AI use case

AI for radio access network energy optimization

Machine learning that predicts traffic per cell and puts radio carriers, cells and hardware components into sleep modes when demand is low, then wakes them before users notice, so a mobile network uses less electricity without losing coverage or quality.

By Len Debets · Last verified 27 September 2026 · 5 public deployments

Up to 8%
Reported energy savings
Telefónica, organization claim.
USD 150,000 to USD 3 million
Indicative value per year
A mobile operator with 10,000 radio sites. Worked example, see how it is calculated.

What problem does it solve?

For operators such as BT Group and Orange, the network uses most of the energy. BT Group says its networks account for around 89 per cent of its total energy consumption, and Orange puts IT and networks at around 85% of its energy requirements. In a mobile network, part of that power keeps radio capacity switched on in periods when it is not needed, such as quiet nights: this is the capacity that cell sleep features switch off.

Vendors ship power saving features (carrier shutdown, micro sleep, deep sleep), but switching them on with fixed schedules leaves savings on the table in quiet cells and risks quality in busy ones. The settings differ per cell, traffic patterns change with events, holidays and new sites, and nobody can tune tens of thousands of cells by hand. Energy prices and net zero targets make it a cost and climate priority: Orange stepped up its energy saving measures during the 2022 energy crisis, and BT Group calls network energy efficiency integral to its net zero ambition.

How does it work?

  1. Learn each cell's rhythm. Models forecast traffic per cell and carrier from history, calendar effects and local events.
  2. Choose the saving action. For each forecast quiet period the system picks the deepest power saving mode that the cell can use safely: switching off capacity carriers, micro sleep, deep sleep of radio units or shutdown of idle components.
  3. Protect quality. Coverage layers stay on, neighbouring cells absorb the remaining traffic, and the system watches live load so sleeping capacity wakes within seconds if demand rises.
  4. Tune per cell. Thresholds are adjusted per cell from measured quality and savings, rather than one setting for the whole network.
  5. Report. Energy saved, quality indicators and wake up events are reported per site and cluster, so engineers can see where savings cost quality and adjust.
Audience
Back office
Autonomy
Autonomous
Adoption
Early adopters
Channels
API and system to system, Internal tools

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

Value benchmarks for AI for radio access network energy optimization
KPIMedianReported rangeData pointsClaimed by
Energy savingsToo few to pool
Not pooled: up to 8%
0plus 1 up to1 organization

Value drivers: Lower cost to serve.

Indicative value

A mobile operator with 10,000 radio sites

USD 150,000 to USD 3 million

Radio network electricity cost avoided per year

How this is calculated

Formula: sites * kwhPerSite * savingShare * pricePerKwh. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
Radio sites in scope sites, sites10,00010,000The reference operator.
Electricity use per radio site per year kwhPerSite, kWh per site per year30,00050,000Editorial assumption for a multi band macro site, including radio, power and cooling equipment. Replace with metered data.
Share of site electricity saved by AI driven sleep modes savingShare, fraction of site energy0.0050.024BT Group expects 4.5m kWh a year across EE's more than 19,500 sites, about 0.5 to 0.8% of the assumed site consumption estate wide, with a per site ceiling of up to 2 kWh a day (about 730 kWh a year, 1.5 to 2.4% of the assumed site consumption). The range spans that estate wide expectation up to the per site ceiling. Telefónica reports savings of up to 8% of a 5G site's 24 hour consumption, but in a single site test, and Nokia expects planned network energy cost savings of 8 to 10% for Safaricom; both are above this range and are not used to set it.
Electricity price pricePerKwh, USD per kWh0.10.25Editorial assumption. Replace with your own contracted price.

What it leaves out: Site electricity cost only. It leaves out carbon value, longer battery backup during grid outages, software licence and integration costs, and any quality impact that has to be compensated.

Who already uses it?

5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

BT Group

United Kingdom · Telecommunications · 2024

ScaledGrade B

After trials in each of the UK's home nations, BT Group rolled out cell sleep software to more than 19,500 EE mobile sites. It puts 4G capacity carriers to sleep during quiet periods that machine learning has predicted for each site, wakes them automatically at busy times and within seconds when traffic surges unexpectedly, and can use a deeper sleep mode overnight. The sleep functions come from the radio equipment suppliers; BT Group's site data drives the statistical algorithms that control them. BT Group published an expected annual energy saving rather than a measured result.

No outcome disclosed.

Telefónica

Spain · Telecommunications · 2022

ProductionGrade B

Telefónica has activated power saving features in its mobile networks for more than a decade. The early ones for 2G and 3G used static parameters; the current ones for 4G and 5G use AI and machine learning to predict traffic, set thresholds, shut down cells in low traffic hours and check quality. The platforms were first tested in O2 Germany in 2021. Telefónica Spain was the first operator to test Ericsson's Radio Deep Sleep Mode at a 5G site in Madrid, where the company reports savings of up to 8% of the site's 24 hour consumption and up to 26% in low traffic hours.

  • Energy savings: up to 8%, total 24 hour consumption of one 5G test site in Madrid, Radio Deep Sleep Mode
    "Supported by Artificial Intelligence and Machine Learning algorithms, the company achieved savings of up to 8%, considering the site’s total 24-hour consumption, and up to 26% in low traffic hours."
    Claimed by: organization

Indosat Ooredoo Hutchison

Indonesia · Telecommunications · 2025

ProductionGrade C

After a pilot in the live network, Indosat Ooredoo Hutchison deployed Nokia Energy Efficiency, part of Nokia's Autonomous Networks portfolio, across its entire Nokia radio access network footprint in Sumatra, Kalimantan, Central and East Java. The software uses AI and machine learning on real time traffic patterns to adjust or shut idle radio equipment during low demand and includes thermal management to cut cooling energy. It is multi vendor and delivered as a service. No measured savings are published.

No outcome disclosed.

O2 Telefónica Germany

Germany · Telecommunications · 2023

ProductionGrade C

O2 Telefónica Germany chose Nokia's AVA for Energy software, delivered as a service, for the parts of its radio network built on Nokia equipment. The software monitors traffic patterns and throttles back resources such as base stations during low usage, while monitoring quality so that customers do not notice the change. In its test the operator switched off unused radio resources automatically and saw significant savings, but no figure specific to O2 Telefónica Germany is published.

No outcome disclosed.

Safaricom

Kenya · Telecommunications · 2023

ScaledGrade C

After a pilot, Safaricom Kenya rolled out Nokia's AVA Energy Efficiency software across approximately 30,000 5G, 4G and 3G cells. The software uses AI and machine learning to switch off idle and unused equipment automatically during low usage periods, together with Nokia's radio energy efficiency features, while maintaining network quality. The release gives planned energy cost savings, not measured results.

No outcome disclosed.

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • Traffic and quality counters per cell and carrier at 15 minute or finer granularity
  • Site energy metering, ideally per site rather than estimated
  • Configuration and capability data for the power saving features per vendor and software release
  • Calendar of events and planned works

Systems to integrate

  • Radio network management and configuration systems per vendor
  • Performance management and network data platform
  • Energy metering and site management systems
  • Self organizing network (SON) platform where one exists

Complexity: Medium

The power saving features usually exist in the radio equipment already. The work is in reliable per cell traffic data, safe automation across vendors and convincing radio engineers that quality will hold.

  1. 1

    Measure the baseline

    Meter energy per site and record quality indicators before any change, so savings and quality impact can be proven rather than estimated.

  2. 2

    Switch on vendor features with guardrails

    Activate the available sleep features with conservative thresholds in a cluster, and compare with a control cluster.

  3. 3

    Add prediction and per cell tuning

    Replace fixed schedules with traffic forecasts and per cell thresholds, and let the system tune them from measured quality.

  4. 4

    Scale by cluster, not by country

    Roll out region by region with quality checks at each step, and keep special sites (hospitals, stadiums, transport hubs) under manual rules.

  5. 5

    Report savings finance can trust

    Agree the measurement method with finance and sustainability teams up front, so the savings count in budgets and emissions reporting.

Guardrails

  • Coverage layers and emergency service capability are never switched off
  • Automatic wake up on load thresholds, with a maximum wake up time per mode
  • Exclusion lists for critical sites and events, maintained by radio engineering
  • Quality key performance indicators monitored per cell, with automatic rollback when they degrade

KPIs to instrument

  • Energy saved per site and per cluster against a metered baseline or control group
  • Accessibility, retainability and throughput per cell during sleep periods
  • Number and duration of wake up events
  • Customer complaints about coverage in optimized areas

Human in the loop

Radio engineers set the guardrails, exclusion lists and quality thresholds, and review weekly reports of savings against quality per cluster. The system acts on its own within those limits, because sleep and wake decisions across thousands of cells happen too often to approve one by one.

Common failure modes

Savings that exist only in the model
Savings are estimated from switch off time rather than metered. Use metered energy and control clusters.
Quality loss at the edges
Neighbouring cells cannot absorb the traffic and users at the cell edge lose throughput. Monitor edge quality, not just averages.
Events the forecast did not know
A match, a concert or an emergency brings traffic to a sleeping area. Feed event calendars and keep fast wake up paths.
Vendor lock in of the optimizer
A rollout can end up covering only one vendor's part of the network (O2 Telefónica Germany uses Nokia's software on the Nokia part of its radio network, and Indosat Ooredoo Hutchison's rollout covers its Nokia radio footprint in four regions). Check multi vendor support and plan for a view across vendors where networks are mixed.

What are the risks and rules?

EU AI Act

Depends on design

Optimizing energy use is normally minimal risk. Under Article 6(2), Annex III point 2 lists AI systems intended as safety components in the management and operation of critical digital infrastructure as high risk, and public electronic communications networks fall under that infrastructure. Recital 55 limits safety components to systems that directly protect the infrastructure or the health and safety of persons, so an optimizer is not high risk by default, but a design in which it could affect emergency service availability should be assessed against point 2.

Guidance

Controls to put in place

  • Documented guardrails and exclusion lists with an accountable radio engineering owner
  • Change control for thresholds and new power saving modes
  • Monitoring of quality indicators with automatic rollback
  • Metered measurement method agreed with finance and sustainability reporting

Frequently asked questions

How much energy can AI save in a radio network?
The results on this page are site or trial level. Telefónica reports that Ericsson's Radio Deep Sleep Mode, supported by AI and machine learning, saved up to 8% of a 5G site's 24 hour consumption and up to 26% in low traffic hours in a Madrid test. The operators and vendors cited here give network wide figures only as expectations: Nokia expects its software to deliver planned network energy cost savings of 8 to 10% across about 30,000 Safaricom cells, and BT Group expects up to 2 kWh per site per day.
Is this different from the power saving features vendors already ship?
The features are the same; the difference is when and where they are used. BT Group puts capacity carriers to sleep based on quiet periods predicted for each site through machine learning, instead of one fixed schedule for the whole network.
Does it hurt network quality?
It should not if coverage layers stay on, capacity wakes within seconds and quality is monitored per cell with automatic rollback. BT Group says its sleeping carriers wake within seconds without interruption to customers, and Telefónica's platforms periodically review quality so as not to affect network performance or user experience.

How to cite this page

Blits.ai AI Use Case Library, "AI for radio access network energy optimization", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/ran-energy-optimization. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 27 September 2026: First published

Related use cases

Telecommunications

AI for mobile network planning and capacity optimization

Machine learning that forecasts where and when a mobile network will run out of capacity, recommends where to add cells, spectrum or hardware, and continuously tunes radio parameters so existing capacity carries more traffic, with planners approving investments and major changes.

Deployments
5 public, best grade B
Reported cycle time reduction
at least 95%
Deutsche Telekom, organization claim
Telecommunications

Agentic AI for autonomous, intent based network operations

AI agents that run closed loops over a telecom network: they take an intent from the operator (for example a latency or availability target for a service), observe the network, diagnose deviations and execute corrective actions across radio, transport and core, within guardrails set by engineers and with human approval for major changes.

Deployments
6 public, best grade B
Reported cycle time reduction
at least 95%
Deutsche Telekom, organization claim
Telecommunications

AI for predictive network maintenance in telecom

Machine learning that spots the early signs of network failure, such as degrading cells, faulty customer equipment, ageing hardware or planned digging near fibre, and triggers a preventive fix, a remote reset or a targeted intervention before customers lose service.

Deployments
6 public, best grade B
Autonomy
Supervised agent
Telecommunications

AI for telecom fraud detection (SIM swap, IRSF and Wangiri)

AI that protects the operator's own network, revenue and numbers from fraud: it watches call, messaging, roaming and account activity to detect SIM swap and port out takeovers, international revenue share fraud (IRSF) and Wangiri one ring scams, blocks or flags them in real time, and shares risk signals with banks and other businesses that rely on the phone number for security. Scam calls aimed at subscribers are handled by call blocking.

Deployments
4 public, best grade B
Reported detection improvement
30%
Vodafone, organization claim