How AI decides which parts of a mobile network can go to sleep

A mobile network needs enough capacity for busy moments, but that equipment does not have to run at full availability every quiet hour. In June 2024, BT said it had rolled out machine-learning-based cell sleep across more than 19,500 EE sites. Selected 4G capacity carriers could rest when demand was low and reactivate within seconds if traffic increased.[1]

For the person holding the phone, the intended experience is an ordinary connection. Behind it is a changing allocation of radio capacity. The useful AI question is how much equipment can safely rest while the capacity that remains awake handles demand.

Putting spare capacity to sleep is a selective operation

A site can provide coverage and extra capacity through different frequencies and equipment. An ITU technical supplement describes moving users onto an available coverage layer before shutting down an additional carrier. If traffic rises, that extra capacity can return. The same document distinguishes carrier shutdown from reducing radio channels or creating brief gaps in transmission. These are different physical actions, with different control requirements.[2]

The equipment underneath matters. Ericsson’s operations guide describes combinations of antenna, radio and computing resources that can be reduced or shut down. A command called sleep therefore does not imply that an entire mast, its support systems and every service it provides have been switched off.[7]

AI adjusts the decision around the switch

Automatic shutdown itself does not require AI. The learned part can forecast local demand and choose suitable schedules or thresholds. The ITU account includes historical traffic, service patterns, special days and neighbouring capacity. A business district, residential area and transport corridor can need different policies even if their equipment is similar.[2]

BT’s account describes using each site’s traffic history to predict quieter periods. It also retains the ability to respond to unexpected demand. Forecasting helps identify an opportunity; a fast response handles the situation when the opportunity disappears. Both are needed for a network serving people who can change their plans.[1]

A saving needs both a time period and an equipment boundary

Vodafone and Ericsson reported a London trial in March 2025 combining predictive 5G deep sleep, 4G cell orchestration and power-related analysis. Their largest figures refer to selected radio units and particular periods. A 70% reduction during low traffic is not a 70% reduction in the operator’s annual electricity consumption.[3]

The older Madrid example is useful because its 24-hour figure includes the busy part of the day. Neither announcement provides the information needed to compare whole networks or to isolate how much of the saving came from machine learning rather than the equipment’s underlying sleep functions.[4][3]

Consider an invented example. One switchable component draws 1kW when active and 0.4kW asleep. It sleeps for six hours and stays active for eighteen. Other equipment draws 0.5kW continuously. These chosen inputs illustrate the arithmetic; they are not specifications for an actual mast.

The component saves 3.6kWh in a day. Adding unchanged equipment leaves those saved kilowatt-hours intact but reduces the percentage of the site total. A headline about the sleeping component and a bill for the whole site are answering different questions.

Predicting demand is only the beginning of planning

A July 2026 Ericsson Research paper explores a further step: learning what the network will do after a proposed action. Its world model predicts the next state given current conditions and a control choice. That helps a planner consider the effect of moving load between cells, rather than merely extrapolating tomorrow’s traffic.[6]

Its closed-loop control experiments take place in a simulator. Separate replay uses real network records, but the proposed alternative actions were not applied to the live network. More simulated sleep therefore does not establish an observed electricity saving. The study measures sleep time as a proxy, with actual energy depending on the hardware.[6]

The control also works at several speeds. The paper distinguishes slower policy decisions from local mechanisms that can wake capacity rapidly. Its hourly simulation does not model those fast wake-ups. That limitation matters when interpreting simulated service violations or transferring a policy into real operations.[6]

The useful target is an energy saving at an agreed service level

Ericsson’s technical discussion argues for explicit performance reference points: the throughput or other service level an operator intends to provide. Comparing energy without specifying that target can reward a network for doing less useful work. The evaluation needs both the electricity used and the service maintained.[5]

That gives the idea a practical purpose. A network can prepare for crowds while allowing surplus equipment to rest when they are absent. AI helps decide where and when that is possible; the meaningful result is an energy reduction that survives measurement across the chosen equipment, time period and service commitment.

Report an error or suggest a correction