Turbulence intensity (TI) is often used to quantify the strength of turbulence in wind energy applications and serves as the basis of standards in wind turbine design. Thus, accurately characterizing the spatiotemporal variability in TI should lead to improved predictions of power production. Nevertheless, turbulence measurements over the ocean are far less prevalent than over land due to challenges in instrumental deployment, maintenance, and operation. Atmospheric models such as mesoscale (weather prediction) and large-eddy simulation (LES) models are commonly used in the wind energy industry to assess the spatial variability of a given site. However, the TI derivation from atmospheric models has not been well examined. An algorithm is proposed in this study to realize online calculation of TI in the Weather Research and Forecasting (WRF) model. Simulated TI is divided into two components depending on scale, including sub-grid (parameterized based on turbulence kinetic energy (TKE)) and grid resolved. The sensitivity of sea surface temperature (SST) on simulated TI is also tested. An assessment is performed by using observations collected during a field campaign conducted from February to June 2020 near the Woods Hole Oceanographic Institution Martha's Vineyard Coastal Observatory. Results show that while simulated TKE is generally smaller than the lidar-observed value, wind speed bias is usually small. Overall, this leads to a slight underestimation in sub-grid-scale estimated TI. Improved SST representation subsequently reduces model biases in atmospheric stability as well as wind speed and sub-grid TI near the hub height. Large TI events in conjunction with mesoscale weather systems observed during the studied period pose a challenge to accurately estimating TI from models. Due to notable uncertainty in accurately simulating those events, this suggests summing up sub-grid and resolved TI may not be an ideal solution. Efforts in further improving skills in simulating mesoscale flow and cloud systems are necessary as the next steps.
While the number of wind turbines installed offshore in US waters is small, it is expected to continuously escalate for the foreseeable future (Musial et al., 2021). Therefore, it is critical to accurately simulate the wind resource, as well as the turbulence, in offshore environments. Wind power generation is sensitive to atmospheric turbulence in addition to wind speed (Yang et al., 2017; Berg et al., 2019; Vanderwende and Lundquist, 2012; St. Martin et al., 2016; Wharton and Lundquist, 2012). Atmospheric turbulence impacts the loads on the turbine and, ultimately, the life span of the wind turbine (Mücke et al., 2011). Even in the ambient environment without any turbulence, the turbine itself generates turbulence via its wake (Wu and Porté-Agel, 2012; Brand et al., 2011; Porté-Agel et al., 2020; Hansen et al., 2012).
Turbulence is generally largest near the surface where the wind shear and buoyancy are generally largest, and it is closely connected to stability in the planetary boundary layer (PBL) (Rodrigo et al., 2015; Bardal et al., 2018; Garratt, 1994). Compared to conditions over land, the sea surface is relatively smooth, leading to smaller amounts of shear-generated turbulence near the surface. In addition, while the continental PBL is generally unstable during the day, the diurnal variation in marine PBL is not evident but dependent on the thermal gradient between the sea surface and atmosphere above. Thus, the atmosphere is generally less turbulent offshore than onshore (e.g., Bodini et al., 2020). Nevertheless, strong turbulence-producing events such as hurricanes, winter storms, and mesoscale convective systems can impact offshore wind farms.
In wind energy applications, turbulence is often quantified using turbulence intensity (TI) (Bodini et al., 2020; Barthelmie et al., 2007), and TI is the basis of standards used in wind turbine design (e.g., Shaw et al., 1974). Accurately characterizing the spatiotemporal variability in TI should lead to improved predictions of power production. Earlier studies have discussed how TI may influence the power production of turbines (Bardal and Sætran, 2017; Kaiser et al., 2007; Saint-Drenan et al., 2020; Clifton and Wagner, 2014). It has been shown that power production during periods of high or low TI can vary by up to 20 % (Lundquist and Clifton, 2012). However, turbulence and stability measurements over the ocean are made far less frequently than over land as it is challenging to maintain and operate instruments over the open ocean for long durations.
Atmospheric models such as weather prediction and large-eddy simulation (LES) models can potentially bridge this gap as they can simulate turbulence based on atmospheric and surface conditions for any region over the globe. While LES models are too computationally expensive to simulate long periods of time, mesoscale meteorological models are much more efficient and can also estimate turbulent properties such as turbulence kinetic energy (TKE) by applying an applicable turbulence parameterization. Nevertheless, the derivation of TI from atmospheric models has rarely been examined. Since multiple model uncertainties may contribute to TI bias under various conditions, comprehensive observational datasets are desired for model validation and to help quantify the errors.
This study addresses these shortcomings to derive TI from a mesoscale weather model and access its performance in an offshore environment. We implement an online calculation of TI in the standard version of the Weather Research and Forecasting (WRF) model and complete a quantitative assessment of simulated TI using observations collected during a field campaign conducted from February to June 2020 near the Woods Hole Oceanographic Institution (WHOI) Martha's Vineyard Coastal Observatory (MVCO). Simulations incorporating high-resolution sea surface temperature (SST) are performed to evaluate the impact of SST on changing the atmospheric stability and simulated values of TI.
The paper is organized as follows: observational datasets used in this study are described in Sect. 2. Details of model configuration, TI derivations, and experimental design are provided in Sect. 3. Section 4 covers our primary findings including (a) wind and turbulence profiles near the sea surface, along with corresponding air–sea temperature difference; (b) the sensitivity of simulated wind and turbulence to SST forcing; (c) the relationship between TI and the bulk Richardson number; and (d) the role of model-resolved TI. To conclude, a summary and discussion are given in Sect. 5.
A US Department of Energy (DOE)- and Bureau of Ocean Energy Management-supported field campaign was conducted from February through June 2020 near the Woods Hole Oceanographic Institution (WHOI) Martha's Vineyard Coastal Observatory (MVCO; Austin et al., 2000) with the goal of evaluating the performance of DOE's lidar buoys (Gorton and Shaw, 2020; Krishnamurthy et al., 2021; Sheridan et al., 2022). The MVCO is a purpose-built facility for conducting detailed atmospheric and oceanic research. A major component of the MVCO is the Air-Sea Interaction Tower (ASIT), built near the vicinity of the Rhode Island and Massachusetts wind energy areas (Fig. 1). At the site, a suite of wind energy specific measurements were made, including a pair of cup anemometers at the top of the tower (26 m a.m.s.l. – above mean sea level), a wind vane at 23 m a.m.s.l., and a WindCube v2
vertically profiling lidar (hereafter reference lidar) on the main platform located 13 m a.m.s.l. The centers of the reference lidar range gates are 53, 60, 80, 90, 100, 120, 140, 160, 180, and 200 m a.m.s.l. WindCube v2 measures line-of-sight radial velocity sequentially along four cardinal directions, with a zenith angle of 28
TI is defined as the standard deviation of the horizontal wind speed (
The model configuration, including horizontal and vertical grid spacing, spin-up time and selections of atmospheric, SST forcing, and PBL and surface layer parameterizations, can contribute to the uncertainty in simulated offshore wind (Bodini et al., 2021; Siedersleben et al., 2020; Hahmann et al., 2015; Chang et al., 2015; Sward et al., 2023; Optis et al., 2021). In this study, the version 4.2 WRF model is used to simulate offshore near-surface winds. A single model domain centered on the MVCO site is used that encompasses part of the northeastern USA and adjacent oceans (Fig. 1). The choices of model physics parameterizations for this study are consistent with the setup for the 20-year wind resource dataset released by the National Renewable Energy Laboratory (NREL) (Optis et al., 2020). Optis et al. (2020) conducted a series of model sensitivity experiments with respect to surface layer and PBL parameterizations, reanalysis data, and SST forcing. The results of their model assessment indicated that the largest uncertainty is associated with the choice of PBL parameterizations, and the Mellor–Yamada–Nakanishi–Niino (MYNN) boundary layer parameterization (Nakanishi and Niino, 2009) generally outperforms the Yonsei University (YSU; Hong et al., 2006) scheme off the east coast of North America. Hence the MYNN boundary layer parameterization, as well as other parameterizations described by Optis et al. (2020), was used for generating the CA20 dataset as well as our simulations. The model uses a horizontal grid spacing of 2 km and a stretched vertical coordinate with 60 levels. There are approximately 10 model levels between the surface and 200 m. The Eta (Ferrier) microphysics parameterization, MYNN surface layer parameterization, Unified Noah Land Surface Model parameterization (Chen and Dudhia, 2001), and RRTMG longwave and shortwave radiation parameterization (Iacono et al., 2008) are employed. Initial and boundary conditions are taken from NOAA's High-Resolution Rapid Refresh (HRRR; Benjamin et al., 2016) product. The HRRR analysis has several advantages over other coarse-resolution reanalysis products. For instance, (1) the model core of HRRR, the WRF model, is identical with what we use in this study; (2) it has a grid spacing of 3 km, which is very close to the grid spacing used here (2 km) to match the CA20 dataset; and (3) it is constrained hourly by assimilating radar observations including Doppler velocity and reflectivity, which reduces the uncertainties in the prediction of precipitating clouds. All these advantages would primarily mitigate model uncertainties in part due to issues in model balance and spin-up.
Realistically, observed TI occurs across a range of spatial and temporal scales. Application of the WRF model leads to an artificial separation of motions that are explicitly represented by the model and turbulence treated by the boundary layer parameterization; thus both grid-resolved and sub-grid motions can give rise to TI. These two contributions to TI will be referred to as model-resolved and sub-grid TI, respectively. A new algorithm is implemented to extract the wind variation and mean wind speed over a 10 min window, allowing for the calculation of both sub-grid and model-resolved TI. Since the three-dimensional components of turbulent wind speed cannot be obtained through the boundary layer parameterization, the sub-grid TI is derived by leveraging TKE (Eq. 2) prognosed by the MYNN boundary layer parameterization.
The derivation of TI has been proposed in several studies including a recent
document of Larsén (2022). In this report, the relation between TI and TKE is derived using the Kaimal boundary layer turbulence model (Kaimal and Finnigan, 1994). However, since the TI derivation in that work divides horizontal wind into along- and cross-turbine wind components rather than zonal and meridional winds that are used in WRF, we decided to apply the TI formula used in Shaw et al. (1974), Wharton and Lunquist (2011), and Bodini et al. (2020):
The lidar-measured wind variances (
Therefore, Eq. (3) can be approximated by the form of
Summary of comparison between two types of TI output from the WRF model.
To facilitate comparison against observations collected during the field campaign near the MVCO ASIT, the WRF model simulations were performed for February through June of 2020. The “baseline” experiment is a concatenation of a series of 36 h simulations. Each of these simulations is initialized at 00:00 UTC and continuously integrated for 36 h. To avoid model spin-up issues, the first 12 h of each individual simulation is discarded, and the resulting 24 h results are retained for the analysis.
Since the near-surface atmospheric stability can strongly influence turbulence intensity at hub height, the uncertainty in representation of sea surface temperature should also be considered. Despite the HRRR analysis representing convective-scale structures better than other coarse-resolution reanalysis products as mentioned earlier in Sect. 3.1, it does not provide corresponding SST forcing data. Therefore, a sensitivity experiment (named “sstupdate” hereafter) is conducted to examine the variability induced by replacing the SST forcing in the model.
Note that Optis et al. (2020) used two SST products including the Operational Sea Surface Temperature and Ice Analysis (OSTIA, resolution 0.05
The simulated wind speed (WSPD), wind direction (WDIR), TI, TKE, and air–sea
temperature gradient (air temperature at 2 m above the surface (
Here we highlight the comparisons by showing the results from March and May
in 2020 as examples. The panels in Figs. 3 and 4 show a composite month-long comparison of near-surface properties between baseline simulations and the tower-mounted lidar for March and May, respectively. Note the model profiles in the figures only extend to 200 m to match the range of the lidar. The results show the model reproduces the major “convective” (negative values of
Time–height comparison of wind speed (WSPD), wind direction (WDIR), air–sea temperature gradient (
Similar to Fig. 3, but results for May 2020 are displayed.
Scatterplots depicting the relationship between simulated (baseline) WSPD and sub-grid TI with color shadings for representing 80 m TI
and TKE errors in
Metrics including correlation coefficient (CC), root mean square error (RMSE), and bias of 80 m TI, TKE, WSPD, and air–sea temperature difference are given as computed for baseline and sstupdate simulations.
To quantify the model performance for the entire period of study, an assessment of the baseline simulation is carried out that examines the 80 m TI, TKE, WSPD, and air–sea temperature gradients. Note data points from simulations and observations are only counted when wind speed is between 5 and 25 m s
As described in Sect. 3.3, the default SST in the baseline simulation is replaced by the NASA JPL SST analysis product (Chin et al., 2017), which has finer-scale features as opposed to the SST representation in the baseline simulation. This sensitivity experiment is named sstupdate, and the impacts of SST forcing are examined in this section.
The same metrics as those used in the previous section are applied to the sstupdate simulation and given in Table 2. These indicate the replacement of SST forcing has positive impacts on all the examined variables. For instance, the CC, RMSE, and bias for TI as simulated by sstupdate (baseline) are 0.59 (0.56), 0.035 (0.037), and
RMSE profiles of
The probability density functions (PDFs) are given in Fig. 7 to help describe the similarities of each variable among the three datasets. Figure 7a shows that the distribution of observed TI has a wider peak between 0.025 and 0.1 than the simulated distributions. While the baseline simulation has more points with smaller TI values (peak around 0.04), the distributions of TI from the sstupdate show an additional peak around 0.09 and 0.1, and there are more large values of TI. The median TI of sstupdate (0.066) is much closer to observations (0.07) than the baseline (0.057), indicating that the SST representation has a notable impact on TI simulation. The individual impact of SST on the TKE, WSPD, and air–sea temperature difference is displayed in Fig. 7b–d, respectively. The model tends to produce more instances with small TKE than was observed (0.67; Fig. 7b), and sstupdate (0.4) has slightly larger median TKE than the baseline (0.377). Note part of the discrepancy in TKE may be attributed to varying uncertainty in lidar turbulence retrievals as a function of atmospheric stability. Sathe et al. (2015) found hub-height turbulence (
Probability density function (PDF) plots illustrating the results
of
Application of the higher-resolution SST impacts the time of evolution of the winds and turbulence. Figure 8 shows how the cold bias seen in the SST of baseline simulation is effectively reduced during May 2020, which subsequently fixes the cold bias in near-surface air temperature. While there are relatively small differences between the two simulations in the TKE and WSPD, we do find improvement in TI over some periods. For example, during 5–8, 13–14, 17–18, and 30–31 May, TI simulated by the sstupdate simulation are generally higher than what is simulated by the baseline simulation and closer to the observed values. The correction of TKE has a larger impact on TI when WSPD is relatively small (less than 10 m s
Time-series display of data from baseline, sstupdate simulations, and corresponding observations for May 2020. From top to bottom rows, the near-surface air temperature, sea surface temperature, hub-height TKE, WSPD, and TI are displayed.
The impact of SST is also examined in the context of monthly variability. Figure 9 summarizes the metrics calculated for each simulation, variable, and month with the observational data as reference. The analysis suggests that the overall performance of the simulated air–sea temperature gradient (TG) is improved when the higher-resolution SST forcing is used, and the improvement is more prominent in the spring months than in February. Despite only slight impact on the correlation for the simulated TG by replacing the default SST (Fig. 9a), the corresponding RMSE (Fig. 9b) and bias (Fig. 9c) for the TG are considerably reduced. While a relatively small positive impact is shown for TKE and wind speed, the improvement in the bias of TI is prominent, as shown by the reduced RMSE and bias, particularly for April and May.
Metrics of the
Many earlier studies showed that the bulk Richardson number (
Time-series plots of observed and simulated
While simulated results show that the model has good skill for large-TI
(over 0.2) events, there are periods with moderate TI (0.05 to 0.2) where the model misses the observed peaks as indicated by the green arrows in Fig. 10d. It is found that in those time periods, the buoyancy component (
To explore how TI can be related to the buoyancy and shear components of the
Scatterplots mapping the hub-height sub-grid TI in the coordinates of buoyancy (
Although in general, TI decreases as the buoyancy becomes more positive and
the atmosphere becomes stable as identified in the observations and the two
WRF model configurations (Fig. 11a–c), a population of reddish circles (large values of TI) in the lower-right quadrant of the figure does not follow this relationship (Fig. 11a). While both simulations fail to represent these cases (Fig. 11b and c), the fractional difference in simulated TI between the two simulations indicates that the sstupdate simulation generally has larger TI than the baseline simulation (Fig. 11d). This can be attributed to an overall reduction in the cold bias in the baseline simulation as shown in Fig. 9. Furthermore, a larger fractional increase in TI is found in the regime where the conditions are between neutral and slightly stable (buoyancy component
The next step was to locate these specific data points in time by applying
thresholds of buoyancy component
While some of the intermittent turbulent events may cause the occurrence of large TI under small shear and stable conditions, we find that most of the cases with high levels of TI in stable conditions and weak shear can be linked to fluctuations in wind speed associated with mesoscale features of the flow, and they are partially resolved by the mesoscale model, or for which there are timing errors. In Fig. 10e, the mesoscale (model-resolved) TI, described in Sect. 3.2, is denoted by the green line. For the periods such as 3–4, 6–7, and 10–11 February, when the simulated sub-grid TI is much smaller than that of observations, spikes in mesoscale (model-resolved) TI have amplitudes similar to the observed values. This suggests the need to consider both sub-grid and mesoscale TI when modeling TI derived from a mesoscale atmospheric model. Furthermore, the uncertainties in the timing of simulated mesoscale weather events may lead to notable contrasts between the timing of modeled and observed peaks in TI. This can impact the simulated TI in two ways as the mesoscale activity not only generates variance in the winds over larger spatiotemporal scales than the sub-grid turbulence does, but also effectively offsets dynamic and thermodynamic conditions near the sea surface in the boundary layer that influence the simulated sub-grid TI generated by the boundary layer parameterization. Our analysis demonstrates the utility of identifying scale-dependent uncertainties in TI modeling, which allows us to isolate the causes of errors in the simulated TI.
The metrics used in Table 1 are employed again to assess the impact of
resolved TI on the TI prediction and are listed in Table 3. Results indicate that the addition of resolved TI reduces the bias for baseline configuration (
Metrics including the correlation coefficient (CC), root mean square error (RMSE), and bias for TI that are computed based on total (sub-grid and resolved) TI simulated by baseline and sstupdate simulations.
Algorithms to derive TI are successfully implemented in the WRF model and tested offshore using a multi-month dataset collected at the MVCO using a combination of Doppler lidar, tower, and buoy data. Simulated TI is divided into two components depending on scale, comprising sub-grid and grid resolved. To obtain simulated TI, we calculate the square root of the sum of horizontal wind variances and then divide this by the mean wind speed over a 10 min window. While sub-grid TI is diagnosed from parameterized turbulence kinetic energy (TKE) through the MYNN PBL parameterization, resolved TI is estimated by using the model-resolved wind variances.
The modeled TI computed over a wide range of atmospheric conditions is analyzed and validated by using a variety of observations collected at the offshore tower at the MVCO between February and May in 2020. The primary findings include the following:
The model's PBL scheme reacts reasonably well to changes in the vertical temperature gradient near the sea surface despite the fact that the simulated TKE is generally smaller than what is observed by the lidar, especially for events with large TKE values. The modified WRF model slightly underestimates TI, and the error is mainly attributed to relatively large negative bias in TKE as the predicted wind speed generally agrees with observation. The WRF model has difficulty predicting periods of weak TI due to smaller air–sea temperature differences and fluxes within the simulated boundary layer. Overall cold bias in the SST of baseline simulation is effectively reduced by substituting it with more accurate SST forcing. This subsequently reduces model biases in near-surface air temperature as well as in hub-height WSPD and TI. A regime of large observed values of TI during periods with positive buoyancy and weak shear is identified, but the values are not captured by the model. Many of the events occurred in conjunction with mesoscale weather systems, but directly summing up sub-grid and resolved TI does not improve the TI prediction. This is because the primary source of uncertainty in those events is caused by the unrealistic representation of mesoscale weather systems, including timing errors, in the model.
Our analysis suggests additional model constraints are required to further
improve model representation of TI and TKE in mesoscale or finer-scale cloud
(system) processes. Approaches that couple mesoscale model with data
assimilation techniques to improve skill in the simulated mesoscale flow
features and cloud predictions (Tai et al., 2020, 2021; Gaudet et al., 2022) could be very beneficial.
The codes implemented in the WRF model for the presented simulations as well as the scripts for data analysis can be made available upon request. In the case of interest, please contact the authors.
The observational and simulated datasets used in this study can be made available upon request.
SLT: conceptualization, methodology, execution, analysis, and writing. LKB: conceptualization, methodology, supervision, and review. RK: observational data processing, consulting, and review. RN: observational data processing. AK: observational data consulting and review.
The contact author has declared that none of the authors has any competing interests.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Pacific Northwest National Laboratory is operated by Battelle for the US Department of Energy under contract DE-AC05-76RLO1830.
This research was supported by the National Offshore Wind Research and Development Consortium under agreement number 147502.
This paper was edited by Sukanta Basu and reviewed by three anonymous referees.