Variability in North Sea wind energy and the potential for prolonged winter wind drought

The United Kingdom is committed to substantially increasing offshore wind capacity in its drive to decarbonise electricity production and achieve net zero. If low wind episodes—or ‘wind drought’ events—occur during high energy demand periods, energy security may be threatened without alternative supply. The challenge of managing the variability of wind power will increase into the future as its share in the energy mix increases. This study focuses attention on the North Sea as a centre of current and planned offshore wind resource, and on the winter season, given the occurrence of weather patterns that risk security of supply. We use a large climate model ensemble, providing a dataset an order of magnitude larger than the reanalysis‐based observations, to better sample wind drought events. This leads to a more robust estimate of their frequency and persistence and their dynamical teleconnections compared with the observational record. We define week‐long wind drought events, based on a 20th percentile threshold in 10 m wind speed, during which wind power is estimated to be around half that in a typical week of winter. Wind drought events of up to two consecutive weeks have been observed, but the model indicates a 1‐in‐40 chance of three or more continuous weeks of wind drought each winter, with the single most prolonged simulated event lasting 5 weeks. There is a doubling of the likelihood of these prolonged wind drought events during El Niño, indicating that monitoring and predicting the state of the tropical Pacific may be useful in assessing the risk of wind drought events in an upcoming winter.

world, behind China (Global Wind Energy Council, 2022), and is projected to grow rapidly as the UK government pursues decarbonisation of the power system by 2035 (HM Government, 2021). It has set a target of 40GW of offshore wind by 2030, close to quadrupling capacity this decade (HM Government, 2020). Variability in wind power presents a growing challenge to the management and operation of the power system (Bloomfield et al., 2016;Guerra et al., 2022), particularly where there is high penetration of wind power (wind generation as a large fraction of total electricity generation). Recent episodes of exceptionally low winds, such as in late February to early March 2021, when Britain's wind energy capacity remained below 20% for 11 days (Staffell et al., 2021), and then over a more prolonged period across northern and western Europe in summer to early autumn 2021 (Copernicus Climate Change Service, 2021), have acted as conspicuous reminders of the importance of taking variability into account when designing energy system security. Typical mitigation measures include balancing wind power supply with alternative generation, and to this end requirements for alternative capacity must be properly assessed.
To ensure resilience of the power system now and in the coming years as offshore wind generation grows, better understanding of the severity, frequency, and duration of low wind episodes would be useful. Variability in winds is likely to dominate over trends in the next few decades (Pryor et al., 2020), and hence having improved information on present day characteristics of wind drought is valuable. The observational and reanalysisbased record of wind speeds is limited in its length, and by definition extreme low wind episodes are rare and therefore difficult to characterise. Extended 20th century reanalysis products present the opportunity to sample a wider range of events, but large, possibly spurious, trends in wind speeds and related fields and interproduct inconsistencies have been reported Wohland et al., 2019). Here, we use the UNSEEN (UNprecedented Simulated Extremes using ENsembles, Thompson et al., 2017) approach to address this problem. UNSEEN is a framework through which a large ensemble of initialised climate model simulations is used to provide a synthetic but realistic event set that greatly increases the sample size of extreme events and gives more robust information about their likelihood and properties.
In this study, we apply UNSEEN to wind drought events over the North Sea in winter, which is the season of primary concern, given the occurrence of weather patterns that produce both high energy demand (low temperatures) and low wind power supply in the United Kingdom (Thornton et al., 2017). The study domain takes in a broad area across the North Sea as this is a global hotspot of current and planned offshore wind farms (Akhtar et al., 2021), which serve the United Kingdom and some other European countries. We first evaluate the fidelity of the climate model for this application before using the large ensemble to assess the severity, frequency, and duration of winter wind drought events, using the basic unit of a week of low winds as the timescale of analysis. Finally, an examination of the wider conditions associated with wind drought events is undertaken to investigate what remote factors may contribute to prolonged wind drought.

| DATA, DOMAIN AND TIME PERIOD OF ANALYSIS
The spatial domain used in this study is a box, 53 N-59 N, 1 E-7 E, that encompasses much of the North Sea ( Figure 1, inset). This domain is representative of the region, on a spatial scale commensurate with large-scale meteorological drivers of variability in winds, but does not cover some individual wind farms, such as those close to the east coast of Scotland and northern England. Nevertheless, comparisons with publicly available wind energy data on monthly (BEIS, 2022b) and daily (Drax, 2022) time resolutions suggest that this box average is meaningful in terms of energy production, at least in Great Britain. The focus season is winter, December-January-February (DJF), when energy demand is highest.
The climate model data are taken from the hindcast ensemble of the Met Office Decadal Prediction System version 4 (Hermanson et al., 2023;in prep). This system uses the Met Office global coupled climate model HadGEM3-GC3, which has an atmospheric grid resolution of $60 km in the midlatitudes, and an ocean resolution of 0.25 (Williams et al., 2018), which places it among the higher resolution models submitted to the sixth phase of the Coupled Model Intercomparison Project (Lin & Yu, 2022). The hindcast ensemble consists of 10 members and is initialised each 1 November, making the winter season (DJF) months 2-4 of the hindcast. The lead time of 1 month is beyond the deterministic range of near-term weather prediction for North Sea wind and so ensemble members are well spread with little predictable signal remaining (as shown in Figure 1), yet the climatology of the model is still expected to be near to that observed. We use 60 years of data, 1961-2020, where 1961refers to December 1960to February 1961. This gives a total of 600 (60 years Â 10 members) winter seasons. For this study, we use near-surface wind speed at 10 m, available from the model at daily resolution. Daily mean wind speed is averaged over the North Sea domain. With spatial averaging, note that variability is reduced as compared with a single gridbox or group of wind farms (refer to Figure S1 to compare gridpoint variability with box-average variability).
In addition, an analysis of 1.5 m air temperature, sea surface temperature (SST) and mean sea level pressure (MSLP) is undertaken to inspect the wider conditions associated with low winds over the North Sea. These variables have been linearly detrended at gridpoint level to place the focus on variability.
The model's fidelity is tested (Section 3) by comparing its 10 m winds against those from ERA5, the fifthgeneration atmospheric reanalysis from the European Centre for Medium Range Weather Forecasts (Hersbach et al., 2020). Reanalyses are valuable and widely used datasets in renewable energy applications, thanks to their complete and consistent data coverage, but are subject to a range of uncertainties derived from the numerical model, the observations, and data assimilation methods (Ramon et al., 2019). Global and multiterritory studies, including North Sea data points, have evaluated reanalysis wind speed data for wind power applications, and found that among the available products, ERA5 is generally in best agreement with in situ observations (Gruber et al., 2022;Gualtieri, 2021;Olauson, 2018;Ramon et al., 2019). Over the North Sea, wind speed underestimates in ERA5 have been reported for some locations (Kalverla et al., 2020), although the homogenous surface of sea areas may improve reanalysis performance when compared with the complex topography of mountainous regions or coastal zones (Gualtieri, 2021). ERA5 has been widely accepted as a suitable data source for use in offshore wind power assessment and modelling (Gualtieri, 2021;Hayes et al., 2021;Soares et al., 2020). Instantaneous zonal and meridional winds are available on a 0.25 latitudelongitude grid, or $30 km, at hourly resolution from ERA5, and these are converted to wind speed before calculating a daily mean, equivalent to the model data. The data are regridded onto the coarser model grid prior to finding the North Sea area mean.
The winter season is split into non-overlapping week-long blocks. This time frame represents sustained lulls which, at present, cannot be accommodated by storage or flexible demand (Jenkins et al., 2018) and so require additional generation from alternative sources (Devlin et al., 2017). Means of 10 m wind speed and the other variables are calculated over each 7-day period, giving 13 data points per winter. A timeseries of the weekly mean wind speeds is shown in Figure 1.

| TESTING THE MODEL AGAINST REANALYSIS-BASED OBSERVATIONS
Under the UNSEEN framework (Thompson et al., 2017), the first step is to determine whether the model ensemble realistically simulates the North Sea near-surface wind speeds during winter. If the model and reanalysis data are shown to be statistically indistinguishable then the F I G U R E 1 Weekly mean 10 m wind speed (ms À1 ) during the winter season over the North Sea domain (inset). For each year, there are 13 data points for ERA5 (black dots) and 13 weeks Â 10 members for the model (grey dots), bias shifted to ERA5 (see Section 3). The threshold for wind drought, 6.75 ms À1 , see Section 4, is marked with a horizontal line, and model wind drought weeks are highlighted in blue. The lower horizontal line marks the minimum from ERA5, 4.46 ms À1 , below which model weekly mean wind speeds are classed as unprecedented.
model ensemble can be regarded as multiple realisations of the current climate, providing a synthetic but realistic event set that is an order of magnitude greater in sample size than that available from the reanalysis. This large event set better samples extremes, and from this, a more robust assessment can be made of wind drought characteristics in the current climate.
To test the model fidelity, the mean, standard deviation, skewness and kurtosis moments of the distribution of 10 m weekly mean winds are compared with the reanalysis. Bootstrapping methods are used to provide confidence intervals on the model statistics (Thompson et al., 2017). The model is repeatedly sampled with replacement across its members, to produce many samples equal in length to the reanalysis timeseries: 60 years Â 13 weeks, that is, for each year, the 13 weeks of winter are taken from a randomly selected ensemble member. This gives a single pseudo-timeseries with 60 Â 13 data points from which the moments of the distribution are calculated. The process is repeated 10,000 times, producing a distribution of each statistic. If the reanalysis value lies within the central 95% of the distribution, then the model and observations are regarded as statistically consistent. The results of the fidelity testing are displayed in Figure 2. They show that, with the exception of the mean (Figure 2b), the model distribution is consistent with the reanalysis. The model wind speeds are biased high, around 0.8 ms À1 with respect to the reanalysis. We correct this bias through taking the difference between the reanalysis and the ensemble mean. The bias through December to February is stable and so we use the seasonal mean bias, and apply that constant mean shift to each ensemble member, as in previous studies (Jain et al., 2020;Kay et al., 2020;Squire et al., 2021).

| WIND DROUGHT SEVERITY: MAGNITUDE, FREQUENCY AND DURATION
From the set of bias-adjusted 7-day mean wind speeds, the chance of unprecedented low winds-below the ERA5 minimum weekly mean of 4.46 ms À1 (Figure 1)is estimated to be $3% each winter, that is, a 1 in $30-year event. During a lull of this magnitude over the period of a week, there would necessarily be high dependence upon alternative power sources to balance the system. It is important to understand that episodes of extremely low winds, lower than any yet recorded, are perfectly possible in the current climate, and that significant capacity to cover severe shortfalls will need to be built into the system. Repeated or prolonged exposure to deficient wind power supply may present additional system impacts, particularly as penetration of wind energy increases. To investigate the frequency and persistence of low wind episodes we define a wind speed threshold for 'wind drought' events to be the 20th percentile of ERA5 daily 10 m winds over the North Sea domain, 6.75 ms À1 . When the weekly mean wind speed falls below that threshold, the week is categorised as a wind drought event (Figure 1). The 20th percentile wind speed threshold is chosen as it represents conditions with substantially reduced wind power. There is a cube relationship between wind speed and wind power (e.g., Harrison & Wallace, 2005;Zeng et al., 2019), implying that a small reduction in wind speed would lead to a large reduction in wind power. In practice, calculating wind power relies on the highest possible spatial and temporal resolution data, as well as accurate turbine specifications. However, as a rough approximation, we estimate that during a F I G U R E 2 Fidelity testing of model wind speeds. Winter 7-day mean 10 m wind speed samples are compared with ERA5 reanalysis over the North Sea domain. Distributions of 7-day mean wind speeds are given in (a), with the moments of the distributions in (b-e). The distributions of the model moments are taken from the 10,000 bootstrapped samples, and the vertical line in each marks the reanalysis statistic, with its position in the model distribution noted in the title.
week when the average wind speed is below this threshold, the total power produced over the week falls by around a half compared with a typical winter week. For a full description of the process used to obtain this estimate of a 50% reduction in wind power, please refer to the Supporting information S1.
The ensemble dataset of 10 m winds is mined for information on the frequency and persistence of wind drought events. Each year there is a > 60% chance of at least 1 week of wind drought during winter (Figure 3a). The maximum number of non-consecutive wind drought weeks observed during a single winter is 4, which occurs twice in 60 years, compared with a maximum of five simulated by the model, which occurs four times in 600 simulated winters (Figure 3a). We find that in both the model and reanalysis, wind drought events are more common in the second half of the winter, with over 40% of wind drought events occurring in the final 4 weeks (February), well above the $30% that would be expected if wind drought events were distributed evenly throughout the winter. In $50% of winters, in both the model and reanalysis, the duration of the longest wind drought event is a single week (Figure 3b). However, continuous multi-week episodes are also possible, and the model samples more prolonged events than have been recorded. In the reanalysis data, wind drought events lasting a maximum of 2 consecutive weeks have occurred, whereas the ensemble indicates that a continuous series of up to 5 weeks of wind drought is possible (Figure 3b). There is a 2.5%-or 1-in-40-chance each winter of a wind drought lasting 3 or more weeks. The 4-week and 5-week wind drought episodes occur only once each in the model set of winters, but such long-lasting events, particularly during the winter season when demand is high, have the potential for significant impacts on the energy sector. Figure 4a illustrates the known coincident relationship (Bloomfield, Brayshaw, et al., 2018;Brayshaw et al., 2012;Thornton et al., 2017) between wind drought and lowtemperature anomalies over the United Kingdom and large parts of Northern Europe. These anomalies are in the coldest 20th percentile of weekly winter temperatures, and therefore increased energy demand would be expected at the very time that supply from wind drops off.

| DYNAMICAL CONDITIONS DURING WIND DROUGHT EVENTS
The large set of wind drought events sampled by the ensemble offers an opportunity to investigate the dynamics of low wind speeds over the North Sea in a more F I G U R E 3 Probability of the frequency and persistence of wind drought episodes. (a) Total number of wind drought weeks per winter in ERA5 (grey) and the model (pink). This is a simple count of the number of wind drought weeks per winter, with no differentiation between events that are isolated single weeks or clustered to form multi-week episodes. As noted in the legend, there are 60 data points in the reanalysis and 600 in the model corresponding to the number of winter seasons in each dataset. (b) Duration of longest wind drought episode during each winter, in weeks. This records the single longest event during each winter, disregarding any other wind drought events that may occur within the same winter. Again, 60 (ERA5) and 600 (model) data points are presented. Inset is a magnified section-note the difference in y-axis scale-showing that the model simulates events of longer duration than observed. robust manner than with the limited events in the reanalysis, and in particular the prolonged (3+ week) events that do not exist in the reanalysis. During wind drought events, high pressure dominates over the North Sea and wider northern European sector (Figure 4b), with negative MSLP anomalies to the south. This pattern projects onto negative North Atlantic Oscillation (NAO)-type conditions, associated with increased blocking, anomalous easterlies and low wind speeds, and the low temperatures over northern Europe seen in Figure 4a.
Notable during the longer (3+ week) duration wind drought events are the low MSLP anomalies in the Aleutian region of the North Pacific (Figure 4b). Intensification of the Aleutian low is a well-documented response to El Niño conditions (Alexander, 1992;Bjerknes, 1969). For these longer wind droughts there is indeed evidence of an El Niño pattern in the SST anomalies (Figure 4c), with some locally significant positive anomalies in the eastern equatorial Pacific and negative anomalies in the west at the 95% confidence level. The El Niño-Southern Oscillation (ENSO) is known to influence the North Atlantic/European sector such that anomalies driven by moderate El Niño events tend to project onto negative NAO conditions, particularly in late winter (Ayarzagüena et al., 2018;Fraedrich & Müller, 1992;Ineson & Scaife, 2009;Toniazzo & Scaife, 2006), which is also the period when wind drought events are found here to be more frequent. Using DJF seasonal mean SSTs in the Niño 3 region as an ENSO index, the ensemble is sub-selected for El Niño conditions, defined here as Niño 3 area-mean SSTs >1 standard deviation above the mean, which yields 71 winters. Of these winters, four contain a wind drought event of three or more continuous weeks. This constitutes an approximate doubling of the chance, from 2.5% to 5.6%, of a wind drought event lasting 3+ weeks, compared with the full set of winters. Having four out of 71 winters with prolonged wind drought events, by chance, would be rare (within the top 10th percentile of 10,000 randomly sampled sets of 71 winters), suggesting that El Niño may have a role to play. Although the sample of events is small, that prolonged wind drought events occur during El Niño is also consistent with what we know about the influence from the tropical Pacific fostering conditions over northern Europe conducive to blocking and weaker winds.
F I G U R E 4 Wider environmental conditions during winter wind drought events. Composite anomalies of (a) 1.5 m air temperature (K), (b) mean sea level pressure (MSLP; hPa) and (c) sea surface temperature (SST; K). Wind drought events of one or more (1+) weeks' duration are displayed on the top row, two or more (2+) weeks' duration in the middle and three or more (3+) weeks on the bottom row. Mean wind speeds during each week of the event satisfy the wind drought threshold, and composite anomalies are found by taking a mean across the weeks of the event and all events of the specified duration. The number of events in each category is noted in the title. The data have been linearly detrended. Stippling indicates significance at the 95% confidence level. This is calculated by randomly selecting a set of weeks to match the wind drought event count and duration, for example, to calculate 2+ week significance: randomly select 60 Â 2 weeks, 13 Â 3 weeks, 1 Â 4 weeks and 1 Â 5 weeks, make a composite mean, then repeat the process 10,000 times. At each gridpoint, the position of the wind drought composite value within the distribution of random composites is calculated, and if it lies outside the central 2.5%-97.5%, it is marked as significant.

| DISCUSSION
This study quantifies the chances of wind drought events over the North Sea region in the current climate by using a large, initialised ensemble of climate model simulations, which allows a more robust estimate of their likelihood compared with using observations alone. We find that more severe winter wind drought episodes than experienced to date are possible in the current climate. The model simulates unprecedented, record low weekly mean wind speeds, with a likelihood of $3% each winter, that is, a 1 in $30-year event. In addition to the magnitude of deficient winds, other aspects of wind drought severity include event frequency and duration. To investigate this, the 20th percentile of the reanalysis daily wind speeds was chosen as a threshold for wind drought events. While the effect of the threshold on resulting wind power would require wind speed data at higher spatial and temporal resolution to determine, we estimate that average wind power over the broad North Sea region during a wind drought week would be approximately half of that expected from a typical winter week. The North Sea domain used in this study covers the exclusive economic zones (EEZs) of several countries and information on wind drought events is likewise at this scale. Planning at the scale of EEZs or individual wind farms would require further refinement of the analysis presented here, with a more targeted assessment of model performance against local reanalysis or observational data, and characterisation of smaller-scale wind drought events.
We find that a maximum of 5 weeks of non-consecutive wind drought is possible over the 13-week winter season, compared with a maximum of four currently seen in the reanalysis. A potentially more significant finding is the heightened risk of prolonged wind drought events. There is a 1 in 40-year chance that wind droughts will last longer than the current record of 2 weeks and events of 5 weeks' duration are possible. The characteristics of wind drought events presented here could be useful in stress testing existing mitigation measures and planning future alternative supply. This could complement new tools and datasets that are being developed to assess resilience in future electricity systems (Dawkins et al., 2021). An interesting extension to this study would be to consider solar energy potential during wind drought events (Bett & Thornton, 2016), which could be helpful in characterising wider 'renewable energy drought' (Jurasz et al., 2021;Otero et al., 2022). More information on the spatial structure of anomalies in winds and irradiance could inform requirements for European power system interconnectivity (Otero et al., 2022).
An examination of the dynamics associated with wind drought events shows that, as expected, an area of high pressure is located over the North Sea and the wider region across northern Europe. These conditions bring low temperatures to large parts of Europe, which would stimulate increased demand for electricity in several countries simultaneously at the time when North Sea wind capacity is reduced. In the Aleutian region, low pressure is intensified during drought weeks, which in the longest duration events are associated with El Niño. There is a doubling (from 2.5% to $5%) of the likelihood of prolonged (3+ week) wind drought events occurring during El Niño compared with the full set of winters. The sample size of very long wind drought episodes is small and further investigation would benefit from the use of additional climate model ensembles. Nevertheless, the simulations of more extreme wind droughts than have yet been observed demonstrates a benefit of using large ensembles over observations alone in assessing the conditional likelihood of wind drought events. That an El Niño event could promote conditions conducive to prolonged low winds over the North Sea is consistent with the known teleconnection between ENSO and the NAO. Given the long-range predictability of El Niño events (Barnston et al., 2019) it may also be worth using seasonal wind forecasts to assess the risk of wind drought in upcoming winter seasons.
Projections of future changes in the midlatitude jet streams and storm tracks are uncertain (Pryor et al., 2020;Shaw et al., 2016) and hence likewise is the outlook for windiness over the North Sea and northern Europe. Global and regional declines in near surface wind speeds have been reported, with some degree of recovery in recent years (Dunn et al., 2022;Zeng et al., 2019). However, over the next decades at least, and certainly on net zero timescales, climate variability is expected to be much larger than any climate change signal in wind speeds (Pryor et al., 2020). Our approach helps to provide better information on variability in winds that could be useful in energy planning around Europe and further applied globally.