1. Introduction
Historical time series of wind speeds and associated wind resources and power production in regions with high wind energy penetration exhibit strong evidence for variability linked to internal climate modes and changes in land use (i.e., surface roughness) (Pryor et al. 2020c; Wohland et al. 2021). As described in Pryor et al. (2023, hereinafter Part I) and Pryor et al. (2020c), a number of previous studies have used direct wind speed output from global (low resolution) Earth system models (ESMs) to make projections of possible future wind resources. However, most archives of ESM output include only daily or monthly mean wind speeds at 10 m above ground level, far below typical wind turbine hub heights (Pryor et al. 2020a). Further, the resolution of such models prohibits representation of the scales of motion critical to dictating wind resources. To overcome these shortcomings, either statistical or dynamical downscaling is used to obtain more accurate local-scale projections. The former is the focus of this paper, whereas Part I focuses on application of dynamical downscaling.
Statistical downscaling is the process of using statistical transfer functions to link large-scale predictors drawn from ESMs to the local target variable of interest (predictands) using statistical relationships (transfer functions) (Ekström et al. 2015; Lanzante et al. 2018). Examples of the statistical functions used include regression models (Huang et al. 2015; Hutengs and Vohland 2016), principal component analysis (Benestad et al. 2015; Davy et al. 2010), stochastic weather generators (Raynaud et al. 2020), synoptic classification schemes (typically using cluster analysis) (Bermúdez et al. 2020), and artificial neural networks (Pryor and Schoof 2019).
Most previous studies that have sought to address possible changes in wind power production have focused on dynamical downscaling the wind climate (Pryor et al. 2020c). Comparatively few statistical downscaling analyses using wind resources as the predictand have been performed for North America. Tree-structured regression was applied to monthly mean winds speeds at 10 m for five sites in the western United States to downscale CMIP3-generation ESMs and found that summertime monthly mean wind speeds in the Northwest declined by 5%–10% whereas small magnitude increases were projected in the winter (Sailor et al. 2008). Pryor and Barthelmie (2014) use probabilistic approaches where dynamically downscaled descriptors of the surface pressure gradients and upper-level vorticity are linked to the moments of wind speed Weibull distributions. They found that mean and 90th-percentile wind speeds over eastern North America exhibited modest changes (i.e., under 5%; less than the downscaling uncertainty) up to the mid-twenty-first century (2040–60).
As described in Part I, the current work is innovative relative to previous research in that the model predictand is the capacity factor (CF) derived from daily expected power production. Capacity factors are the ratio of the amount of electrical power produced normalized by the potential power produced if all wind turbines at a given facility run at their rated capacity. Gross CF derived from daily expected power differs from net CF from operating wind farms that also include power losses due to downtime for maintenance and operations and curtailment for grid integration (Lee and Fields 2021). Use of CF derived from expected power at operating wind farms as the predictand has the advantage that it is a variable with direct industry application, allows the data to be readily anonymized to avoid disclosure of commercially sensitive information, and avoids the need to postprocess downscaled wind speeds using scaling functions or an idealized wind power curve (Pryor et al. 2020c). One disadvantage is that expected power from multiple locations does not necessarily conform to a single parametric distribution (Gaussian, Weibull, Nakagami, etc.), which limits use of probabilistic approaches like those used for wind speed distributions (Pryor and Barthelmie 2014; Pryor et al. 2005). Here we employ a hybrid statistical downscaling method using synoptic classification with variance enhancement based on pressure gradient magnitude.
A synoptic-typing approach is selected for this analysis in part because of evidence that the CMIP6 ESM exhibits substantially improved fidelity with respect to the synoptic climate relative to previous CMIP-generation ESMs (Cannon 2020). Synoptic-typing methods assume different spatial patterns of predictors are associated with different values of predictands (P. G. Dixon et al. 2016), such as regional wind power density (Coburn 2021; Gibson and Cullen 2015). They have been applied in previous analyses of wind resource projections under climate nonstationarity. Millstein et al. (2019) assessed the synoptic situations responsible for high and low wind energy generation for sites in California over the historical period (1980–2015) and their links to interdecadal warming. Kirchner-Bossi et al. (2015) used a modified Lamb classification scheme to downscale wind speeds for two sites in Spain over the period 1880–2014 using the twentieth century reanalysis. Their results illustrate that low and high wind resource (power production periods) vary consistently with synoptic regime that in turn are linked to large-scale dynamics and modes of internal climate variability such as El Niño–Southern Oscillation (ENSO) and the North Atlantic Oscillation (NAO).
Statistical downscaling frequently employs a perfect-prognosis framework, in which the transfer functions linking the observed predictand (in this case, daily power production) to the predictors are derived using reanalysis products and applied to output for the predictors from the ESMs under the assumption that the ESMs reproduce the predictors with high fidelity (Pryor and Schoof 2020). It is also assumed that the relationships linking the large-scale predictors to the local-scale targets remain constant over time and under different climate states, also called stationarity (K. W. Dixon et al. 2016). Thus, another critical assumption in work that employs synoptic typing is that any evolution of the synoptic-scale climate does not involve generation of new synoptic classes that are not present (or not abundant) in the training period (Van Uytven et al. 2020). Both assumptions are tested herein.
The objectives of the current work are as follows:
-
Develop a novel hybrid statistical downscaling method and apply it to quantify possible changes in future wind power production (as measured using CF) at sites across North America using an ensemble of CMIP6 ESM realizations for different Shared Socioeconomic Pathways (SSPs).
-
Analyze the resulting ensemble of daily CF projections to examine the presence of evidence of (i) secular trends in annual mean CF, (ii) changes in median (P50) annual mean CF, and (iii) tendency toward increased frequency of wind “drought” or “bonus” periods and to attribute any differences to changes in the synoptic-scale climate.
-
Examine the extent and timing of the emergence of forced trends (i.e., signals) relative to internal variability (i.e., noise) in the climate system (Hawkins and Sutton 2009; Hawkins et al. 2020).
2. Data
a. Expected power
Analyses presented herein focus on daily expected power production at 22 operating wind farm sites across North America for the period 2015–20. As discussed in Part I, these data represent the power produced by the wind farm each day under the assumption of no curtailment or operations/maintenance stoppages. They are provided by a wind farm owner–operator on the explicit proviso that they be fully anonymized by conversion to CF and in terms of the precise locations. While many of these wind farms are common to the analysis presented in Part I, the study domain considered here is larger, allowing for inclusion of wind farms outside the simulation domain considered in Part I, but an additional criterion is applied to selection of sites. That is, only wind farms with more than 3 years of daily expected power estimates are included in the statistical downscaling to ensure the downscaling algorithms have sufficient training data to condition robust transfer functions. The wind farms considered herein are referred to by the region in which they are located: Northeast (NE; magenta, 6 sites), northern Great Plains (NGP; green, 2 sites), Midwest (MW; blue, 3 sites), southern Great Plains (SGP; red, 7 sites), southern West Coast (SWC; cyan, 3 sites), and northern West Coast (NWC; orange, 1 site) (Fig. 1). Daily expected power values at each operating wind farm are converted to CFs using information on the total installed capacity of the wind farm and the rated capacity of the wind turbines. Daily CF values below 0 are set to NaN (324 days across all sites, or 0.7% of all daily readings), and values greater than the installed capacity (>1) are set to the installed capacity (<100 daily values across all sites, all due to rounding errors). The resulting daily CFs are the predictand in the statistical downscaling.
Map of study domain and the regions from which data are presented. The number of wind farms from which data are presented in each region is shown by the numbers, and the color coding for each region shown is used throughout: magenta for NE, blue for MW, green for NGP, red for SGP, cyan for SWC, and orange for NWC. The background shading shows the 100-m mean wind speed from ERA5.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
To aid in the presentation and interpretation of the results, six exemplar locations are used to illustrate details of the method and results. They are representative of the regions from which they are drawn and are referred to using the regional abbreviations shown above and a digit. For example, NE 3 is the third wind farm in the Northeast region. One wind farm each from the NE, NGP, NWC, and SWC regions is included along with two from the SGP that illustrate differences according to distance from the coast; SGP 5 is far inland and is also representative of behavior of wind farms in the MW region.
b. ERA5
The European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis dataset is used as a source of predictors for use in transfer function training and testing. ERA5 is produced by assimilating a large suite of in situ and remote sensing observations, with atmospheric variables output at high spatial (0.28° × 0.28°) and temporal (hourly) resolution (Hersbach et al. 2020). ERA5 has been subject to a wide array of validation analyses. It has been reported to exhibit relatively high fidelity for near-surface pressure and temperature fields (Bell et al. 2021), as well as for extratropical cyclone tracks and characteristics (Gramcianinov et al. 2020). Further, ERA5 has been shown to produce optimal results in downscaling applications using circulation analogs as compared with other reanalyses, especially when the outputs are averaged to resolutions that typify ESMs (Horton 2021). Once hourly output for each of the downscaling predictors from ERA5 are averaged to derived daily mean values of each predictor in each grid cell over a 6° × 6° (600–700 km, 25 × 25 grid cells) area centered on each wind farm.
Sea surface temperature and sea level pressure from ERA5 are also used compute monthly indices of the Pacific decadal oscillation (PDO) and the Northern Annular Mode (NAM), following Coburn and Pryor (2021). The results are used to contextualize projected CF changes in a historical period of very low wind power production that impacted wind farms in the California independent system operators (ISOs) and the Electric Reliability Council of Texas (ERCOT) and higher-than-normal production in ISOs over the Northeast (Pryor et al. 2020b). Simultaneous reinforcement of the positive phase of the PDO and NAM is thought to have led to persistent and intense ridging over the West Coast that was a major cause of the 2015 first-quarter (Q1) wind drought (Lledó et al. 2018; Pryor et al. 2020b).
c. ESMs
Output from four ESMs (Table 1) from phase 6 of the Climate Model Intercomparison Project (CMIP6; the latest phase) archive are extracted for the same 6° × 6° area around each wind farm for the historical period (1979–2014) and the twenty-first century projections (2015–2100) under four different SSPs; SSP126 (sustainability track with low emissions), SSP245 (middle-of-the-road emissions), SSP370 (regional rivalries and higher emissions), and SSP585 (rapid fossil fuel development and very high emissions), with increased net radiative forcings of 2.6, 4.5, 7.0, and 8.5 W m−2 by 2100 (O’Neill et al. 2017). These ESMs span a range of different equilibrium climate sensitivities (ECS) (see Table 1). They are selected to (i) analyze multiple realizations from each ESM to assess internal variability in model outcomes, (ii) sample models that span the range of ECS in CMIP6, and (iii) sample models that represent different ESM “families” [i.e., to sample independent models and avoid those with significant dependencies such as shared components or origins (Brunner et al. 2020)] and also because of the necessity to only include models that archive the predictors at a daily time step (Table 1). Between 3 and 10 realizations of each scenario for each ESM are used, for a total of 23 model runs over the historical period and each SSP, yielding 115 ESM-downscaled CF time series.
A summary of the ESMs used in this study. From left to right, the model names, source institutions, atmospheric component resolutions (label “Atmos res,” reported as the number of longitude, latitude, and vertical grid cells), ECS, and references are listed. The boldface portions of the model names correspond to the shortened names used in the text. The countries of origin are italicized after the source institutions. The spectral resolutions of the first three models are T85, T85, and T63, respectively. The ECS are given in degrees Celsius per 2 × CO2.
3. Methods
a. Downscaling
The analysis workflow is shown schematically in Fig. 2. First, a suite of 15 predictors from ERA5 are evaluated in terms of their ability to generate clear, distinct synoptic types that differentiate high and low CF days. The predictors tested are (i) sea level pressure (SLP), (ii) geopotential height Z and wind vector components U and V, plus wind speeds W, air temperature T, and specific humidity Q at 500 and 700 hPa, and (iii) the lapse rate dT and atmospheric thickness dZ calculated using T and Z at 500 and 700 hPa. In this process, daily values from the 15 ERA5 predictors from 1979 to 2020 are used to derive independent synoptic typing based on k-means clustering for each individual wind farm site. The degree to which the different clusters discriminate varying CF is tested and used to select both the best predictor for the synoptic typing and the optimal number of clusters. Since each ESM has a different grid resolution, ERA5 output is regridded to match the grid spacing for each ESM {i.e., regridded from 25 × 25 to 5 × 5 [CNRM and MIROC6], 3 × 3 [MPI], and 3 × 5 [U.K. Earth System Model (UKESM)] cells}, and the synoptic typing is repeated for each set of grid dimensions. This regridding is necessary to avoid features being identified in synoptic classes derived using the ERA5 that are subgrid scale (and thus unresolved) in the ESMs.
A summary of the workflow for this study.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
Many methods for deriving synoptic types are used throughout the literature, including predefined algorithms such as the Lamb classification scheme (Jones et al. 2013) and more complex methods such as self-organizing maps (SOM) (Berkovic 2018). Synoptic typing has been widely applied in atmospheric science, and k-means clustering has generally been found to be one of the better-performing techniques (Broderick and Fealy 2015). Initial tests of other methods for defining the synoptic types yielded similar results to those derived using k-means clustering. For example, consistent with past work that showed k-means clustering is more robust than the Lamb scheme (Coburn 2021), k-means clustering generated synoptic types that more clearly differentiated days with high versus low CF. Use of SOM produced similar results in terms of differentiation of CFs by synoptic type but at the cost of increased complexity.
The k-means clustering method is an unsupervised classification approach in which a sample of n predictor fields is separated into k clusters, or groups (Zscheischler et al. 2012). Initial nodes (i.e., centroids of each cluster) and the number of groups (the number of synoptic types) have to be defined a priori. The former means that, for each predictor, the initial spatial patterns with which all daily predictor patterns are first compared must be supplied. As each daily field is added, the clusters are redefined to maximize internal similarity within each cluster while maximizing external differences across clusters using Euclidean distance as the similarity metric. A given daily predictor pattern may begin as part of one group but end up in another, similar group by the conclusion of the process. The end product is k clusters that each have a membership of nk days (where the sum of nk is the total number of days of data). The spatial pattern representing each cluster may be deduced by either finding the mean of all members or by taking the daily pattern closest to the cluster centroid as representative of that cluster. The k-means clustering method is computationally efficient, but the synoptic typing derived using k-means clustering naturally exhibits a dependence on k and the initial nodes (Wang and Su 2011). Sensitivity analyses are performed here by comparing results for k = 5–25 and by bootstrapping 1000 times for different randomly selected initial nodes. The use of two predictors in the synoptic typing did not yield improvements in terms of discriminating different daily CFs and thus is not pursued.
Clusters that exhibit PF, EV, and S approaching 1 are defined as optimal. Based on results from analyses of PF, EV, and S, the following methodological selections are made. SLP is selected as the variable for defining the synoptic clusters. Clustering using SLP patterns yields higher EV and PF than most or all other predictors in analyses for each of the wind farm locations (see examples in Fig. 3). It is also the only predictor to result in positive S values at all sites, indicating better classification and differentiability. To balance the need for explanatory power with pattern separability for physical interpretation, k = 15 is used. While increasing the number of synoptic types (clusters) increases EV (Fig. 3; and PF, not shown), much of the increase in these metrics occurs at lower k values. Further, increasing the number of clusters leads to a decrease in the between type differentiation, decreasing S. Selection of the initial seed days has only a very modest impact on the resulting synoptic clusters.
The (left) explained variance and (right) silhouette scores for each predictor over the number of clusters (5–25) for sites (a) NE 3, (b) SGP 1, and (c) SGP 4. The predictors are shown by colored lines, with SLP as a thick red line. Vertical lines denote the 15-cluster synoptic typing used herein.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
b. Evaluation
Confidence in CF projections is determined by quantifying historical fidelity, and differential credibility between projections from different ESMs is assigned based on the relative ability of that ESM to reproduce the frequency of the different synoptic types in the historical climate, as well as the associated pressure gradients. This fidelity assessment is applied prior to making projections of daily CFFINAL. A chi-squared χ2 test (Wilks 2011) is applied to identify differences in the relative frequency of the synoptic types in the historical (1979–2014) and contemporary (2015–20) climate in ERA5. In these analyses a confidence level of 95% (i.e., p value of 0.05) is used. ESM realizations are evaluated relative to ERA5 in terms of their ability to correctly (i) capture the frequency of the different synoptic types during 1979–2014, (ii) reproduce pressure gradients around each wind farm, and (iii) reproduce the cumulative distribution functions (CDFs) and seasonality of the ERA5-downscaled daily CF values and the CFOBS from 2015 to 2020.
c. Projections
Because the downscaled projections from each ESM for a given SSP are consistent across the different realizations, in the majority of the analyses we present only the mean value of downscaled CFs from realizations with an ESM. As in Part I, CF projections are presented in terms of (i) evaluation the presence of secular trends, (ii) changes in the 10th and 50th percentile annual mean CFs, and (iii) the probability of extended periods of anonymously low/high CFs (so-called wind droughts and wind-bonus periods). Linear trends in CFs calculated using monthly CFs over all months and for each season [spring (March–May), summer (June–August), autumn (September–November), and winter (December–February)] using nonparametric Theil–Sen regression (Wilks 2011) and p = 0.05.
P50(AEP) is the annual energy production (AEP) expected to be equaled or exceeded in one-half of years, and P90(AEP) is the AEP expected to be equaled or exceeded in 9 of 10 years on average. These properties are used in wind farm financing (Pryor et al. 2018). To be consistent with the literature and avoid confusion, here we present P50(CF) as the mean annual CF expected to be equaled or exceeded in one-half of years and P90(CF) as the value expected to be equaled or exceeded in 9 of 10 years on average. These values are computed for each wind farm and ESM realization as follows. Daily CFFINAL time series are used to compute annual mean CFs, and from those the P50(CF) and P90(CF) are computed for each vicennium period (i.e., 20-yr blocks; 1980–99, 2000–19, 2020–39, 2040–59, 2060–79, 2080–99). Differences in P50(CF) and P90(CF) in each vicennium are compared with the baseline period (1980–99).
Extended periods of anomalously low or high production, hereinafter referred to as “wind droughts” or “wind bonuses,” are assessed for two example sites: NE 3 and SGP 1. A wind drought is identified when the 30-day running-mean CF falls below the 20th percentile of the historical samples for that day of year in 1980–2009 computed using output from all ESM realizations. A wind bonus is identified when a given time sequence (e.g., 1–30 January 2040) lies above the 80th percentile of 30-day running-mean values in all downscaled results for 1980–2009. Because the threshold used to identify wind droughts or wind-bonus periods samples across all ESM realizations, a downscaled time series from a given ESM can systematically fall below the 20th percentile if the downscaled CFs exhibit a systematic negative bias. Thus, this analysis captures both systematic bias and temporal trends in the prevalence of wind droughts/wind-bonus periods.
CF projections developed here are necessarily a product of the synoptic climate, and thus any projected changes must be due to alteration of the frequency, tracking, or morphology of synoptic-scale systems. We further assume that global climate nonstationarity does not lead to emergence of synoptic types that are not present in the historical climate. Since the cluster analysis is inclusive (i.e., each day belongs to one cluster), we can test this assumption using the MAE statistic used in the clustering procedure to allocate each day to a cluster. If the MAE computed across all days is higher or lower in the historical (1979–2014) and future (2071–2100) periods, then it respectively implies the synoptic patterns are becoming less or more similar to the originally defined synoptic types. A Student’s t test (Wilks 2011) is used to assess if group MAE scores are significantly different between the time periods. We seek to attribute projected changes in CFs to either modifications to the frequency with which synoptic types occur or the intensity of the associated pressure gradients. In this analysis, the difference in the frequency of occurrence of the synoptic types is computed by subtracting the frequency of occurrence for the 36-yr period at the end of the century (2065–2100) from the historical period (1979–2014). The difference in the mean magnitude of the pressure gradient–derived CF adjustment [Eq. (3)] in the two periods is also quantified.
As described above, wind climates exhibit important low-frequency variability due to the action of internal climate modes and in common with all regional projections exhibit additional dependence on the climate model and the SSP (Hawkins and Sutton 2009; Hawkins et al. 2020). This leads to difficulty in projecting when/whether externally forced trends (i.e., signals) will emerge relative to internal variability (i.e., noise). Alternatively, this issue can be articulated in terms of the relative importance of different sources of diversity in climate projections and thus the projection uncertainty. Here we use a similar approach to projection uncertainty quantification to that in Lehner et al. (2020). The full range of P50(CF) values from each of the 115 downscaled realizations in each vicennium is decomposed and averaged by ESM (to quantify the model contribution to the range) and scenario (i.e., SSP contribution to the range), with the remainder after removing the ESM and scenario contributions being designated as internal variability (i.e., range due to internally forced climate variability).
4. Results
a. Historical validation and stationarity assessment
The synoptic typing from ERA5 using the different grids, wherein ERA5 is remapped to the resolutions of the ESMs, are highly self-similar (Fig. 4). This implies that the remapping of the ERA5 to each of the ESM grids does not greatly distort the synoptic types. The synoptic typing also appears to be stable through time. The MAE computed between patterns on all days and the cluster centroids is similar for samples drawn from 2065 to 2100 and 1979 to 2014. Student’s t-test scores indicate no difference in the MAE for any wind farm or any of the synoptic types. Further, the 90th percentile of MAE scores (i.e., the largest MAE that indicates the patterns that are least like the cluster centroid) also exhibits no consistent shift for each site, type, or ESM. This indicates that synoptic patterns detected in the historical climate are broadly consistent with those of the future climate in ESMs, with no discernable shift toward emergence of new synoptic types.
Frequencies of the synoptic types from ERA5 and the ESM in the historical period. Three results are shown for ERA5 resulting from mapping to the different ESM resolutions: 5 × 5 (dark green), 3 × 3 (light green), and 3 × 5 (moderate green) for CNRM and MIROC6, MPI, and UKESM, respectively. Note that the scales vary between panels. The types are ordered according to the median CFs of all days in that type from low to high, which are shown by the red line and scale on the right axis, with black error bars denoting the 95th-percentile range of CFs from the 1000 iterations of k-means clustering. Also shown by the colored bars (magenta for CNRM, blue for MIROC6, green for MPI, and gold for UKESM) is the model-mean bias for each ESM in frequency for each synoptic type. Site names are color coded to match the regions that they represent, as in Fig. 1.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
The relative frequency of the 15 synoptic types identified in ERA5 at each wind farm exhibit marked spatial variability but are all physically interpretable. For example, NE 3 and SGP 1 exhibit peaks in the frequency of type 11 or 12 (Fig. 4), which for both sites are patterns where cyclones are nearby (though not directly centered over the site), causing higher wind speeds and CFs. The distribution of days across the different synoptic types is relatively even at SWC 2, but highly concentrated at NGP 2 (Fig. 4). The synoptic types derived using ERA5 SLP fields are also demonstrably able to differentiate CFs. For example, as shown in Fig. 4, the median CF and range from the 2.5th to 97.5th percentile CF from the 1000 bootstrapped samples (referred to as the 95th-percentile range) at SGP 1 ranges from less than 10 to greater than 80 across the 15 types. At all wind farms, low CF is associated with anticyclone-dominated types (closer to 1) and higher CF under cyclone-dominated synoptic types (closer to 15). The estimates of daily CFFINAL for 2015–20 from downscaling of ERA5 exhibit some fidelity with regard to both the seasonality and the probability distribution of CFOBS (Fig. 5). Performance is worst for CFs at wind farms in the northern Great Plains (Fig. 5). The CFOBS exhibits relatively low variability across calendar months, but that variability, and indeed the variability across the entire CDF of daily CF values, is poorly reproduced by the downscaling method.
(left),(right center) Mean monthly CFs and (left center),(right) CDFs of CFs downscaled from the historical ESM simulations and ERA5 (1979–2014) and from observations (2015–20, black-dashed lines). Line coloring denotes the ESM (magenta = CNRM, blue = MIROC6, green = MPI, and gold = UKESM), and ERA5 (solid black). Site labels are color coded to match the regions that they represent, as in Fig. 1.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
We postulate two possible causes for the incomplete agreement between the seasonal cycle of CFs downscaled from ERA5 for 1979–2014 and the observations from 2015 to 2020. The first relates to subgrid-scale phenomena such as thermo-topographic flows that are not fully reproduced in daily mean fields of the SLP and the action of low-frequency variability in internal climate modes that may render the synoptic-scale conditions in 2015–20 an imperfect analog of those in 1979–2014. Examination of the synoptic-type frequencies in ERA5 from these two time periods using a χ2 test indicates very few statistically significant differences at p = 0.05, and the shift in type frequencies at individual wind farms is small (<10 percentage points across all 15 synoptic types).
b. Evaluation of ESM downscaling in the historical period
Most ESMs reproduce key aspects of the synoptic climate. The frequency of the synoptic types assessed using SLP fields from the ESM closely matches those from ERA5, consistent with Cannon (2020). At virtually all wind farms, biases in the synoptic-type frequencies relative to those in ERA5 are less than ±0.05. That is, if type 1 occurs on 15% of days over the historical period in ERA5, the frequency of that type in the ESM historical output is between 10% and 20%. However, MIROC6 exhibits biases of 0.1 (i.e., bias > 10 percentage points) or more from the ERA5 frequencies in NE 3, SGP 1, SGP 4, and SWC2 (Fig. 4). This bias in the synoptic-scale climate tends to be associated with excess presence of high-CF types that are dominated by intense low pressure systems, at the expense of synoptic types that are characterized by low to moderate CFs. This is consistent with previous work that found that MIROC6 exhibits excess energy at the synoptic scale in other global regions (Fernandez-Granja et al. 2021).
Probability distributions of daily maximum pressure gradients, used in the variance inflation [Eq. (3)], from the ESM exhibit similarity to those from ERA5, particularly at more easterly wind farms (NE 3; SGP 4) (Fig. 6). However, at the more westerly sites located in more complex terrain (NGP 2, NWC 1, and SWC 2), the lower-resolution ESMs (MPI and UKESM) exhibit a modal value that is 2–5 hPa lower than in ERA5, potentially due to terrain effects.
Distributions of pressure gradient values from 0 to 25 hPa for six example sites. Line coloring is as in Fig. 5. The line styles for the three sets of ERA5 results indicate the data resolution: H = 5 × 5 (dashed), L = 3 × 3 (dotted), and M = 3 × 5 (dot–dashed) grid cells to match CNRM and MIROC6, MPI, and UKESM, respectively.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
Over all 22 wind farms, downscaled CFs in the historical climate from MPI exhibit the highest fidelity, and those from MIROC6 exhibit comparatively poor fidelity. There is also marked geographic variability in the skill of the downscaled CFs. CFOBS (2015–20) from NE 3 and NWC 1 exhibit summertime minima and winter maxima with larger seasonal amplitude at NE 3 (Fig. 5). Downscaled CFs from ERA5 for 1979–2014 broadly capture this seasonality as do the downscaled ESMs, although CNRM and MIROC6 both exhibit a positive bias in mean monthly CF during all calendar months, largely due to excess frequency of high-CF synoptic types in MIROC6 (Fig. 4) and excess pressure gradients in CNRM (Fig. 6). CFOBS in the SGP also show a summer minimum but coupled with a spring maximum. Downscaling of the ERA5 reanalysis broadly captures the probability distribution of CFs, though downscaling results from the ESMs lack the summertime minimum of CF (Fig. 5). The seasonal cycle of observed CFs in the Southwest (SWC 2) exhibits peaks in the late spring and a minimum in the winter. The peak monthly CF is delayed by one month in ERA5 downscaling and is shifted to the summer in downscaling results from all ESMs. MIROC6 downscaling results are positively biased in all months, again largely due to excess frequency of high-CF synoptic types. Due to the poorer representation of the relative frequency of the synoptic types in the ESMs during the historical period at wind farms in the NGP and NWC (Fig. 4), in the following, analyses of CF projections focus primarily on wind farms in the NE, SGP, and SWC.
c. Trends in CF
Monthly mean CFs from the 22 wind farms at both the seasonal and annual time scales are dominated by negative trends in downscaling of output from the highest forcing scenario (SSP585) (Fig. 7). In the SGP, positive trends of up to 1.2 percentage points per decade are manifest in the spring, while other seasons are characterized by declines. The annual mean CFs are projected to decline by from 0 to −0.2 percentage points per decade for an overall difference from 2020 to 2100 of up to −1.6 percentage points. Conversely, results for the Northeast wind farms exhibit evidence of increases in winter (trends from −0.08 to 0.41 percentage points per decade at NE 3), and declines in all other seasons (Fig. 7). Projections for CFs at SWC 2 indicate positive trends in spring and negative trends in summer. This is in contrast to results from a variable-resolution ESM that projected increased wind resources in the summer rather than spring (Wang et al. 2018). However, consistent with Wang et al. (2018) the statistically downscaled CFs indicate declines in all other seasons of between −0.63 and −1.56 percentage points per decade in winter.
Linear trends in seasonal and annual mean CFs, reported in CF percentage points per decade. Trends are the mean for each ESM and the ensemble mean of all model runs (EM) [spring (March–May), summer (June–August), autumn (September–November), and winter (December–February)] and for all months (annual). To aid legibility, grid boxes are colors coded by trend sign and magnitude [≤−1 (orange), ≈0 (yellow), and ≥+1 (blue)]. Site names are color coded to match the regions that they represent, as in Fig. 1.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
Differences in P50(CF) and P90(CF) for vicennium periods relative to 1980–99 illustrate a clear and consistent primary dependence on the ESMs used in the downscaling with a secondary, but consistent, dependence on the SSP (Fig. 8). Consistent with expectations and the secular trend analyses, the differences are larger at the end of the century. The sign of differences is also consistent between P50(CF) and P90(CF). For example, at NE 3, downscaling of MIROC6 for all SSPs results in an increase in both P50(CF) and P90(CF), while downscaling of all three other ESMs leads to lower P50(CF) and P90(CF) by the end of the century. The two best-performing ESMs for NE 3 in the historical climate are MPI and UKESM (Fig. 5). The difference in P50(CF) is up to 5 percentage points for UKESM under SSP585 (Fig. 8). The mean of downscaled P50(CF) results for NE 3 from the MPI realizations under SSP585 are 2.5 percentage points lower in 2080–99. Thus, both of the ESMs for which the downscaling performed best in the historical climate indicate declines in P50(CF) for the Northeast by 2081–2100.
Difference in (left) P50(CF) and (right) P90(CF) values in each vicennium vs 1980–99. Colored lines indicate mean results from each ESM, as in Fig 5. Line styles differentiate the SSPs [SSP126 (dashed), SSP245 (dash–dotted), SSP370 (dotted), and SSP585 (solid)]. Horizontal lines show −2.5, 0, and 2.5 percentage points. Site names are color coded to match the regions that they represent, as in Fig. 1.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
The two wind farms in the SGP exhibit divergent trends in P50(CF) and P90(CF). The more coastal wind farm exhibits large magnitude increases in downscaling of MPI and CNRM, small magnitude declines for UKESM, and no change when MIROC6 is downscaled. Downscaling of MPI and CNRM exhibited best performance in the historical climate for this location, thus this analysis implies increases in P50(CF) of 2–8 percentage points by 2080–99 under SSP585. Results for the more-inland SGP wind farm are symptomatic of virtually no change in P50(CF) or P90(CF) (Fig. 8). P50(CF) and P90(CF) in the middle to end of the twenty-first century in the southern West Coast region (SWC 2) are −0.5 to −4 percentage points lower than in 1980–99 (Fig. 8) and are of largest magnitude in downscaling results from UKESM, which performed relatively well in the historical climate (Fig. 5). Declines in P50(CF) and P90(CF) of −0.5 to −2 percentage points are found by the end of the current century in downscaling of all ESMs for sites in the Midwest and the northern Great Plains (not shown) but are considered rather uncertain due to the poor fidelity in the historical climate.
d. Projected changes in wind drought and wind bonus
Extended wind-drought and wind-bonus production periods can greatly impact power utilities and directly affect project planning and feasibility. Projections of the frequency of these periods of anomalously high and low production tend to follow the patterns set by the site-specific historical CF patterns and projected trends discussed above (Figs. 9 and 10). Generally, wind farms for which the downscaling results indicated increased CFs also tend to be associated with an increased frequency of wind-bonus periods, while those with declines in P50(CF) are also characterized by an increased frequency of wind droughts. However, downscaling results for some wind farms indicate more nuanced changes.
Wind-drought frequency in each vicennium from 1980 to 2099 (1980–99, 2000–19, …, 2080–99) for wind farms (top) NE3 and (bottom) SGP 1 presented by season for each ESM as well as the multimodel mean for four SSPs. Redder colors indicate higher drought frequency and blue colors indicate lower frequency than in the 1980–2009 period (the expected value of 0.2 is the mean of the seasonal values in the period 1980–2009). Filled circles indicate differences that are statistically significant at p = 0.05, and open circles show results for p = 0.1.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
The bias in summertime CFs in downscaling of MIROC6 at both NE3 and SGP 1 (Fig. 5), leads to a negative bias in wind-drought occurrence in the summer in the period that persists over all future vicennium periods (Fig. 9). Downscaling results for the other ESMs indicate annual drought and bonus frequencies that are close to 0.2 (occurring 20% of the time when averaged over the four seasons) in the 1980–2009 period, indicating that these three ESMs are performing in a consistent manner between the three ESMs and also in time when downscaled to produce CF estimates. However, bias in the historical period in terms of wind-drought or wind-bonus periods does not seem to uniformly dictate the time trajectory. For example, downscaling of UKESM at NE 3 leads to a negative bias in wind-drought frequency in 1980–99 and exhibits no change through time for most SSPs. Conversely, downscaling of UKESM at SGP 1 exhibits a negative bias in wind-drought frequency in 1980–99 and trends toward a frequency of >0.2 by the end of the century to a degree that scales with the SSP forcing.
Changes in the probability of wind-drought (Fig. 9) and wind-bonus (Fig. 10) periods at NE 3 and SGP 1 through time are broadly consistent with patterns in the seasonal trend analysis and tend to become more intense in stronger radiative forcing scenarios. Downscaling of CNRM shows large-magnitude increases in wind-drought probability for NE 3 in summer and autumn toward the end of the twenty-first century. The probability of occurrence increases to about 0.4 under SSP585, which implies a doubling in the prevalence (Fig. 9). Increasing probability of wind drought in autumn is consistent across downscaling of all ESMs at both NE 3 and SGP 1, while summer increases and winter decreases are consistent across all ESMs at SGP 1. Periods of high production decrease in frequency in the summer and autumn at NE 3 (i.e., shifting from about 20% to nearly 5%) and increase in winter (i.e., shifting from 20% to nearly 30%) (Fig. 10). The variability in the probabilities of wind-drought and wind-bonus periods between decades, seasons, and models (Fig. 10) illustrates the continued importance of internal climate variability in impacting the frequency of occurrence of wind droughts and wind-bonus periods.
As in Fig. 9, but for wind-bonus frequency.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
e. Attribution of projected changes in CF
Figures 11 and 12 show the differences in mean CFs in 2065–2100 versus 1979–2014 by SSP and ESM and the decomposition of these differences into forcing due to shifts in the frequency of different synoptic types and changing pressure gradients. Changes in CFs at NWC 1 and NGP 2 are modest and thus are not shown here, although NWC 1 will be discussed in the context of changes shown at SWC 2. At most sites, the modest projected decreases in CFs are driven primarily by declines in the maximum pressure gradient. For example, at SGP 1, downscaling of the CNRM model suggests the frequency of the various synoptic types shifts in a manner that would prompt higher CFs, but that change is (more than) offset by declines in the maximum pressure gradient and thus CFRESID. Declines in the pressure gradient are stronger under higher warming scenarios, indicating that externally forced changes are driving large-scale weakening of pressure gradients in the ESMs. Conversely, trends in downscaled CFs due to shifts in the frequency distributions of synoptic types vary significantly by site, region, and climate forcing. The relative frequencies of the synoptic types are generally projected to shift from high CF types (12–15) toward midtier CF types (7–12) rather than from the highest and lowest CF types. For example, large declines in the highest CF types (13–15) are compensated for by increased frequency of types 5–11 at SWC 2. Even though some sites are expected to experience greater frequencies of high-CF types (i.e., low pressure systems, such as at NE 3), they are projected to be weaker than in the historical climate.
For NE 3 and SGP 1, differences in CFs 2065–2100 vs 1979–2014 by SSP expressed as a trend in percentage points per decade, along with a preliminary attribution achieved by decomposition into changes in (left) synoptic types and pressure gradients. (right) The difference in the frequency of each synoptic type under the high forcing scenario (SSP585). Note that the scales vary between panels. Results are colored to reflect the ESM, as in Fig. 5, and the line style in the left panels indicates the attribution to changes in synoptic type (dotted) and pressure gradients (dashed). The boxplots show the range of trends across all downscaled ESMs from changing synoptic types (light gray) and changes in pressure gradients (dark gray).
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
As in Fig. 11, but for SGP 4 and SWC.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
Most wind farms (and regions) exhibit CF trends that are consistent with broader changes caused by future climate warming that generally favors weaker pressure gradients and slowed circulation (Fabiano et al. 2021). For example, the Northern Hemisphere storm track is expected to shift northward as warming continues, with the largest changes in autumn and winter (Shaw et al. 2016). This is most consistent with the trends in CFs at northeastern sites (NE 3), where changes in CFs, as well as associated drought and bonus periods, are strongest in autumn and weak to positive in winter. Coastal areas of the southern Great Plains (SGP 4) experience strong increases in CFs in some models (especially MPI), driven by increasing pressure gradients that may arise due to increasing land–sea thermal contrasts.
Projections of CFs exhibit larger differences from the historical climate for the southern West Coast region than in other regions, with contributions from both shifts in the frequency of individual synoptic types and pressure gradient intensity (Figs. 11, 12). Thus, an analysis is undertaken focused on causes of historical reductions of CFs in this region and whether there is evidence for increased frequency of those conditions. For example, the wind drought that impacted wind farms in the California ISO (CAISO) and ERCOT during Q1 of 2015 (Pryor et al. 2020b) was associated with large-scale ridging over the West Coast (Lledó et al. 2018) (Fig. 13). Such ridges block transient cyclones (i.e., shift the synoptic patterns toward lower CF types) and reduce pressure gradients. This ridging is linked to large-scale internal climate modes of variability such as the PDO and NAM (Gibson et al. 2020). Ridging along the U.S. West Coast associated with the wind drought in Q1 of 2015 is linked to the positive phase of the PDO (early winter) and NAM (late winter) (Fig. 13). When month indices of the PDO and NAM are computed from each ESM realization following Coburn and Pryor (2021), the PDO does not show a secular trend, but NAM exhibits a tendency toward increased frequency of the positive phase (Fig. 13). Thus, the large-magnitude declines in wintertime CFs over the SWC (and also NWC 1) (Fig. 7) may be partly a manifestation of a shift in the ESMs toward a higher frequency of positive-phase PDO and associated large-scale ridging (Fig. 7) that tends to steer transient cyclones away from the southwestern United States.
(top) The 500-hPa geopotential height (m) from ERA5 in January 2015 during a wind drought that impacted CAISO and ERCOT. Absolute value of the monthly mean net CF (%) and deviation of monthly mean CFs from the long-term average (percentage points) for the four highlighted ISOs are respectively 5 and −14 for CAISO, 23 and −9 for ERCOT, 37 and +7 for the New York ISO (NYISO), and 32 and +3 for the New England ISO (NEISO) (Pryor et al. 2020a). A 12-month running-mean index time series of the (middle) PDO and (bottom) NAM that caused the strong ridging event in January 2015 from ERA5 (black; 1980–2020) and the ESM under the SSP585 is computed following Coburn and Pryor (2021). Individual model realizations are shown in gray, and the model means are shown by the colored lines as in Fig. 5.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
f. Uncertainty analysis
As shown above, there is marked divergence in CF projections from different ESMs and under different SSPs. Here, the 92 projections of P50(CF) in each vicennium period are decomposed into three sources of uncertainty; ESM, SSP scenario, and internal variability. The results indicate that the action of internal climate modes will continue to play a critical role in dictating projection spread at all locations up to at least the middle of the twenty-first century. After that point, the ESM and the SSP scenario play increasingly important, or even dominant, roles at all locations (Fig. 14). This is because greenhouse gas concentrations and resultant warming from the differing SSPs are very similar over the early part of the century and are consistent with increased dependence on SSP toward the end of the century in projections of air temperature and precipitation (Lehner et al. 2020). P50(CF) from downscaling of the different ESMs and SSP during 2000–19 spans a range of 4.3–13.4 across the 22 wind farms. This range increases in each vicennium to 5.8–24.7 in 2080–99. The increase in the spread of P50(CF) estimates with time is very marked at SWC 2 and SGP 4. Thus, assuming the divergence in the projections from the different ESM realizations and SSP represents the uncertainty in the P50(CF) projections; the uncertainty increases toward the end of the century. The contribution to this uncertainty arising from the different emissions scenarios tends to be small in the first vicennium (2000–19) and grows through the end of the century (2080–99), although it only rises to 10%–25% at most wind farms, regardless of ESM. One exception is SWC 2 where, much like all southern West Coast sites, the climate forcing scenario contributes up to 40% of the P50(CF) projection uncertainty in the late twenty-first century.
Decomposition of CF projection uncertainty into that deriving from the SSP scenario (blue), ESM (teal), and internal climate variability (yellow) for each vicennium. Site names are color coded to match the regions they represent, as in Fig. 1.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
5. Conclusions
Daily CFs at 22 wind farms across North America are used as predictands to develop downscaling models that are then applied to 115 ESM realizations to the end of the current century. The new statistical downscaling approach is based on synoptic typing with variance inflation scaled by the pressure gradient intensity and is unique in the literature because the predictand reflects actual expected power at operating wind farms. The 23 realizations from four ESMs and five climate forcing scenarios (historical and Shared Socioeconomic Pathways) are downscaled and validated against observed CF data and synoptic conditions for the historical period (1979–2014) from ERA5. These ESMs are able to capture the synoptic-scale climate with relatively high fidelity, leading to some skill in reproducing the seasonal cycle and probability distribution of daily CFs. However, skill varies between ESM and from region to region. Over much of North America, MIROC6 exhibits excess cyclogenesis, contributing to the excessive power production (CF) projected for the historical period. Downscaled CFs from all ESMs exhibit poorest agreement with CFOBS over the northern Great Plains.
Future changes in CFs downscaled from the ESM realizations are quantified using secular trend analyses applied at the seasonal and annual time series, by computing P50(CF) and P90(CF) in vicennium periods, and in terms of the likelihood of anomalously high and low wind power production periods. The downscaling method is also used to attribute changes in CFs to changes in synoptic-type frequency and/or intensity.
Projected changes in CFs, and hence power production, at operating wind farms across North America are highly ESM dependent and season specific. For example, CF declines over the Northeast are most marked in summer and autumn, and the increases in the SGP and along the SWC are manifest in spring. Downscaled projections of CFs from the different ESMs are generally consistent in some regions (e.g., wind farm SGP 1 in the southern Great Plains). However, they are less consistent in terms of magnitude of change for the example wind farm, SWC 2, and at some wind farms the projected changes in CFs span zero (e.g., SGP 4) (Figs. 7; 8). Further, downscaling of MPI leads to the largest-magnitude positive trend in CFs at SGP 4 while downscaling of UKESM projects the largest-magnitude declines in CFs at SWC 2 (Fig. 8). However, in most regions, these changes in CFs are small relative to the variability introduced by internal climate modes.
Declines in CFs (and increased frequency of seasonal wind droughts) projected for later in the current century are primarily the result of declining pressure gradients that weaken all synoptic types. In contrast, SGP 4 exhibits evidence for increasing pressure gradients and CFs in most ESM realizations. Internal climate variability will continue to play an important role in dictating P50(CF) in each decade or vicennium at all sites 2050 and beyond in most regions. Projected large-magnitude CF declines in the SWC may be linked to a trend toward the positive phase of NAM and enhanced wintertime ridging over the West Coast. Thus, while the forcing scenario plays a larger role in the spread of P50(CF) toward 2100 at SWC 2, it may do so, in part, through the impact of forced climate change on internal modes of variability like NAM. The spread of projected CFs at a given wind farm for a specific vicennium period derives primarily from the ESM downscaled, but internal variability is also important, even under high warming scenarios. This reemphasizes the importance of efforts to improve model fidelity with respect to representation of internal climate modes in ESM (Coburn and Pryor 2021) and efforts to understand teleconnections from those internal climate modes to local and regional climates and wind power production (Coburn 2021; Schoof and Pryor 2014; Shepherd et al. 2022).
This study combines dynamical and statistical downscaling approaches and uses the convergence of results from the two independent approaches to assign higher credibility to the resulting projections. It is worth recalling the relative strengths and weaknesses of dynamical and statistical downscaling approaches. In Part I, WRF is nested within the MPI ESM and used to model the flow conditions using a grid spacing of 12 km for the period 2009–49. Wind speeds at wind turbine hub height are then converted in power output using site-specific wind turbine power curves. Modeled CFs from the initial part of the simulation are then evaluated relative to observed CFs. The results indicate this model chain exhibits some skill in reproducing the spatial variability of power production but also exhibits higher root-mean-square error relative to the observations than similar modeling performed with 2 km grid spacing for 3 years in the historical climate. The modeled daily CFs are then used to examine possible changes in wind power production at the operating wind farms to the midcentury under a high climate forcing scenario. As described in Part I, this research methodology has the advantage that dynamical causes of flow variability are described using fundamental equations of motion, but the computational costs necessitate that only one model chain, one ESM, and one SSP can be sampled. The statistical downscaling approach presented here directly uses the CFs from operating wind farms to construct the transfer functions linking conditions at the synoptic scale to power production. The computational efficiency of this research methodology allows us to sample more of the envelope of uncertainty by including a range of ESMs, multiple realizations from each ESM, and sampling across climate forcing pathways. It thus permits decomposition of the projection uncertainty into that deriving from the climate forcing, the climate model, and the internal climate variability. However, the statistical downscaling is predicated on stationarity of the synoptic types, requires large data volumes to train the transfer functions, and cannot fully represent the dynamical linkages between atmospheric conditions, land surface conditions, and power production. The overarching research framework thus seeks to leverage the relative strengths and complementarity of the downscaling approaches. A summary of the P50(CF) projections from Part I where the WRF Model is used to downscale the MPI ESM along with results from the statistical downscaling presented herein is given in Fig. 15. The mean values of projected changes in P50(CF) to the midcentury from statistical downscaling of the four different ESMs under the SSP585 scenario range from −1.8 to +1.3 percentage points. For wind farms in the four regions common to the statistical and dynamical downscaling, northern Great Plains, southern Great Plans, Midwest, and Northeast, the results indicate an inconsistent sign of difference in P50(CF) across the ESMs and/or the different ESM realizations to 2040–60. However, like the dynamical downscaling, the majority of the projections made using statistical downscaling indicate evidence for fairly small magnitude declines in P50(CF) of up to a maximum of 1.9 percentage points (Fig. 15). After 2050, the statistical downscaling projections of CFs suggests larger-magnitude differences in P50(CF) relative to 1980–99 under the SSP585 climate forcing scenario. In 2080–99, all P50(CF) projections, sampled across ESMs and realizations, for the southern West Coast and Midwest regions are below those in 1980–99, by −3.9 to −0.5 percentage points (Fig. 15). Projections for the single wind farm in the northern West Coast indicate some evidence for increases at the end of the century. For the Northeast and the northern Great Plains, projections from three of the four ESMs indicate lower P50(CF) in 2080–99 than 1980–2099, while in the southern Great Plains, the results are mixed and exhibit a clear dependence on proximity to the coast and the ESM.
Synthetic precis of projected differences in the median annual CF [P50(CF)] for the midcentury and end of the century for the pairs from Part I and this paper. Results are for the SSP585 and are summarized across wind farms in a given region (see Fig. 1). Results for MPI-WRF dynamical downscaling are the differences of 2040–49 vs 2010–19. Results for the statistical downscaling are shown for 2040–59 and 2080–99 vs 2000–19, and the range denotes the variations across the downscaled ESMs. The shading denotes the direction of change: red indicates that all projections from the five ESMs are lower in the future, blue indicates that all projections are higher, and yellow indicates that the projected changes span 0.
Citation: Journal of Applied Meteorology and Climatology 62, 1; 10.1175/JAMC-D-22-0047.1
This work highlights the high value of data sharing from wind farm owner–operators and the utility in making industry-relevant projections that can be readily monetized and contextualized in other sources of CF gains (see Part I). Future work would be greatly enhanced by sharing from a wider range of operating wind farms and by the availability of longer CFOBS time series for statistical downscaling model training.
Acknowledgments.
This work is supported by the U.S. Department of Energy (DE-SC0016605) and used computing resources from the National Science Foundation Extreme Science and Engineering Discovery Environment (XSEDE) (the allocation award to author Pryor is TG-ATM170024). The suggestions of two reviewers and the editor are gratefully acknowledged.
Data availability statement.
Power production data used in this analysis are proprietary. Model output from ERA5 is available from the Copernicus database (https://cds.climate.copernicus.eu/cdsapp#!/home) and is obtained via the data request form from two specific datasets: ERA5 hourly data on single levels from 1959 to present and ERA5 hourly data on pressure levels from 1959 to present. Mode indices were calculated from ERA5 monthly averaged data on single levels from 1959 to present. CMIP6 Earth System Model data are available from the Earth System Grid Federation catalog (https://esgf-node.llnl.gov/search/cmip6/).
REFERENCES
Bell, B., and Coauthors, 2021: The ERA5 global reanalysis: Preliminary extension to 1950. Quart. J. Roy. Meteor. Soc., 147B, 4186–4227, https://doi.org/10.1002/qj.4174.
Benestad, R. E., D. Chen, A. Mezghani, L. Fan, and K. Parding, 2015: On using principal components to represent stations in empirical–statistical downscaling. Tellus, 67A, 28326, https://doi.org/10.3402/tellusa.v67.28326.
Berkovic, S., 2018: Wind regimes and their relation to synoptic variables using self-organizing maps. Adv. Sci. Res., 15, 1–9, https://doi.org/10.5194/asr-15-1-2018.
Bermúdez, M., L. Cea, E. Van Uytven, P. Willems, J. F. Farfán, and J. Puertas, 2020: A robust method to update local river inundation maps using global climate model output and weather typing based statistical downscaling. Water Resour. Manage., 34, 4345–4362, https://doi.org/10.1007/s11269-020-02673-7.
Broderick, C., and R. Fealy, 2015: An analysis of the synoptic and climatological applicability of circulation type classifications for Ireland. Int. J. Climatol., 35, 481–505, https://doi.org/10.1002/joc.3996.
Brunner, L., A. G. Pendergrass, F. Lehner, A. L. Merrifield, R. Lorenz, and R. Knutti, 2020: Reduced global warming from CMIP6 projections when weighting models by performance and independence. Earth Syst. Dyn., 11, 995–1012, https://doi.org/10.5194/esd-11-995-2020.
Cannon, A. J., 2020: Reductions in daily continental-scale atmospheric circulation biases between generations of global climate models: CMIP5 to CMIP6. Environ. Res. Lett., 15, 064006, https://doi.org/10.1088/1748-9326/ab7e4f.
Coburn, J., 2021: Climatological teleconnections with wind energy in a midcontinental region. J. Appl. Meteor. Climatol., 60, 305–322, https://doi.org/10.1175/JAMC-D-20-0203.1.
Coburn, J., and S. C. Pryor, 2021: Differential credibility of climate modes in CMIP6. J. Climate, 34, 8145–8164, https://doi.org/10.1175/JCLI-D-21-0359.1.
Davy, R. J., M. J. Woods, C. J. Russell, and P. A. Coppin, 2010: Statistical downscaling of wind variability from meteorological fields. Bound.-Layer Meteor., 135, 161–175, https://doi.org/10.1007/s10546-009-9462-7.
Dixon, K. W., J. R. Lanzante, M. J. Nath, K. Hayhoe, A. Stoner, A. Radhakrishnan, V. Balaji, and C. F. Gaitán, 2016: Evaluating the stationarity assumption in statistically downscaled climate projections: Is past performance an indicator of future results? Climatic Change, 135, 395–408, https://doi.org/10.1007/s10584-016-1598-0.
Dixon, P. G., M. Allen, S. N. Gosling, D. M. Hondula, V. Ingole, R. Lucas, and J. Vanos, 2016: Perspectives on the synoptic climate classification and its role in interdisciplinary research. Geogr. Compass, 10, 147–164, https://doi.org/10.1111/gec3.12264.
Ekström, M., M. R. Grose, and P. H. Whetton, 2015: An appraisal of downscaling methods used in climate change research. Wiley Interdiscip. Rev.: Climate Change, 6, 301–319, https://doi.org/10.1002/wcc.339.
Fabiano, F., V. L. Meccia, P. Davini, P. Ghinassi, and S. Corti, 2021: A regime view of future atmospheric circulation changes in northern mid-latitudes. Wea. Climate Dyn., 2, 163–180, https://doi.org/10.5194/wcd-2-163-2021.
Fernandez-Granja, J. A., A. Casanueva, J. Bedia, and J. Fernandez, 2021: Improved atmospheric circulation over Europe by the new generation of CMIP6 Earth System Models. Climate Dyn., 56, 3527–3540, https://doi.org/10.1007/s00382-021-05652-9.
Gibson, P. B., and N. J. Cullen, 2015: Regional variability in New Zealand’s wind resource linked to synoptic-scale circulation: Implications for generation reliability. J. Appl. Meteor. Climatol., 54, 944–958, https://doi.org/10.1175/JAMC-D-14-0273.1.
Gibson, P. B., D. E. Waliser, B. Guan, M. J. DeFlorio, F. M. Ralph, and D. L. Swain, 2020: Ridging associated with drought across the western and southwestern United States: Characteristics, trends, and predictability sources. J. Climate, 33, 2485–2508, https://doi.org/10.1175/JCLI-D-19-0439.1.
Gramcianinov, C. B., R. M. Campos, R. de Camargo, K. I. Hodges, C. G. Soares, and P. L. da Silva Dias, 2020: Analysis of Atlantic extratropical storm tracks characteristics in 41 years of ERA5 and CFSR/CFSv2 databases. Ocean Eng., 216, 108111, https://doi.org/10.1016/j.oceaneng.2020.108111.
Hawkins, E., and R. Sutton, 2009: The potential to narrow uncertainty in regional climate projections. Bull. Amer. Meteor. Soc., 90, 1095–1108, https://doi.org/10.1175/2009BAMS2607.1.
Hawkins, E., D. Frame, L. Harrington, M. Joshi, A. King, M. Rojas, and R. Sutton, 2020: Observed emergence of the climate change signal: From the familiar to the unknown. Geophys. Res. Lett., 47, e2019GL086259, https://doi.org/10.1029/2019GL086259.
Hersbach, H., and Coauthors, 2020: The ERA5 global reanalysis. Quart. J. Roy. Meteor. Soc., 146A, 1999–2049, https://doi.org/10.1002/qj.3803.
Horton, P., 2021: Analogue methods and ERA5: Benefits and pitfalls. Int. J. Climatol., 42, 4078–4096, https://doi.org/10.1002/joc.7484.
Huang, H.-Y., S. B. Capps, S.-C. Huang, and A. Hall, 2015: Downscaling near-surface wind over complex terrain using a physically-based statistical modeling approach. Climate Dyn., 44, 529–542, https://doi.org/10.1007/s00382-014-2137-1.
Hutengs, C., and M. Vohland, 2016: Downscaling land surface temperatures at regional scales with random forest regression. Remote Sens. Environ., 178, 127–141, https://doi.org/10.1016/j.rse.2016.03.006.
Jones, P. D., C. Harpham, and K. R. Briffa, 2013: Lamb weather types derived from reanalysis products. Int. J. Climatol., 33, 1129–1139, https://doi.org/10.1002/joc.3498.
Kirchner‐Bossi, N., R. García‐Herrera, L. Prieto, and R. M. Trigo, 2015: A long‐term perspective of wind power output variability. Int. J. Climatol., 35, 2635–2646, https://doi.org/10.1002/joc.4161.
Lanzante, J. R., K. W. Dixon, M. J. Nath, C. E. Whitlock, and D. Adams-Smith, 2018: Some pitfalls in statistical downscaling of future climate. Bull. Amer. Meteor. Soc., 99, 791–803, https://doi.org/10.1175/BAMS-D-17-0046.1.
Lee, J. C. Y., and M. J. Fields, 2021: An overview of wind-energy-production prediction bias, losses, and uncertainties. Wind Energy Sci., 6, 311–365, https://doi.org/10.5194/wes-6-311-2021.
Lehner, F., C. Deser, N. Maher, J. Marotzke, E. M. Fischer, L. Brunner, R. Knutti, and E. Hawkins, 2020: Partitioning climate projection uncertainty with multiple large ensembles and CMIP5/6. Earth Syst. Dyn., 11, 491–508, https://doi.org/10.5194/esd-11-491-2020.
Lledó, L., O. Bellprat, F. J. Doblas‐Reyes, and A. Soret, 2018: Investigating the effects of Pacific sea surface temperatures on the wind drought of 2015 over the United States. J. Geophys. Res. Atmos., 123, 4837–4849, https://doi.org/10.1029/2017JD028019.
Mauritsen, T., and Coauthors, 2019: Developments in the MPI‐M Earth System Model version 1.2 (MPI‐ESM1.2) and its response to increasing CO2. J. Adv. Model. Earth Syst., 11, 998–1038, https://doi.org/10.1029/2018MS001400.
Millstein, D., J. Solomon-Culp, M. Wang, P. Ullrich, and C. Collier, 2019: Wind energy variability and links to regional and synoptic scale weather. Climate Dyn., 52, 4891–4906, https://doi.org/10.1007/s00382-018-4421-y.
O’Neill, B. C., and Coauthors, 2017: The roads ahead: Narratives for shared socioeconomic pathways describing world futures in the 21st century. Global Environ. Change, 42, 169–180, https://doi.org/10.1016/j.gloenvcha.2015.01.004.
Pryor, S. C., and R. J. Barthelmie, 2014: Hybrid downscaling of wind climates over the eastern USA. Environ. Res. Lett., 9, 024013, https://doi.org/10.1088/1748-9326/9/2/024013.
Pryor, S. C., and J. T. Schoof, 2019: A hierarchical analysis of the impact of methodological decisions on statistical downscaling of daily precipitation and air temperatures. Int. J. Climatol., 39, 2880–2900, https://doi.org/10.1002/joc.5990.
Pryor, S. C., and J. T. Schoof, 2020: Differential credibility assessment for statistical downscaling. J. Appl. Meteor. Climatol., 59, 1333–1349, https://doi.org/10.1175/JAMC-D-19-0296.1.
Pryor, S. C., J. T. Schoof, and R. J. Barthelmie, 2005: Empirical downscaling of wind speed probability distributions. J. Geophys. Res., 110, D19109, https://doi.org/10.1029/2005JD005899.
Pryor, S. C., T. J. Shepherd, and R. J. Barthelmie, 2018: Interannual variability of wind climates and wind turbine annual energy production. Wind Energy Sci., 3, 651–665, https://doi.org/10.5194/wes-3-651-2018.
Pryor, S. C., R. J. Barthelmie, and T. J. Shepherd, 2020a: 20% of US electricity from wind will have limited impacts on system efficiency and regional climate. Sci. Rep., 10, 541, https://doi.org/10.1038/s41598-019-57371-1.
Pryor, S. C., F. W. Letson, and R. J. Barthelmie, 2020b: Variability in wind energy generation across the contiguous United States. J. Appl. Meteor. Climatol., 59, 2021–2039, https://doi.org/10.1175/JAMC-D-20-0162.1.
Pryor, S. C., R. J. Barthelmie, M. S. Bukovsky, L. R. Leung, and K. Sakaguchi, 2020c: Climate change impacts on wind power generation. Nat. Rev. Earth Environ., 1, 627–643, https://doi.org/10.1038/s43017-020-0101-7.
Pryor, S. C., J. J. Coburn, R. J. Barthelmie, and T. J. Shepherd, 2023: Projecting future energy production from operating wind farms in North America. Part I: Dynamical downscaling. J. Appl. Meteor. Climatol., 62, 63–80, https://doi.org/10.1175/JAMC-D-22-0044.1.
Raynaud, D., B. Hingray, G. Evin, A.-C. Favre, and J. Chardon, 2020: Assessment of meteorological extremes using a synoptic weather generator and a downscaling model based on analogues. Hydrol. Earth Syst. Sci., 24, 4339–4352, https://doi.org/10.5194/hess-24-4339-2020.
Sailor, D. J., M. Smith, and M. Hart, 2008: Climate change implications for wind power resources in the northwest United States. Renewable Energy, 33, 2393–2406, https://doi.org/10.1016/j.renene.2008.01.007.
Schoof, J. T., and S. C. Pryor, 2014: Assessing the fidelity of AOGCM‐simulated relationships between large‐scale modes of climate variability and wind speeds. J. Geophys. Res. Atmos., 119, 9719–9734, https://doi.org/10.1002/2014JD021601.
Sellar, A. A., and Coauthors, 2020: Implementation of U.K. Earth System Models for CMIP6. J. Adv. Model. Earth Syst., 12, e2019MS001946, https://doi.org/10.1029/2019MS001946.
Shaw, T. A., and Coauthors, 2016: Storm track processes and the opposing influences of climate change. Nat. Geosci., 9, 656–664, https://doi.org/10.1038/ngeo2783.
Shepherd, T., J. J. Coburn, R. J. Barthelmie, and S. C. Pryor, 2022: Exploring ENSO-induced anomalies over North America in historical and future climate simulations that use HadGEM2-ESM output to drive WRF. Climate, 10, 117, https://doi.org/10.3390/cli10080117.
Tatebe, H., and Coauthors, 2019: Description and basic evaluation of simulated mean state, internal variability, and climate sensitivity in MIROC6. Geosci. Model Dev., 12, 2727–2765, https://doi.org/10.5194/gmd-12-2727-2019.
Van Uytven, E., J. D. Niel, and P. Willems, 2020: Uncovering the shortcomings of a weather typing method. Hydrol. Earth Syst. Sci., 24, 2671–2686, https://doi.org/10.5194/hess-24-2671-2020.
Voldoire, A., and Coauthors, 2019: Evaluation of CMIP6 DECK experiments with CNRM-CM6‐1. J. Adv. Model. Earth Syst., 11, 2177–2213, https://doi.org/10.1029/2019MS001683.
Wang, J., and X. Su, 2011: An improved k-means clustering algorithm. Proc. 2011 IEEE Third Int. Conf. on Communication Software and Networks, Xi’an, China, Institute of Electrical and Electronics Engineers, 44–46, https://ieeexplore.ieee.org/document/6014384.
Wang, M., P. Ullrich, and D. Millstein, 2018: The future of wind energy in California: Future projections with the variable-resolution CESM. Renewable Energy, 127, 242–257, https://doi.org/10.1016/j.renene.2018.04.031.
Wilks, D. S., 2011: Statistical Methods in the Atmospheric Sciences. Vol 100, Academic Press, 704 pp.
Wohland, J., D. Folini, and B. Pickering, 2021: Wind speed stilling and its recovery due to internal climate variability. Earth Syst. Dyn., 12, 1239–1251, https://doi.org/10.5194/esd-12-1239-2021.
Zscheischler, J., M. D. Mahecha, and S. Harmeling, 2012: Climate classifications: The value of unsupervised clustering. Procedia Comput. Sci., 9, 897–906, https://doi.org/10.1016/j.procs.2012.04.096.