Deep Learning Based Handover for High-Speed Connected Vehicles in Ultra-Dense Networks

In Vehicle-to-Everything (V2X) applications, vehi-cles pass through network cells at high speeds, with ultra- dense deployments of base stations, frequent handover becomes a prominent issue to be addressed. Especially in V2X road- safety use cases, maintaining a long-term stable connection to the cell is more important than having high immediate throughput. Therefore, the traditional handover scheme that selects the base station with the highest signal strength without considering the remaining dwell time may lead to even more frequent handovers. In this paper, we propose a new handover scheme that aims to select the target cell with maximal remaining dwell time, such that after handover the user equipment will camp on the new cell without the need to handover again soon. The remaining dwell time for the potential target cells is predicted based on the user equipment's local measurement of received signal strength, driving speed and direction using ID-CNN deep learning networks. The new handover scheme does not require any private user information to be exchanged with the BS. Validation results show that our ID-CNN-based deep learning network provides high-accuracy predictions of remaining dwell time achieving an R2 score of 0.96. Network level simulation results demonstrate that the deep learning network based handover can reduce 74% of unnecessary handovers in ultra-dense scenarios.