How Can AI Put Network Towers to Sleep and Save Energy?

The upcoming 6G networks promise lightning-fast speeds, but they also demand huge amounts of energy. The biggest culprit? Radio units (RUs) in base stations stay on even during low traffic, wasting power. Researchers tackled this with artificial intelligence, using an algorithm called TD3 that works with continuous actions, unlike older discrete methods like DQN. TD3 lets RUs switch to various sleep levels more precisely, avoiding an explosion of possible actions. Simulations showed the new approach cut energy use by over 50% compared to an always‑on baseline and outperformed DQN by about 6%. Moreover, TD3 proved more stable and learned faster. In short, AI could put network towers to sleep when they’re not needed, shrinking the ecological footprint of future mobile networks.

The TD3 (Twin Delayed Deep Deterministic Policy Gradient) method is a variant of continuous reinforcement learning, designed to handle large action spaces without discretization. In an O‑RAN context, RUs can operate in multiple power states: full‑on, micro‑sleep, deep‑sleep, etc. Each state has a specific energy consumption and a different impact on latency and responsiveness. TD3 employs two “twin” critic networks to reduce variance in Q‑value estimation, while a delayed policy update mitigates over‑fitting. This yields faster convergence compared to DQN, which suffers from combinatorial explosion when trying to model all sleep states as discrete actions. Moreover, TD3’s stability is enhanced by gradient clipping and slowly updated target networks. Experimental results show that, in high‑cell‑density scenarios, the algorithm maintains QoS (Quality of Service) within acceptable limits while cutting average power consumption by over 50%. This approach is especially promising for future 6G architectures, where flexibility and energy efficiency will be essential requirements.