Structural defects in the Kelmarsh open SCADA dataset: a provenance audit with file reconstruction and downstream consequences
Abstract. The Kelmarsh wind farm dataset, covering six turbines in the United Kingdom from 2016 to 2024, has been used in several dozen wind energy studies since 2022. The studies we read that use it check its quality by removing implausible values, which cannot reveal errors in how the files are organised or how measurements are defined. We examined all 54 annual turbine files for such errors with nine reproducible checks, fixing our hypotheses in advance, and found four kinds. First, most measurements start partway through the record, in one of twelve steps between 2016 and 2023, so whole-year missing-data rates overstate real gaps by up to 73 percentage points. Second, between November 2017 and April 2018 the main wind speed measurement switches from the average of two sensors to a single sensor, making the turbines appear 1.2 to 9.1 % more productive at a given wind speed; the single sensor alone shows a change of −0.6 to +1.5 %. Third, the 2023 and 2024 files store each year in 81 overlapping segments with measurements only in the first copy of each time step, leaving 97.6 % of rows without wind speed or power values. Keeping the first copy recovers the year; keeping the last retains less than 1 % of the data and overstates the annual capacity factor (output as a share of the maximum possible) by 25 to 53 percentage points. Files tested from a second wind farm by the same owner share this structure. Fourth, the energy, voltage, wind speed and temperature records contain smaller inconsistencies and placeholder values. One pre-specified prediction failed, seven others failed as originally worded and one could not be evaluated as specified; all are reported. We provide corrected files, a table flagging suspect values, and the analysis code.
The paper checks the Kelmarsh wind farm dataset for errors in how its files are built. It finds four problems: signals that start on different dates, a wind speed channel that changes meaning in 2018, repeated rows in the 2023 and 2024 files, and some invalid values. The work is careful, and the code and corrected files are shared. Three points need work before publication.
1. Separate the wind speed effect from real changes
The paper states that the change in the wind speed channel makes the turbines look 1.2 to 9.1 % more productive after 2018. This number mixes three effects: the channel change itself, a real change in the turbines, and a 2017 baseline whose definition is partly unclear. The authors' own breakdown isolates the channel effect at 28 to 88 kW, and this is the number that should appear in the abstract. The analysis should also use more than one wind speed bin and include an air density correction.
2. Explain a contradiction in the wind speed data
The paper states that the main wind speed channel equals sensor 2 from May 2018. In 2023, however, sensor 2 has 1,901 negative readings, including the error code −15 m/s, while the main channel has none (Table 7). If the two channels were equal, these values would appear in both. The authors should explain how error codes were handled in the comparison. They should also justify why sensor 2 is recommended over sensor 1, since nothing in the paper shows which sensor is correct.
3. Make the paper easier to read
The paper reports every pre-specified prediction, including those that failed. This is commendable, but the bookkeeping takes a lot of space for very small misses, such as 88.9 % against a 90 % target. As a result, the four main findings get lost and the abstract is crowded with numbers. Moving the prediction tracking (Sect. 4.6 and Table A1) to an appendix and limiting the abstract to the four findings would make the paper much clearer.