In the United States, the towers of new wind farms are often more than 90 meters tall, and the rotors themselves can reach another ~60 meters. The enormous height of these structures means that wind speeds high above the surface can affect wind generation. The ability to accurately characterize wind speeds at these heights is essential for estimating potential wind energy production, for effectively integrating wind farms into the power grid, for increasing grid reliability and resilience, and even for the design of new wind farms themselves.
It is challenging and expensive to measure wind speed at tower height (~90 meters), and publicly available data from tall towers or remote sensing are very limited. Therefore, researchers and practitioners commonly use meteorological models to characterize wind speeds at tower height and above. But how well do these models work? Can they accurately characterize wind profiles and what are their limitations?
A new study from Berkeley Lab and the University of California, San Diego evaluates three commonly used weather models for their ability to characterize hourly wind speed profiles (at tower height) across Texas. The study was published in the journal Wind Energy.
We use a novel approach to evaluate these weather models to overcome the lack of wind speed observations at tower heights: We transform the modeled wind speeds into wind generation using plant-specific characteristics and then evaluate the modeled wind generation against a unique data set of 7 years of hourly generation records from multiple than 100 wind farms in Texas (Figure 1 shows a map of all the farms used in this study). This comparison allows us to evaluate the ability of weather models to represent wind generation sources.
The analysis found that, by some metrics, these weather models performed well and accurately represented the hourly profiles of wind energy resources across Texas. However, the paper also identifies key areas where the accuracy of these models is limited. Note that our results mostly focus on two of the three models we evaluated (ERA5 and MERRA2) because we only evaluated the HRRR model in a subset of the observation period.
- The models present wind generation with relatively small bias and decent hourly correlation. Specifically, the average error across all hours and plants, expressed as capacity factor, was 1.1% for ERA5 and 6.4% for MERRA2 (meaning, for example, that a plant with a recorded capacity factor of 40% would be modeled with a 46.4% capacity factor using MERRA2). Average hourly correlation coefficients range from 0.7 to 0.9 depending on the model and time of day (see Figure 2).
- Errors were larger on an hourly basis than on a daily basis. For example, the probability that the modeled capacity factor from ERA5 was within ±20% of the observed capacity factor was 78% on an hourly basis, but 99% on a daily basis. Again, this is a capacity factor, so if, for example, a recorded capacity factor for a day was 80%, there was a 99% probability that the ERA5 model capacity factor was between 60% and 100% for that day and race. MERRA2 followed a similar pattern, but with slightly lower accuracy on both the daily and hourly time frames.
- Improved resolution of the geographic model reduced errors and improved correlation. ERA5 (horizontal resolution ~30 km) performed better than MERRA2 (horizontal resolution ~50 km) and HRRR (horizontal resolution 3 km) had the best performance metrics. Although the resolution varies dramatically between models, other methodological differences within the models likely contributed to the differences in performance.
- The correlation between the modeled and recorded generation decreased during the night hours, especially during the summer period.
The models were very good at representing the daily wind generation. Simply put, the models were almost always able to successfully distinguish between days of high wind generation and days of low wind generation (by the capacity factor prediction metric within ±20% of the recorded capacity factor). For individual hours, the models were often able to successfully distinguish between hours of strong and weak wind sources, but missed the capacity factor by >20%, roughly 20% to 30% of the time.
We hypothesize that the drop in correlation between modeled and recorded generation during the night was due to the incorrect representation of the atmospheric boundary layer in these models. At night, the boundary layer often shrinks to heights that would impinge on wind turbines. In addition, the ability of models to represent certain meteorological phenomena during the night, such as low-level jets, can further degrade model-observation agreement. Further research is needed to confirm these hypotheses.
Increasing the model resolution improves the correlation metrics overall, but does not solve the problem of reduced correlation during the night.
Practitioners using these weather models to represent wind generation profiles should be aware of their limitations. Further analysis would be beneficial to determine how well these models represent wind profiles in other regions.
Finally, we note that this analysis does not concern weather forecasts, but only meteorological models of real-time weather conditions. We expect the quality of error metrics to decrease with forecasts. Further prediction research is needed to determine their ability to represent hourly generation.
This approach is based on the assumption that there is much more uncertainty in the modeled wind speeds themselves than in the translation between wind speeds and power output. Our conversion of wind speed to generation was relatively simple, and while it involved the use of plant-specific power curves, it did not take into account plant control, local variations in wind speed, shed losses, and other types of losses. If wake losses were less consistent during the night than at other periods, it is possible that this contributed to the diurnal patterns we observed. This means that excitation losses are typically a small fraction of the annual generation (0 to 10%) and so are unlikely to have caused the type of observed changes in model performance. Additionally, our approach cannot account for losses related to plant maintenance or operational shutdowns due to weather (such as winter icing). However, the observed error patterns do not suggest that these operational problems play a major role. We note that if losses were included in the modeled generation, the models would likely have a small negative bias rather than a small positive bias.
Another key limitation is that this assessment applies to the Texas region, but may not necessarily apply to other regions, particularly regions with complex and/or mountainous terrain.
The article in Wind Energy, “Limitations of reanalysis data for wind energy applications” is “open access” and available to all: https://doi.org/10.1002/we.
If you have any questions, please contact Dev Millstein, [email protected], at Lawrence Berkeley National Laboratory.
We acknowledge the financial support of the US Department of Energy’s Office of Energy Efficiency and Renewable Energy.
Courtesy of Berkeley Lab, Electricity Markets & Policy
Do you appreciate CleanTechnica’s originality and reporting? Consider becoming a CleanTechnica member, supporter, technician or ambassador – or a patron on Patreon.