D4.2 - Initial Evaluation of BeGREEN O-RAN Intelligence Plane, and AI/ML Algorithms for NFV User-Plane And Edge Service Control Energy Efficiency Optimization

This document, BeGREEN D4.2 deliverable, presents the development status and the initial validation of the BeGREEN Intelligence Plane and the methods assisted by proposed Artificial Intelligence (AI) and Machine Learning (ML) to enhance the energy efficiency of the Radio Access Network (RAN) and Edge domains. This work builds on the concepts established in BeGREEN D4.1 [1], by introducing and refining the architecture, mechanisms, and use cases designed to reduce energy consumption without impairing network performance. The deliverable also reports an initial evaluation of the proposed solutions. BeGREEN D4.2 is structured in two main chapters. The first one deals with the architecture of the BeGREEN Intelligence Plane, a cross-domain framework designed to integrate AI/ML processes within the O-RAN architecture with the objective of enhancing the decision-making process of rApps and xApps. As the main novelty, the Intelligence Plane incorporates the AI Engine, which hosts ML models and associated services, offloading them from the RAN Intelligent Controllers (RICs). By decoupling the AI/ML services from the O-RAN control loops, the BeGREEN Intelligence Plane offers a modular and reusable framework that allows for independent model development and deployment. Additionally, the deliverable discusses extensions to allow the integration of Edge and Core domains, and of RAN technologies which are currently beyond the scope of traditional O-RAN implementations like Relays, Reconfigurable Intelligent Surfaces (RIS), and Integrated Sensing and Communication (ISAC). Finally, it is examined how to manage and mitigate conflicts across the RICs between contradictory optimisation policies. The second main chapter is dedicated to the evaluation of AI/ML-based solutions that enhance energy efficiency, including the Intelligence Plane itself. First, it presents dimensionality reduction techniques to minimize data inputs for ML models without impacting the model accuracy. This approach reduces data processing overhead and improves the energy efficiency of ML models. The compute resource allocation in virtualized RAN (vRAN) scenarios is also analysed, and Reinforcement Learning (RL) algorithms to dynamically decide on resource allocation according to network load and power consumption patterns are proposed. Another proposed strategy is the carrier on/off switching technique, which uses ML to predict traffic patterns and switch off unnecessary 5G carriers during low-demand periods, significantly reducing energy usage by offloading traffic to 4G carriers. The deliverable also details AI/ML approaches controlling fixed and UE-based relays to ensure that network resources are utilized more effectively, leading to both improved coverage and reduced energy consumption. In particular, the proposed mechanisms address coverage hole detection (CHD), fixed relay placement, candidate Relay UE (RUE) identification and dynamic relay activation and deactivation. Regarding the Edge domain, two main uses cases and methods are considered. The first one proposes a dynamic allocation of the compute resources dedicated to the User Plane Function (UPF) according to the forecasted traffic demand. The second one proposes a Bayesian online learning algorithm addressing the joint orchestration of vRANs and Edge AI services, aiming at minimizing overall power consumption while meeting the service's performance constraints. Initial evaluations of these AI/ML-assisted solutions show promising results in terms of energy savings and performance optimization in both the RAN and Edge domain. The deliverable includes use case studies, technology characterization and experimental validations demonstrating how these algorithms can achieve substantial energy reductions in both the RAN and Edge domains. In the case of the Intelligence Plane, the reported validation assesses the baseline architecture and operations, focusing on the AI Engine and the Non-Real-Time RIC. The final evaluation of the Intelligence Plane and the AI/ML-assisted mechanisms will be reported in the upcoming BeGREEN D4.3.