Deep Learning for Optimizing Virtual Resources Allocation in MEC-based UAV Environments

Abstract

Real-time applications requiring ultra-low latency are driving the deployment of Mobile Edge Computing (MEC) closer to end users. However, when deployed on unmanned aerial vehicles (UAVs), MEC infrastructures face critical constraints in processing, storage, and energy resources. This paper proposes a virtualization-aware MEC framework that integrates GRU-based CPU and memory demand forecasting with dynamic virtual re source allocation in UAV-assisted environments. The architecture combines UAVs, MEC servers, and a central orchestrator to support Collision Detection and Avoidance (CDA) applications through proactive resource provisioning. The proposed approach was evaluated on two heterogeneous UAV datasets, namely OpenSky ADS-B traces and AU-AIR. Experimental results show that the integration of predictive deep learning with virtualization improves resource allocation efficiency compared with non virtualized baselines. At the same time, GRU achieves more accurate and stable forecasting than RNN and LSTM models. These findings show the effectiveness of proactive virtualization aware orchestration for the UAV-assisted MEC systems.

Keywords