Improved Characterization of the Magnitude and Causes of Spatio-Temporal Variability in Wind Resources

The feasibility of wind energy installations at given locations is dictated in large part by the wind resource (‘power in the wind’). In 2014 average capital expenditures for wind project development in the US was ~1.7 million USD per MW of installed capacity, a decrease of nearly one-third from the peak in 2009-2010, but financing considerations continue to play a key role in determining whether individual projects will be realized.

Variability of, and uncertainty in, the wind resource and thus the range (i.e. worse to best case dispersion) of likely annual electrical power production, increase project risk and financing costs. Two specific metrics are used to quantify the viability of wind projects in terms of annual energy production (AEP) (i.e. the expected amount of electricity generated in each year):

• P50: AEP projected to be equaled/exceeded on 50% of years during wind farm operation.
• P90: AEP that is associated with a 10% risk of not being reached.

This research is designed to improve quantification of the wind resource and variability therein (i.e. P50 and P90 AEP) and drivers of that variability. Our research domain is centered over the eastern two-thirds of North America and thus encompasses a substantial fraction of the total wind resource and wind turbine installed capacity (nearly 50 GW by the end of 2016).

We are addressing three linked key research questions:

1. To what degree do reanalysis products capture the spatial patterns of near-surface wind speeds and both the magnitude and dynamic causes of the intra-annual (seasonal to daily variability) and inter-annual (year-to-year, decade to decade) variability of wind climates?
2. How spatially coherent are periods of high wind speeds (and wind speeds below power producing levels) and to what degree are high-resolution (12 km and 4 km) simulations with WRF adding-value by improving estimates of wind resource magnitude and variability?
3. To what degree are high-resolution simulations of wind resources dependent on the precise computational platform used to undertake those simulations, and does that uncertainty and overall simulation uncertainty significantly impact P50 and P90 AEP?

Our research analyzes a triad of wind speed time series in order to address these research questions:

(a) 5-minute data from over 500 sonic anemometers deployed as part of the National Weather Service Automated Surface Observing System. These data are resampled at either 10-minute or hourly time steps and used to evaluate the reanalysis output and WRF simulations, in addition to being included in the spatial coherence analysis (research questions 1 and 2).

(b) Hourly wind speed data from the NCEP-NCAR (NNR) and MERRA-2 reanalysis products are used to characterize spatio-temporal variability of near-surface wind speeds (research questions 1 and 2). NNR is typical of conventional global reanalysis products while MERRA-2 also assimilates an unprecedented array of remote sensing data streams, and is available at a resolution of 0.625°´~0.5° (longitude-by-latitude) versus NNR which is relatively low resolution (2.5°´2.5°).

(c) 10-minute output from multi-year WRF simulations conducted at 12 km and 4 km spatial resolution using WRF on two computational platforms (a Cray and an academic cloud) are used to address research questions 2 and 3.

Initial results of our research indicate:

1. MERRA-2 exhibits relatively good agreement with ASOS observations of 10-m wind speeds, particularly given the challenges in comparing low resolution model grid versus point values (i.e. the inherent ‘differences’ incurred due to different resolutions) and in the 90^th percentile wind speed values that most closely represent the ‘power in the wind’ that can be captured by wind turbines. The data further indicate that not only does the northeastern USA (currently a region of known high wind resources but low wind energy penetration) have high near-surface 90^th percentile wind speeds, but exhibit lower differences between AEP P50 and AEP P90 than characterize parts of the country that have experienced higher wind energy penetration to date.
2. Diagnostic analyses designed to attribute intra-annual to inter-annual wind climates in MERRA-2 to climate modes is ongoing. NNR indicates that the phase of individual climate modes (ENSO, AO and PNA) all significantly contribute to variations in 90^th percentile wind speeds over much of the eastern 2/3 of North America (of a magnitude of the order of +/-10%). However, mode interactions are non-negligible, and indeed that the magnitude and even sign of the wind climate response to the PNA phase differs when the phase of the AO is also considered.
3. WRF simulations at 4 km are skillful in reproducing the spatial variability in AEP P50 and P90. There is value-added in simulations at convection permitting scales (4 km) relative to 12 km in terms of the improved skill in the absolute magnitude of the wind resource but the difference in the inter-annual variability of wind speeds (and hence P50 and P90 AEP) between the 12 km and 4 km runs is very modest.
4. Time periods with 90^th percentile wind speeds above/below normal are coherent at relatively large spatial scales (in the ASOS observations and WRF simulations), but periods with large-scale (supra-regional) calming of wind speeds are infrequent and short-lived in the regions of highest wind resource.
5. WRF is a stable and widely used numerical model and great effort has been expended in ensuring it represents both the state-of-the-art in atmospheric science and can be compiled and run on a wide array of computing system architectures. However, we show that in a simulation of an entire calendar year (2008) there are non-trivial differences (persistent biases) in simulation of 10-m and hub-height wind components and wind speeds across the two computational platforms (the DoE Cray XC40 computer named Cori and the academic cloud system built on Dell components referred to as Aristotle). Consistent with our expectations based on the way in which the domain is discretized for MPI applications and disparities in the compilers, these differences in simulated wind climates are most evident in regions within the domain that exhibit land-surface discontinuities. They are of sufficient magnitude to be a non-trivial component of the overall uncertainty in wind resource estimations and thus AEP P50 and P90.