Achieving High Energy Efficiency for Network Slicing-Enabled 5G O-RAN Base Stations

With the rapid advancement of 5G networks, data-intensive services has been surging and significantly increasing energy consumption. It is a critical challenge in addressing energy efficiency for 5G access networks. In this paper, we propose an effective load balancing algorithm that optimizes energy efficiency as users access multiple services from base stations (BS). Utilizing network slicing, the physical network is segmented into multiple independent virtual networks. We have applied graph neural networks (GNN) to extract features from dynamic traffic, model relationships between different services, and manage resource competition among BSs and network slices, treating each network slice as a distinct node. The GNN model predicts nodes workload, guiding the allocation of frequency domain resources and facilitating load balancing. Additionally, to reduce operational times and minimize disparities among slices at each BS, we propose a deep reinforcement learning (DRL) strategy. This strategy is integrated with the C-states mechanism of CPUs on general x86 computing platforms to further enhance energy efficiency. In simulations, our GNN algorithm demonstrated high precision in predicting and allocating frequency domain resources, achieving a high accuracy of 99.8%. Our DRL algorithm showed a maximum global energy efficiency improvement of 12% compared to benchmark algorithms. Overall, our approach enhanced energy efficiency by an average of 3.155 B/J under conditions of resource competition, significantly outperforming strategies that allocate equal physical resource blocks (PRB).