D4.3 Final Architecture and Evaluation of BeGREEN O-RAN Intelligence Plane, and AI/ML Algorithms for NFV User-Plane and Edge Service Control Energy Efficiency Optimisation

This document, BeGREEN D4.3, consolidates BeGREEN’s final results on AI/ML-assisted energy efficiency for 5G/6G-ready RANs. It refreshes the Intelligence Plane introduced in BeGREEN D4.2, quantifies the AI/ML plane’s own energy footprint, validates closed-loop control with O-RAN components, and reports end-to-end gains across vRAN, live 5G NSA, relay-enhanced RANs, and edge scenarios hosting UPF and AI services. Chapter 2 revisits the Intelligence Plane, keeping the core AI Engine with near/non-RT RIC integration while adding hooks for CF-mMIMO and ISAC. Interfaces from BeGREEN D4.2 are preserved; new policies and telemetry enable sensing-assisted control and user-centric coordination. We analyse the energy cost of the AI/ML plane itself—training, serving, storage, orchestration—against the network-side savings it drives, showing that BeGREEN’s lightweight models achieve favourable trade-offs between size, accuracy, and inference rate. Final validations benchmark serving workloads, automate dataset creation and model training with energy metering, and verify AI Engine–RIC workflows on a representative cell on/off loop, including conflict management procedures. Additional scenario validations will be reported in BeGREEN D5.3 as part of PoC#1. Chapter 3 presents the final evaluation of five AI/ML-assisted methods to enhance energy efficiency. In vRANs, we study cache isolation and LLC utilisation and introduce MemorAI, a lightweight controller that steers cache/CPU to active threads, lowering platform power without harming throughput or latency. For carrier on/off, a compact classifier uses traffic/context features to trigger sleep/activation with guardrails, delivering energy savings without SLA regressions in offline and lab tests. In relay-enhanced RANs, we define RUE identification, compare activation policies, and quantify coverage, sum-rate, and uniformity gains versus no-relay baselines, including detailed power consumption. For UPF, we characterise compute–power curves, validate experimentally, and build a traffic-aware ML policy for scaling and placement that reduces energy on realistic traces at constant QoS. Finally, we jointly orchestrate vRAN and edge-AI services with a multi-objective coordinator that maintains QoS while cutting total energy in static, heterogeneous, and dynamic scenarios. Chapter 4 concludes that BeGREEN delivers an O-RAN-aligned Intelligence Plane extended for CF-mMIMO/ISAC, practical guidance to ensure the AI plane yields net energy savings, and validated AI-driven procedures that reduce energy without QoS penalties. The accompanying assets—configurations, policies, measurement scripts—are ready for replication in PoCs (D5.3) and for transfer to standardisation and industry pilots.