Deep Resource Localization with Traffic Prediction for Adaptive Multi-access Edge Computing Management

Artificial intelligence (AI)-native services require real-time processing of large-scale data and a stable user experience. Multi-access edge computing (MEC) networks address these requirements by deploying distributed servers close to users. MEC servers have varied user densities and traffic loads depending on location and time, which necessitates localized resource management to align with regional traffic patterns. Resource localization requires predicting sudden traffic surges or declines in advance and dynamically allocating resources to enhance server availability and maintain service quality. This study proposes a Deep Edge LOcalization (DELO) framework, which comprises a deep-learning-driven network traffic prediction model and server resource allocation agent. The proposed prediction model extracts regional traffic characteristics via clustering and employs a temporal convolutional network-long short-term memory architecture to improve accuracy. The predicted traffic data serve as input for the resource allocation agent uses proximal policy optimization to balance server load. DELO adapts to local variations in traffic by offloading sudden traffic surges to proximate servers and enhances server resource utilization. We evaluated the performance of the proposed framework on the Milan dataset (Italy), imitating the real-world with MEC servers. The accuracy of the prediction model improves by up to 48.9% compared to that of existing time-series models. The allocation agent is evaluated against various reinforcement learning baselines for resource localization and the agent achieves 61.4% and 39.3% lower operational cost comparing to static and Greedy allocation methods, respectively. DELO also enhances server availability by up to 60.7% and 74.7% compared to static and Greedy allocation approaches, significantly improving network capacity and stability.