fruitpunch

Guide · AI & data

AI for Wind Energy: Forecasting, Maintenance and Blade Inspection

By the FruitPunch research deskUpdated November 15, 20268 min read

In short

AI helps wind power in four main ways: forecasting output, predicting component failures from SCADA data, spotting blade damage in drone images and steering turbine wakes. Google reported that DeepMind's 36 hour forecasts raised the value of 700 MW of US wind by roughly 20 percent, and a Nature Energy field trial of wake steering raised farm energy by 1.2% across all wind speeds. The stakes keep growing: GWEC counts a record 165 GW of wind added in 2025 and 1,299 GW installed worldwide.

This address once hosted FruitPunch AI, an Eindhoven AI for good community. FruitPunch is a new, independent publication and is not that organization. This guide was written from public sources; the history page explains the change.

AI is used in wind energy to forecast how much power a farm will produce, to predict turbine faults before they happen, to find blade damage in drone photos and to coordinate turbines so they waste less wind on each other. The best known result came from Google, which in February 2019 reported that it had applied DeepMind machine learning to 700 megawatts of wind power capacity in the central United States, predicted output 36 hours ahead and said that "machine learning has boosted the value of our wind energy by roughly 20 percent." The value came from selling power in advance with more confidence, not from turbines spinning faster. The fleet these tools serve is large and still growing quickly, which is why small percentage gains add up. This guide covers each use with named studies, open datasets and free software.

Key facts

  • 165 GW: record new wind capacity added worldwide in 2025, up 40% on the previous record year (GWEC Global Wind Report 2026).
  • 1,299 GW: global installed wind power at the end of 2025, with 138 countries now generating from wind (GWEC).
  • 8%: wind's share of global electricity generation in 2024 (IEA Global Energy Review 2025).
  • Roughly 20%: the increase in the value of Google's wind energy credited to DeepMind forecasts (Google DeepMind).
  • 1.2% ± 0.4%: farm energy gain across all wind speeds from model based wake steering in a utility scale trial (Howland et al., Nature Energy, 2022).
  • 0.96 recall: crack detection by a production drone inspection model that processed more than a million damages (Iyer et al., arXiv 2207.11186).

How big is the wind power fleet that AI works on?

The Global Wind Energy Council reports that the sector installed a record 165 GW of new capacity in 2025, a 40% increase on the previous record year, bringing global installed wind power to 1,299 GW. The same report counts 28,395 wind turbines installed across 57 countries in the year. Each of those machines produces a stream of sensor data every few minutes, and each blade needs regular inspection, which is the raw material for the AI methods below.

In generation terms, the International Energy Agency puts wind at 8% of global electricity in 2024, after output grew by about 180 TWh that year. The IEA noted that the 8% annual growth rate was "the lowest in the last two decades" because of a high base and permitting challenges. In 2025, the IEA's Global Energy Review 2026 said lower than normal wind speeds, particularly in Europe, tempered renewable growth, while wind and solar together reached 17% of world electricity.

Looking ahead, the IEA's Renewables 2025 forecast expects cumulative onshore wind additions of 732 GW over 2025 to 2030 and offshore additions of 140 GW, though it cut its offshore forecast 27% from the previous year.

Indicator Figure Source
New wind capacity, 2025 165 GW (record) GWEC
Global installed wind, end 2025 1,299 GW GWEC
Countries generating wind power 138 GWEC
Wind share of world electricity, 2024 8% IEA
Wind generation growth, 2024 About 180 TWh IEA
Forecast onshore additions, 2025 to 2030 732 GW IEA
Forecast offshore additions, 2025 to 2030 140 GW IEA

How does AI forecast wind power output?

Wind farms are paid more when they can promise a fixed amount of power to the grid in advance. A forecasting model takes weather forecasts and the farm's own history of output, then predicts generation hours or days ahead. Google described training a neural network on widely available weather forecasts and historical turbine data and configuring it "to predict wind power output 36 hours ahead of actual generation." Based on those predictions, the system recommended hourly delivery commitments a day in advance.

Two caveats apply. The 20% figure is a value gain reported by the operator for its own fleet, not a peer reviewed accuracy score, and it depends on the electricity market where the power is sold. Forecasting also cannot remove variability: when the wind drops across a whole region, as the IEA described for Europe in 2025, a better forecast helps grid operators plan but does not produce the missing energy.

How does AI predict turbine failures before they happen?

Every modern turbine logs SCADA data (supervisory control and data acquisition): power, rotor speed, temperatures of the gearbox, generator and bearings, pitch angles and status codes, usually as 10 minute averages. Predictive maintenance models learn how a healthy turbine behaves under given wind conditions and raise an alert when readings drift away from that pattern, for example a bearing running hotter than expected for its load.

The hard part is evaluation, because real fault records are rare and operators rarely publish them. The open CARE to Compare dataset addresses this with 95 datasets "containing 89 years of SCADA time series distributed across 36 different wind turbines from the three wind farms." Of those 95 datasets, 45 contain a labeled anomaly event leading up to a turbine fault and 50 show normal behavior, so a model can be scored on both early warnings and false alarms. One of its three farms is an onshore site in Portugal based on data from the EDP open data platform, according to the dataset description; the other two are anonymized offshore farms in Germany.

How do drones and computer vision inspect turbine blades?

Blades take damage from lightning, rain erosion on the leading edge, cracks and manufacturing defects. Drones photograph each blade surface, and computer vision models sort thousands of images into damage types and severity so engineers only review the flagged ones.

Published results show both what works and where it is hard:

Image based inspection is not the only approach. Vibration based structural health monitoring looks for shifts in a blade's natural frequencies, shapes and damping that could signal hidden damage, a method that has to separate real damage from the effect of changing wind and temperature.

What is wake steering and does it raise output?

A turbine leaves a slower, turbulent wake behind it, so turbines downwind produce less. Wake steering deliberately turns (yaws) an upstream turbine slightly away from the wind to push its wake aside. The upstream machine loses a little power and the ones behind it gain more. Models, increasingly data assisted, decide the best angle for each wind direction and speed.

Field trials give real numbers:

A gain of about 1% sounds small, but it comes from a software change on existing turbines, so it applies across the whole operating life of a farm without new hardware.

Which projects, open datasets and tools can people use?

All of the resources below are free to download. Check each license, since some restrict commercial use.

Resource Type What it offers
Kelmarsh wind farm data SCADA dataset 10 minute SCADA and events data from 6 Senvion MM92 turbines in the UK, 2016 to mid 2021, released by Cubico under CC BY 4.0
Penmanshiel wind farm data SCADA dataset Same format for 14 Senvion MM82 turbines in the UK, 2016 to mid 2021
CARE to Compare Fault detection benchmark 95 labeled datasets from 36 turbines on three farms, with a scoring method for anomaly detection
DTU drone inspection images Image dataset Drone images from 2017 and 2018 of the same Nordtank turbine at DTU in Roskilde, Denmark, CC BY NC 3.0
FLORIS Software Python wind farm simulation for wake models and controls, developed since 2013 with U.S. Department of Energy support
OpenOA Software Python library for assessing wind plant performance from operational time series data

A practical starting project is to train a normal behavior model on Kelmarsh SCADA data, predicting gearbox or generator temperature from wind speed and power, then test whether its alerts match the fault events in CARE to Compare. For wake steering, FLORIS lets anyone simulate a farm layout and compare yaw strategies before touching real turbines.

What did the old FruitPunch AI wind energy page cover?

This URL once belonged to FruitPunch AI, an Eindhoven based AI for Good community that is not connected to the current FruitPunch site. Its archived AI for Wind Energy page from December 2021 described a challenge to identify blade damage from ambient vibration data, using data from DTU (Technical University of Denmark) and the startup Tarucca. The page said environmental and operational conditions cause 5 to 10% variations in blade frequency, which is what makes damage hard to separate from normal changes. This guide is an independent overview of the same topic.

Frequently asked questions

How much does AI improve wind power forecasting?

The most cited figure is Google's: DeepMind forecasts made 36 hours ahead raised the value of 700 MW of US wind by roughly 20 percent. That is a market value gain reported by the operator, so results elsewhere depend on local electricity markets and data quality.

Can AI make existing wind turbines produce more power?

Yes, modestly. Coordinated wake steering raised farm energy by 1.2% across all wind speeds and 3.0% at 6 to 8 m/s in a multi month field experiment published in Nature Energy in 2022.

Is drone blade inspection with AI accurate?

It can be within one fleet. One production model reached a recall of 0.96 on cracks, but BladeNet scored 0.995 average precision on Ørsted images and only 0.223 on DTU images, so models need testing on each new turbine type.

What data does predictive maintenance use?

Mostly SCADA data recorded by the turbine itself, typically 10 minute averages of power, speeds, temperatures and status codes. Public examples include the Kelmarsh and Penmanshiel datasets and the labeled CARE to Compare benchmark.

How much wind power is installed worldwide?

GWEC counts 1,299 GW at the end of 2025, after a record 165 GW was added that year. The IEA puts wind at 8% of world electricity in 2024.

Spotted an error in this guide? Email editorial@fruitpunch.ai

Keep reading