Evaluating the Impact of Inter-Annual Variability on Long-Term Wind Speed Predictions
Abstract. Assessing the wind resource and its associated uncertainties is essential for the profitability of a wind farm, with inter-annual variability in wind speed being a key factor. To estimate the wind resource at a potential wind farm site, a year-long wind measurement campaign is typically conducted and combined with long-term – often numerical – reference data using the Measure-Correlate-Predict (MCP) approach. This process accounts for systematic errors in the reference data and captures the long-term wind variability of wind speed. Since wind conditions vary from year to year, the selection of a single measurement year within the MCP framework can significantly influence the predicted wind resource. In this study, we systematically evaluate the impact of the measurement year on wind speed predictions using long-term met mast measurements. We also investigate whether classical and advanced machine learning methods can mitigate this sensitivity. Our results reveal that the variation in predicted wind speed due to the chosen measurement year ranges from 1 % to 14 %, depending on the site and correlation method, with an average of 6.5 %. Excluding years with exceptional wind conditions reduces the mean to 4.2 %. Among the methods selected, the correlation method SpeedSort, along with the advanced machine learning models Random Forest and AdaBoost, most effectively mitigates the influence of inter-annual wind variations in long-term referencing compared to classic linear regression. Additionally, the findings indicate that AdaBoost and Random Forest are especially beneficial for sites with heterogeneous and complex terrain. Furthermore, the study highlights the need for quality-controlled, long-term datasets across a variety of sites with differing terrain complexities to better understand and manage the effects of inter-annual wind variability in diverse wind climates.
General comments:
This paper makes a good contribution to the literature on evaluating long-term variability in the wind resource, an ongoing challenge for wind energy science. The authors make good use of a set of long-term (by wind energy standards) tall-tower data sets to compare several methods to estimate multidecadal wind speed variability from much-less-than-multidecadal in-situ observations. They apply several MCP methods and three widely used ML methods to data from sites in different geographical and climatological settings. The ideal outcome would be to determine the best method(s) to use given a particular geographic or climatological setting, but given the complexity of the wind resource, the results are a little more nuanced than that, and understandably so. Nonetheless, the authors describe how the methods compare to each other across these settings and are able to offer some tentative conclusions and recommendations for estimating long-term variability from short-term records.
Specific comments:
(1) Lines 100-105: The ERA5 data set begins in 1940 but your analysis begins in 1950. Given your goal of characterizing long-term variability, why exclude this additional 10 years of reanalysis data?
(2) Table 2: Are all the sites freestanding met towers or are they towers in the vicinity of a wind turbine, which might possibly be affected by wakes from some wind directions?
(3) Line 210: The headings in Table 4 list 2010-2016 and 2012-2016, but the text says the MER is for 1950-2020. I suggest adding that bit of info to the table heading, just as a reminder to readers.
(4) Lines 220-227: I was confused at this point about why you would want to reduce the influence of interannual variability when creating a long-term reference data set. You talk about this a bit later in the paper, but maybe add a note here that you’ll come back to this in section 3.4.1? In general, I think it might be useful to say a little more in the paper about the importance of long-term “data” such as reanalyses, which in theory can include ENSO and other climate patterns that influence long-term variability at a site.
(5) Line 240: You note here, and in several other places in the paper, that the MNER doesn’t seem to depend much on terrain complexity. Do you have any hypotheses as to why? I encourage you to include a bit more discussion of this - even if only possible hypotheses - in the Discussion and Conclusions section (e.g., lines 380-383). Perhaps that discussion can also touch on your comments about sensitivity to sample size and data gaps (lines 259-263).
Technical comments:
(6) Figure 9: The figure caption should say 2012-2016, not 2010-2016.
(7) Figure 10: The figure caption shows 2010-2016 (MNER2010-2016 %) twice; the second one should be 2012-2016.