Audit of a widely used wind farm dataset reveals hidden errors that made turbines look more productive than they really were
Since 2022, dozens of wind energy studies have drawn on the Kelmarsh open dataset — six turbines, eight years of records, freely available and widely trusted. Researchers routinely cleaned the data by filtering out implausible values, a standard step that seemed sufficient.
A new pre-registered audit suggests it wasn’t. Working through all 54 annual turbine files with nine reproducible checks, researchers identified four categories of structural defects — problems baked into how the files were organized and how measurements were defined, invisible to any conventional quality-control routine.
A dataset at the heart of wind energy research
The Kelmarsh dataset covers six turbines operating in the United Kingdom from 2016 to 2024. Since its public release in 2022, it’s appeared in several dozen published studies — a remarkable adoption rate for a single open dataset. The appeal is straightforward: real operational data, a long time span, freely available.
They also reported, with unusual transparency, which of their nine pre-specified predictions failed, partially failed, or couldn’t be evaluated as written.
Researchers using it typically validated the records by removing implausible values — negative power outputs, wind speeds outside physical bounds, and similar anomalies. That approach catches bad numbers. It can’t catch problems with how files are structured or how measurements are defined.
The new audit took a different approach entirely. All 54 annual turbine files were examined using nine pre-specified, reproducible checks, with hypotheses committed to before the results were examined. No prior study had applied anything close to that level of structural scrutiny.
Missing data that was never really missing
The first defect concerns when measurements actually begin. Most variables don’t start at the opening of the record — they’re introduced in one of twelve staggered steps spread between 2016 and 2023.
That staggered rollout has a concrete consequence. Calculating missing-data rates across a full year means counting the period before a sensor was introduced as a gap, when it was never a gap at all. The audit found this inflates whole-year missing-data rates by up to 73 percentage points. Studies that reported high rates of missing data in the Kelmarsh files may have been measuring a file-structure artifact, not actual sensor outages. The turbines weren’t offline. The measurements simply hadn’t started yet.
A sensor switch that inflated turbine performance
The second defect is subtler and potentially more consequential for energy modeling. Between November 2017 and April 2018, the primary wind speed measurement quietly changed. Before the switch, it represented the average of two sensors; afterward, a single sensor took over.
That change made turbines appear 1.2 to 9.1 % more productive at any given wind speed. The single sensor alone showed a variation of −0.6 to +1.5 %, meaning the blended effect of the switch — not just sensor-to-sensor variability — is the dominant source of the apparent performance gain. Power-curve estimates built across that boundary would absorb this shift without any warning. A turbine that looks more productive after 2018 might simply be measured differently.
Overlapping file segments that nearly erased two years of data
The third defect affects the most recent records and carries the steepest practical risk. In the 2023 and 2024 files, each year is stored across 81 overlapping segments, with measurements appearing only in the first copy of each time step.
Keeping the first copy recovers the full year. Keeping the last copy — a reasonable default in many data-processing pipelines — retains less than 1 % of the wind speed and power values. That’s not a minor data loss. Researchers who retained the wrong copy would overstate the annual capacity factor by 25 to 53 percentage points. A second wind farm operated by the same owner was found to share this identical file structure, suggesting the problem may not be isolated to Kelmarsh.
What the audit team is releasing — and why it matters
The researchers didn’t stop at identifying the defects. They published corrected files, a table flagging suspect values, and the complete analysis code — enough for any independent team to verify or extend the work.
They also reported, with unusual transparency, which of their nine pre-specified predictions failed, partially failed, or couldn’t be evaluated as written. One prediction failed outright. Seven others failed as originally worded. One couldn’t be assessed as specified. All are documented. That kind of candor is rare, and it models the open-science practices the broader research community has long advocated.
The findings point to a gap in how open SCADA datasets are currently maintained. Value-level filters — removing implausible numbers — are standard practice. Structural validation checks are not. Wind farm operators and dataset curators may need to add that second layer before the next generation of studies compounds errors the first generation unknowingly inherited. As more open datasets enter circulation and more research builds on them, the Kelmarsh audit offers a template worth watching: pre-registered, reproducible, and honest about its own limits.
You can learn more information through the complete study: Park, A., Yin, L., Lee, C., Wong, A., and Choi, Y.: Structural defects in the Kelmarsh open SCADA dataset: a provenance audit with file reconstruction and downstream consequences, Wind Energ. Sci. Discuss. [preprint], https://doi.org/10.5194/wes-2026-166, in review, 2026.
Daniel García is an Editor-in-Chief with strong expertise in structural work and engineering principles. He combines this technical foundation with deep knowledge of energy, spatial design, and emerging technologies, bringing a forward-thinking and analytical approach to editorial leadership.
