Scalable machine learning-based approaches for energy saving in densely deployed Open RAN
A new federated TD3 algorithm achieves 43.75% faster convergence and slashes network energy consumption.
A team of researchers, including Xuanyu Liang, Ahmed Al-Tahmeesschi, and Cicek Cavdar, has published a paper proposing novel machine learning methods to tackle the massive energy consumption of densely deployed cellular base stations in Open Radio Access Networks (Open RAN). Their work leverages the flexibility of Open RAN's disaggregated architecture to embed AI directly into network operations. They propose three different Deep Reinforcement Learning (DRL)-based solutions, implemented as xApps, to intelligently control the active/sleep modes of radio units (RUs) from the Near-Real-Time RAN Intelligent Controller (RIC).
To address scalability across large networks, the researchers' key innovation is a federated DRL framework. This system uses a central aggregator as an rApp in the Non-Real-Time RIC, coordinating with local DRL agents deployed as xApps. This distributed approach reduces the data and computational burden of a fully centralized AI model. In simulations across layouts with 6 to 24 RUs covering 500-1000 meter regions, their proposed federated Twin Delayed DDPG (TD3) algorithm significantly outperformed centralized baselines.
The results are compelling for network operators facing soaring energy costs. The federated TD3 solution demonstrated up to 43.75% faster convergence during training. More importantly, it achieved more than a 50% reduction in overall network energy consumption while successfully maintaining required coverage and quality of service. The federated approach itself is also more efficient, requiring 37.4% lower energy for the training process compared to centralized training, and it produced more robust control policies. This research provides a practical blueprint for deploying scalable AI to make next-generation Open RAN deployments both high-performance and sustainable.
- Proposes a federated Deep Reinforcement Learning (DRL) framework with a TD3 algorithm, deployed as rApps and xApps in the Open RAN Intelligent Controller.
- Simulation results show over 50% reduction in network energy use and 43.75% faster algorithm convergence versus centralized baselines.
- The federated approach also cuts AI training energy by 37.4% and improves policy robustness while maintaining service quality across 6-24 radio unit deployments.
Why It Matters
This could dramatically reduce the massive energy footprint and operational costs of 5G/6G networks, making dense cellular coverage more sustainable.