Denizer, BirkanBirkanDenizerLandsiedel, OlafOlafLandsiedel2025-11-282025-11-282025-10IEEE 50th Conference on Local Computer Networks, LCN 2025https://hdl.handle.net/11420/59279As remotely controlled and autonomous vehicles become widely available, the demand for high Quality of Service over cellular networks for their remote control and monitoring is becoming increasingly important. Accurate prediction of available uplink bandwidth is essential to mitigate bandwidth fluctuations and avoid impacting real-time applications, ensuring reliable and low-latency video streams. In particular, bandwidth overpredictions lead to packet losses, retransmissions, and significant latency increases, especially during network handovers, as network buffers fill up. Prior bandwidth prediction approaches lower absolute or relative errors but fail to address the impacts of overpredictions and the associated latency spikes.This paper introduces CapAware, a bandwidth prediction approach explicitly designed to minimize capacity violations (i.e., overpredictions) and reduce latency spikes during network handovers for uplink streams. It utilizes an efficient neural network architecture with an integrated handover prediction mechanism and a learnable capacity-aware loss function. CapAware predicts network handovers with a 92.4% F1 score and improves efficiency by 24.4% using its custom loss function with predicted handover information. Compared to deep-learning baselines, CapAware improves network efficiency (i.e., utilizationto-capacity violation ratio) by 4.7% and 34.9% on 5G SA datasets.enBandwidth predictionhandover predictioncapacity-awareutilizationoverprediction5GComputer Science, Information and General Works::004: Computer SciencesConference Paper10.1109/lcn65610.2025.11146351Conference Paper