The United States Atlantic coast is a development site for upcoming offshore wind projects. There are 15 leasing areas located throughout the Atlantic Outer Continental Shelf, where a number of offshore wind farms are planned to be developed . Therefore, characterizing offshore boundary layer winds in the region has risen in importance. Accurate forecasting will provide developers with a better understanding of local wind patterns, which can inform wind farm planning and layout decisions . Additionally, improved weather prediction will allow for real-time adjustments of turbine operation to increase their operating efficiency and protect them against unnecessary wear and tear .

The Mid-Atlantic Bight (MAB) is an offshore cold pool region spanning the eastern United States coast from North Carolina up through Cape Cod, Massachusetts, and it overlies the offshore wind leasing areas. The cold pool forms during the summer months, when the ocean becomes strongly stratified and the thermocline traps colder water near the ocean floor. During the transition to winter, as sea surface temperatures (SSTs) drop, the stratification weakens and the cold pool breaks down. Thus, the cold pool generally persists from the spring through the fall. Southerly winds that drive surface currents offshore will result in coastal upwelling of this colder water, and, at times, strong winds associated with storm development can mix the cold pool upward, cooling the surface and influencing near-surface temperatures and winds .

Accurate representation of the MAB in forecasting models is important because SSTs are closely tied to offshore winds. Horizontal temperature gradients between land and the ocean, as well as vertical temperature gradients over the ocean – which can form, for example, when SSTs are anomalously cold, as with the MAB – help define offshore airflow. In particular, variations in temperature can lead to or impact short-lived offshore events occurring on hourly timescales, such as sea breezes and low-level jets (LLJs). Sea breezes are driven by the land–sea temperature difference, which, if strong enough (around 5 ∘C or greater), can generate a circulation between the water and the land . With a relatively colder ocean, as during summer months, this leads to a near-ground breeze blowing landward, with a weak recirculation toward the ocean aloft . Similarly, the near-surface horizontal and air–sea temperature differences dictate the strength of stratification over the ocean. Studies have found a robust link between atmospheric stability and LLJ development, so accurately representing SSTs is key to modeling near-surface stability and, accordingly, LLJs . A Weather Research and Forecasting model (WRF) analysis of the region conducted by found that in the MAB leasing areas specifically LLJs occurred in 12 % of the hours in June 2010–2011. Both LLJs and sea breezes can affect individual wind turbine and whole farm operation, so forecasting them correctly can improve power output and turbine reliability .

Typical climate and weather model initialization and forcing inputs are reanalysis products, such as ERA5 and MERRA-2, which are global data sets that assimilate model output with observations to create a comprehensive picture of climate at each time step considered . These data sets primarily include global SST products that are produced at lower temporal and spatial resolutions than what can be available via regional, geostationary satellites. These coarser-resolution data sets, therefore, do not capture observed hourly and, in many cases, diurnal fluctuations in SSTs, which may influence their ability to properly force sea breezes and LLJs. Some preliminary comparisons between weather simulations, forced with different SST products, indicate that this particular input can have a significant impact on modeled offshore wind speeds .

Few studies have examined the impact of finer-temporal-resolution SST products specifically on wind forecasting, and to the authors' knowledge, none so far have focused on the Mid-Atlantic. There have been studies looking at numerical weather prediction model (NWP) sensitivity to SST, but they have considered other regions or different, often coarser-spatial- and coarser-temporal-resolution products . In this article, we explore the effects of forcing the Weather Research and Forecasting model (WRF), an NWP used for research and operational weather forecasting, with different SST data sets characterized by different spatial and temporal resolutions in the Mid-Atlantic region during the summer months. Specifically, we address differences in model performance on monthly timescales and then contrast characterization effectiveness during shorter wind events. Section  lays out the data, model setup, and methods used in this study. Section  explains the findings of our simulations, and Sect.  explores their implications. Finally, Sect.  summarizes the intent of the study as well as its findings.

We first validate three different SST data sets against observations taken at an array of buoys off the Mid-Atlantic coast during the months of June and July 2020. Following validation, we select two of the three data sets for use as inputs to two different model simulations in the MAB region, which are identically configured aside from the SST data. August is not considered due to data availability constraints at the time of the study. The output data are compared with in situ measurements taken at buoys (SSTs) and floating lidars (winds) in the region. A 140 m hub height is assumed, based on the analysis of regions with moderate wind resource by . We evaluate performance primarily via a set of validation metrics calculated on monthly timescales. We then flag specific events during which the model generally captures regional winds, but output from the two simulations deviate significantly (defined in this study as 1 or more standard deviations from their mean differences) from one another. Again, validation analysis is performed for these periods.

We first assess how well each SST data set compares with buoy data. Following SST validation, we evaluate the model's wind characterization performance when forced with different SST data. We assume a hub height of 140 m and compare output winds at this altitude against measurements taken via floating lidars off the Mid-Atlantic coast. Specifically, we analyze monthly performance and then select several shorter periods during June and July 2020 during which we compare wind characterization accuracy.

After evaluation of the SST data sets, only OSTIA- and GOES-16-forced WRF simulations are compared, as during July – one-half of our study period – MUR25 SSTs validate significantly worse against buoy measurements than the other two products. OSTIA SSTs validate the best out of the three data sets. Average, simulated hub-height winds across monthly periods validate similarly for both OSTIA and GOES-16. At an event-scale temporal resolution – that is, on the order of hours – GOES-16 and OSTIA perform comparably, with GOES-16 marginally outperforming OSTIA.

In Sect. , we saw varied performance between the two SST products, depending on the variable in consideration (SST or winds) and the timescale (monthly or event scale). SST in OSTIA validated better than GOES-16 across the buoy array. However, hub-height wind speeds at each lidar location, for each month, point to a similar performance by both data sets. More nuanced events that occurred over hourly-to-daily timescales, which were selected based on differences between simulation output and overall WRF performance, also indicate comparable performance between the two simulations, with, in general, a slightly stronger performance using GOES-16 SSTs. Events such as these are of importance to wind energy forecasts because we found that they often correlate with wind ramps, during which times wind speeds fall below the rated power section of turbine power curves. In this region, power output is very sensitive to fluctuations in wind speed. Therefore, obtaining the most accurate wind forecast within this regime is important.

Although OSTIA SSTs at the lidars validate better than GOES-16, modeled hub-height (140 m) winds do not demonstrate the same trend. Compared with the 140 m wind heat maps for the events (Figs. , , and ), good wind characterization performance does not necessarily correspond to good SST validation (Appendix ), which implies that SSTs across the full study area, not just localized temperatures, may have an impact on winds in the leasing area.

To further examine the disjointedness between SST and wind speed point validations, we consider model performance during each event in conjunction with how well their corresponding SSTs validate with observations across the buoy array (where data are available). During the 21–22 June event, 140 m wind speed bias and EMD in both simulations are the lowest of all the cases considered (Fig. ), but there is an overall cold SST bias (-0.31 ∘C for OSTIA, -0.84 ∘C for GOES-16) across the buoy arrays, and SST correlation is very poor (Fig. ). During the 10–11 July event, OSTIA SSTs validate better than GOES-16 (Fig. ), but the two simulation's hub-height wind speeds validate similarly, with relatively poor performance across all validation metrics compared with the other flagged events. During the 18 July event, while SSTs from both products validated the best out of the three events (Fig. ) and the models captured the overall trend in wind speeds with correlations of 0.88 for both simulations, EMD and the magnitude of the bias were both greater than the 21–22 June event.

Across all three events there is no clear relationship between wind forecasting ability and SST correctness. However, as with the monthly analysis, the relatively larger differences between SST and wind speed validations in the GOES-16 event simulations compared with those of OSTIA suggest that GOES-16's finer temporal and spatial resolutions may more accurately influence larger-scale dynamics throughout the region, leading to greater forecasting skill. Though beyond the scope of this study, more widespread in situ observations and validation of the two data sets against observations would help address how large of an effect this is. Such a study would additionally allow for a better understanding of how much a nuanced representation of diurnal SST cycling (as in the GOES-16 product, Fig. ) influences the accuracy of wind forecasting in the region.

Rapid fluctuations in wind speed, on the order of minutes to hours, are overall not well captured by either simulation. This is most evident during the 10–11 July event, when the magnitude and timing of wind speed changes are most misaligned. Correlation in particular is low (0.62), and RMSE is high (3 ms-1). In comparison, during the 18 July event, when wind speeds slow without any ramping, both simulations perform better; in particular, correlation is relatively strong for both (0.88). Similarly, during the 21–22 June event, which lacks any sharp ramping, correlation is also above 0.75 for both GOES-16 and OSTIA. Shorter, weaker fluctuations in wind speed during this event are not well captured by either simulation (Fig. ), although GOES-16 validates with nominally less error (Fig. ). It is possible that GOES-16's finer resolution may contribute to its slightly stronger performance in this case, as the product does capture hourly fluctuations in SSTs. In contrast, during the 10–11 July storm, winds are so strong and variable that SST resolution does not appear to make a significant difference, and the two simulations both show significant error.

Also of note are the differences in weather between June and July in the Mid-Atlantic region. June average wind speeds are faster than those in July for both simulations. Overall cooler SSTs throughout the region are also present in June, with a weaker temperature gradient in the cold bubble offshore of Cape Cod than in July. The difference in temperature in this region between GOES-16 and OSTIA is more pronounced in July, with GOES-16 showing warmer monthly average SSTs (Fig. ). The corresponding average wind speed and surface pressure differences between the products are also more distinct in July. While the two simulations validated similarly for each month, in the more turbulent July environment, OSTIA outperforms GOES-16. However, due to the greater number of missing pixels that required gap-filling in July as compared with June (36.86 % more over the course of each month, resulting from the increased cloud cover – and the possible stronger coastal cold pool), this indication of product superiority may be questionable. The addition of more refined post-processing to account for incorrectly cloud-filtered pixels, as shown in , has been demonstrated to improve the accuracy of GOES-16 SSTs. A comparison of simulations using OSTIA against those using the sophisticated GOES-16 product should be conducted in the future to evaluate the level of improvement delivered using the latter's more computationally expensive data set.

In this study, we evaluated how SST inputs affect how well WRF forecasts winds off the Mid-Atlantic coast. We initially considered three remotely sensed SST data sets and validated them against in situ observations taken at various buoys in the region. The first data set, OSTIA, has a 0.05∘ resolution, is output at a daily interval, and is assimilated with in situ observations. The second product, GOES-16, has a finer resolution of 0.02∘ and is generated on an hourly interval but does not have the same level of post-processing as OSTIA. The data set contains gaps and lacks assimilation with in situ observations. To account for missing data, we ran an EOF process to statistically analyze trends in each monthly data set and fill its gaps. The third SST data set, MUR25, has a 0.2∘ spatial resolution and a daily temporal resolution, and similar to OSTIA it is assimilated with in situ observations. MUR25 was removed from further analysis after SST validation, as its performance compared with the other two data sets was notably poor.

Following SST validation, WRF was run for June and July of 2020 in the Mid-Atlantic domain depicted in Fig. . The simulation setups were identical with the exception of the input SST field (key model parameters presented in Table ). We compared the characterized 140 m (hub-height) winds against lidar observations by calculating correlation, bias, RMSE, and EMD on both monthly and event-length (hourly) timescales. Times during which there was significant difference in wind speed output from the two simulations and the general WRF performance was satisfactory were flagged as events. Our findings indicated that while OSTIA SSTs validated better against buoy measurements, the model validated comparably when forced with each SST data set, with GOES-16 having output nominally more accurate wind speeds. These findings suggest that the differences in temporal and spatial resolution between the two SST products influence wind speed characterization, with finer resolutions leading to better forecasting skill.

Overall, this study shows that SST inputs to WRF do affect forecasted winds and how well they validate against observations in future leasing areas in the Mid-Atlantic. Although GOES-16 SSTs validate worse against buoy observations than OSTIA SSTs, on both monthly and event timescales, the wind characterization produced using GOES-16 inputs validates as well as, or in some cases better than, the OSTIA characterization, indicating that finer temporal and spatial SST resolution may contribute to improved wind forecasts in this region. To further explore this, a higher-quality post-processed GOES-16 data set assimilated with in situ observations, such as that created by , could be evaluated in future studies; generating a comparable data set was beyond the scope of this study.

This step of the research is in its early stages, and more cases need to be considered – on both monthly and event scales. Future work should focus on identifying and analyzing more promising events – in particular, wind ramps and low-pressure systems, as they have shown to produce larger differences between GOES-16- and OSTIA-forced model output. Wind direction validation, as well, should be considered in the future, as it was beyond the scope of this study.