Wind Turbine Failure Prediction: Promising in Studies, Unproven in Fleet Operations

The use case

In a study of six wind farms in Italy and Romania, a neural-network model flagged a generator bearing problem on 26 April 2016. The operator detected it on 26 June, and the turbine then stood idle for 36 days (the authors' own account, without a control group)⁵. That is the promise of the use case: learn from routine operating data what normal looks like, and warn engineers weeks before a drivetrain part fails.

The data is usually SCADA (supervisory control and data acquisition): temperatures, speeds and power logged every ten minutes. Here, AI means machine-learning models such as autoencoders, LSTM networks (neural networks for time series) and random forests that learn from this data¹². Older technologies do similar work and are not credited as AI: fixed-threshold alarms, rule-based vibration monitoring hardware and scheduled inspections. Image-based blade inspection is out of scope.

Where it is running

No named commercial deployment with confirmed machine learning and published results was found. Confirmed machine-learning use appears in studies on real turbines, mostly with anonymised operators. Named commercial rollouts do not state that machine learning is used.

The bp and Longyuan announcements are older than three years; they show that digital monitoring programmes started in 2016 to 2019, but they report no outcomes.

What the evidence shows

Texas turbines. A US national laboratory tested four algorithms on 13 turbines of 1.5 MW with bearing cracks, labelling the last 30 days before failure as faulty. Adding modeled bearing-health data to SCADA cut false alarms by 50%, raised F1 (a combined precision and recall score) by 12% and raised AUC (overall ranking quality) by 6% for the LSTM model (retrospective, no control group, February 2021)¹. A conference paper on a Texas farm of 13 turbines of 1.5 MW reported best-model precision of 0.52, recall of 0.62 and F1 of 0.57, with false alarms falling from 105 to 52 and about 38% of faulty samples missed; only three turbines were used for testing (retrospective, date not shown)³. The mechanism is class imbalance: failures are rare, so models either miss them or raise many false alarms.

Offshore Europe. A preprint using four years of data from a large offshore farm reported unhealthy-to-healthy anomaly ratios of up to 1.69 for generator models and 1.21 for gearbox models, and showed that filtering could cut the volume of anomaly scores by up to 65% (preprint of 20 May 2025, using an operator failure list; not peer reviewed; no precision or recall)². The filtering matters because a flood of alerts is itself a cost.

Low precision on one machine. A study of one 3 MW direct-drive turbine over 11 months found detection precision of 0.02 to 0.04 with recall of 0.78 to 0.95 (single turbine, date not shown)⁴. Catching most faults came at the price of many false alerts.

Operator confirmation. In the Italian and Romanian study, a real-time run in 2017 produced 25 alerts (15 gearbox, 8 generator, 2 main bearing). About half received operator feedback, and over 90% of those were confirmed (authors' own trial, partial feedback, no precision or recall)⁵.

Vendor claims. A vendor case study states that 83% of gearbox failures were predicted 45 days ahead, with USD 35 million in savings (vendor-reported, no method shown). A trade publication repeated the 83% figure in May 2025 as a company statement that was not verified¹⁰¹¹.

What it takes: cost, integration, governance

The models need years of SCADA data plus reliable failure records. Reviews note that labelled failure datasets are limited and that models struggle with data imbalance and with turbines that differ from those they were trained on⁶. Public prices for such services were not found.

The stakes explain the interest. A 2026 review reported that gearboxes and bearings account for roughly 30 to 40% of turbine downtime (reported field data) and that operation and maintenance is about 10 to 30% of the energy cost of an operating offshore farm⁶. A 2026 review citing IRENA (2024) gives a similar range for onshore wind in 2023 (secondary citation)⁸.

On rules, the EU AI Act lists AI used as a safety component in the operation of critical infrastructure, including electricity, as high-risk. Commentary says many energy systems lacking a direct safety function are not, and that a proposed postponement would delay Annex III obligations to 2 December 2027¹⁴¹⁵. Neither source classifies turbine-component maintenance specifically.

Where it falls short

The main question is what would make the use case look less successful than it is presented. The answer is the design of the evidence. Results come from retrospective tests on a handful of turbines (13 in the US work, one in the Irish study), from operators' failure lists, or from the authors' own trials. No independent field evaluation of avoided downtime or maintenance cost was found. Reviews report overreliance on simulated datasets and field reliability limited by data imbalance, domain shift and few labelled failures⁶⁷, and describe academic prognostics as not yet widely adopted in the field⁸.

Effects are uneven. One study missed about 38% of faulty samples³; another produced precision as low as 0.02⁴. Benefits also depend on engineers who triage alerts and act on them, so the model's contribution cannot be separated from the process around it⁵.

No paused programme, withdrawal, lawsuit or regulator action was found for this use case; the search was limited and this is an absence of evidence, not proof of none. The evidence is concentrated in Europe and the United States; Chinese and Indian rollouts are announced without results⁹¹².

Phase: Pilot with results — why

The hypothesis was In wide use. It is not supported: machine-learning use is confirmed only in studies and trials on real turbines, the largest covering 150 turbines⁵, and the named commercial rollouts do not state that machine learning is used and report no results¹²¹³. The lower phase is chosen.

What to watch next

• Peer-review outcome of the offshore autoencoder preprint, which started discussion on 20 May 2025².
• 2 December 2027, the proposed postponed date for Annex III high-risk obligations under the EU AI Act¹⁵.
• The first published detection or loss results from the Aviva and ONYX Insight insurance product announced on 15 July 2026⁹.
• Any operator-published field evaluation reporting avoided downtime, which would end the current reliance on retrospective tests.

Questions to ask before adopting

• Was the model tested on turbines and years it was not trained on, and how many confirmed failures did the test contain?
• What are precision and recall on your turbine models, and how many false alerts would engineers face each month?
• Does the service use machine learning, or fixed thresholds and rules, and can the supplier show the difference in results?
• How much warning does the model give, and can crews, parts and vessels be mobilised within that time?
• Who owns the failure data and the model, and what happens to both if the contract ends?