How AI‑Powered 5G Cloud RAN Can Cut Base‑Station Energy Use by Up to 30%

Background

5G cloud‑native radio access networks (RAN) replace vendor‑specific BBUs, RRUs and antennas with software‑defined CU (Centralized Unit), DU (Distributed Unit) and AAU (Active Antenna Unit) running on generic x86 servers. The open architecture provides unified O‑RAN interfaces, multi‑vendor interoperability and high‑frequency telemetry (e.g., user count, PRB utilization, CPU load) that can be exploited by AI/ML applications.

Reference Architecture

The solution builds on Intel FlexRAN reference design, Intel Xeon D system‑on‑chip processors and the Intel vRAN ACC100 accelerator. These components deliver the compute density required for real‑time radio functions and AI inference workloads.

Intelligent Energy‑Saving Workflow

During system initialization, historical KPI series (e.g., user count, traffic load, PRB usage, CPU utilization, power draw, frequency) are collected from the base‑station protocol stack and the host server to establish cross‑source relationships.

Continuous telemetry feeds a training dataset for time‑series AI/ML models.

The Chronos framework trains inference models and publishes predictions to the RAN intelligent‑energy rAPP/xAPP components.

rAPP/xAPP combines model output with long‑term configuration to forecast future KPI trends and generate power‑state actions.

Models are periodically retrained with newly measured KPI data, enabling continual learning.

When forecasts indicate low load (e.g., few active users, low traffic, PRB utilization < 10 %), the system reduces CPU frequency or powers down cores using Intel P/C‑state mechanisms, achieving second‑level granularity without QoS impact.

Performance Evaluation

Laboratory and field trials show a 15 %–30 % reduction in total base‑station energy consumption while maintaining service quality. The AI‑driven approach also reacts to traffic spikes within seconds, outperforming static rule‑based schemes that typically achieve ≤ 5 % savings.

Chronos Framework Details

Data Processing & Feature Engineering : Over 70 built‑in tools accessible via the TSDataset API for rapid preprocessing (missing‑value imputation, scaling, feature extraction).

Built‑in Models : More than ten time‑series forecasters, detectors and simulators (e.g., LSTM, Prophet, ARIMA) ready for out‑of‑the‑box use.

Hyperparameter Optimization : AutoTSEstimator automates model selection, feature selection and hyper‑parameter tuning, providing a full AutoML pipeline.

Chronos integrates with Intel’s optimization stack—ONNX Runtime, OpenVINO, and the oneAPI AI Analytics Toolkit—to accelerate training and inference on Xeon D CPUs.

System Integration

Telemetry is collected with an Intel‑optimized Telegraf agent and other open‑source collectors, then pushed via standard E2 or proprietary interfaces to an InfluxDB time‑series database.

Trained models run on the same host as the RAN software; predictions are consumed by rAPP (non‑real‑time) and xAPP (real‑time) components.

Power‑state decisions are enacted through Intel P‑state (frequency scaling) and C‑state (core idle) controls, configurable per‑core via the Linux cpufreq subsystem.

Processor Power‑State Control Example

Thresholds used in validation tests:

If predicted user count < 15, traffic < 40 Mbps, PRB utilization < 10 % or CPU utilization < 15 %, the Xeon D processor is set to a low‑power P‑state (≈ 1 GHz).

If any metric exceeds the above limits, the processor is switched to a high‑performance P‑state (≈ 1.9 GHz).

These dynamic adjustments enable fine‑grained energy savings while preserving latency and throughput requirements.

Future Directions

The same AI‑enabled closed‑loop methodology can be extended to 5G core‑network functions, edge compute nodes and data‑center workloads, supporting broader green‑network objectives across the telecom ecosystem.