learning for dwell-time-aware handover optimization in ultra-dense vehicular networks

Hou, YunYunHouZhang, ZhiZhiZhangZhang, YutingYutingZhangGeorgiades, MichaelMichaelGeorgiadesDr. HO Kin-Hon, RoyRoyDr. HO Kin-Hon2026-06-032026-06-032026IEEE Transactions on Vehicular Technology, 2026.0018-95451939-9359http://hdl.handle.net/20.500.11861/27303The development of connected vehicular technologies, such as electric vehicles and autonomous driving, demands ultra-reliable low-latency communication for Vehicle-to-Everything (V2X) safety applications, where stable connectivity outweighs throughput requirements for disseminating small safety-critical packets. In 5G and future ultra-dense networks, shrinking cell coverage and overlapping service areas exacerbate frequent handovers, leading to increased signaling overhead, packet loss, and communication interruptions that jeopardize time-sensitive V2X operations. To mitigate frequent handovers (HOs), this paper proposes a novel dwell-time-aware HO scheme that selects the target cell with the maximal remaining dwell time (RDT), thereby extending vehicle-cell association duration and reducing HO frequency. The RDT prediction leverages a deep learning model trained via a personalized federated learning (PFL) framework using UE-reported measurements, without user location data exchanges. Within this architecture, base stations collaboratively optimize shared model layers while maintaining personalized layers locally to address spatial data heterogeneity. Simulations demonstrate that our framework achieves high RDT prediction accuracy (test R20.91) and reduces HOs by 9.89% compared to traditional RSS-based methods without compromising connection quality significantly, enhancing the stability for V2X safety applications.enFederated LearningVehicular NetworksDeep LearningDwell TimeHeterogeneous DataBase StationLow-Latency CommunicationsStable ConnectionSignaling OverheadSmall PacketsUltra-Dense NetworksGlobal ModelCentral UnitHeterogeneous NetworkTraining CenterTraffic LightSpeed ValuesDeep Reinforcement LearningBase LayerUser EquipmentFast FadingLocal TrainingShadowing EffectCommunication RoundsMulti-Agent Reinforcement LearningOpenStreetMapDedicated Short Range CommunicationArtificial Intelligence TrainingFederated Learning AlgorithmPersonalized federated learning for dwell-time-aware handover optimization in ultra-dense vehicular networksPeer Reviewed Journal Article10.1109/TVT.2026.3697268