Stratospheric influences on subseasonal predictability of

KIT – The Research University in the Helmholtz Association www.kit.edu

1) Institute of Meteorology and Climate Research, Department Troposphere Research, KIT, Germany 2) Institute for Atmospheric and Climate Science, ETH Zurich, Switzerland 3) AXPO Solutions AG, Switzerland

“The ‘Beast from the East fails’ to bite”

Alex Froley, Gas Market Analyst, 18 March 2019

Stratospheric influences on subseasonal predictability of European energy-industry-relevant parameters Dominik Büeler 1,2 / Remo Beerli 3,2, Heini Wernli 2, Christian M. Grams 1,2

S2S / TIGGE Workshop ECMWF | April 2019

Institute of Meteorology and Climate Research (IMK-TRO) Dominik Büeler | [email protected] 2

Stratospheric and tropospheric annularmode variations are sometimes independent ofeach other, but (on average) strong anomaliesjust above the tropopause appear to favor tro-pospheric anomalies of the same sign. Oppos-ing anomalies as in December 1998 (Fig. 1) arepossible, but anomalies of the same sign dom-inate the average (Fig. 2).

To examine the tropospheric circulationafter these extreme events, we define weakand strong vortex “regimes” as the 60-dayperiods after the dates on which the !3.0 and"1.5 thresholds were crossed. Our results arenot sensitive to the exact range of days usedand do not depend on the first few days afterthe “events.” We focus on the average behav-ior during these “weak vortex regimes” and

“strong vortex regimes,” as characterized bythe normalized AO index (22). The averagevalue (1080 days) during weak vortex re-gimes is !0.44, and "0.35 for strong vortexregimes (1800 days). The large sample sizescontribute to the high statistical significanceof these averages (23). During the weak andstrong vortex regimes the average surfacepressure anomalies (Fig. 3) are markedly likeopposite phases of the AO (11) or NAO (14),with the largest effect on pressure gradientsin the North Atlantic and Northern Europe.

The probability density functions (PDFs) ofthe daily normalized AO and NAO indices (24)during weak and strong vortex regimes arecompared in Fig. 4. More pronounced than theshift in means are differences in the shapes of

the PDFs, especially between the tails of thecurves. Values of AO or NAO index greaterthan 1.0 are three to four times as likely duringstrong vortex regimes than weak vortex re-gimes. Similarly, index values less than !1.0are three to four times as likely during weakvortex regimes than strong vortex regimes. Val-ues of the daily AO index greater than 1.0 andless than !1.0 are associated with statisticallysignificant changes in the probabilities ofweather extremes such as cold air outbreaks,snow, and high winds across Europe, Asia, andNorth America (25). The observed circulationchanges during weak and strong vortex regimesare substantial from a meteorological viewpointand can be anticipated by observing the strato-sphere. These results imply a measure of pre-dictability, up to 2 months in advance, for AO/NAO variations in northern winter, particularlyfor extreme values that are associated with un-usual weather events having the greatest impacton society.

Since the NAO and AO are known to mod-ulate the position of surface cyclones across theAtlantic and Europe, we examine the tracks ofsurface cyclones with central pressure less than

1998 - 1999 Northern Annular Mode

Oct Nov Dec Jan Feb Mar Apr

1

10

100

1000

hPakm

0

10

20

30

40

50

Fig. 1. Time-height development of the northern annular mode during the winter of 1998–1999.The indices have daily resolution and are nondimensional. Blue corresponds to positive values(strong polar vortex), and red corresponds to negative values (weak polar vortex). The contourinterval is 0.5, with values between !0.5 and 0.5 unshaded. The thin horizontal line indicates theapproximate boundary between the troposphere and the stratosphere.

Composite of 18 Weak Vortex Events

-90 -60 -30 0 30 60 90Lag (Days)

hPa

A10

30

100

300

1000

km

0

10

20

30

Composite of 30 Strong Vortex Events

-90 -60 -30 0 30 60 90Lag (Days)

hPa

B10

30

100

300

1000

km

0

10

20

30

Fig. 2. Composites of time-height development of the northern annular mode for (A) 18 weakvortex events and (B) 30 strong vortex events. The events are determined by the dates on whichthe 10-hPa annular mode values cross –3.0 and"1.5, respectively. The indices are nondimensional;the contour interval for the color shading is 0.25, and 0.5 for the white contours. Values between!0.25 and 0.25 are unshaded. The thin horizontal lines indicate the approximate boundarybetween the troposphere and the stratosphere.

Weak Vortex Regimes

-1

-10

0

0

2

4

A

Strong Vortex Regimes

-1

0

00

0

2

B

Fig. 3. Average sea-level pressure anomalies(hPa) for (A) the 1080 days during weak vortexregimes and (B) the 1800 days during strongvortex regimes.

R E P O R T S

19 OCTOBER 2001 VOL 294 SCIENCE www.sciencemag.org582

the wind speed at 10 m (in m s!1), following theapproach of Manwell et al (2010) in which wind poweris primarily a function of the volume throughput of airdriving the blades of a turbine. r is the air density,computed here using the ideal gas law (r = P/RT),where P, T and R are the pressure, temperature andspecific gas constant for dry air (287.058 J kg!1 K!1)respectively. For reasons of data availability, we usepressure at mean sea level and temperature at 2 m.

Seasonal means of power density were produced foreach gridpoint by averaging over power densitiescomputed using the daily-mean output from theGloSea5 hindcasts and 6-hourly means from the ERAInterim verifying reanalysis. Data at finer temporalresolution (which was not available) would have beenpreferred for this analysis, since the mean of U3 is onlyequal to the cube of daily mean U in the absence ofsub-daily variability. The correlation with power

0.3 0.4 0.5 0.6 0.7 0.8

Figure 1. Correlation between GloSea5 ensemble mean and ERA Interim sea level pressure for DJF compiled from 20 years ofsimulation. Mask (white areas) applied to correlations not significantly greater than zero at 10% level.

Between ERAI NAO andERAI 10m wind speed

Between ERAI NAO andERAI 1.5m air temperature

Between G5 NAO andG5 10m wind speed (480 samples)

Between G5 NAO andG5 1.5m air temperature (480 samples)

Between G5 EM NAO andG5 EM 1.5m air temperature (20 samples)

Between G5 EM NAO andG5 EM 10m wind speed (20 samples)

–0.8 –0.6 –0.4 –0.2 0.0 0.2 0.4 0.6 0.8

Figure 2. Correlations between NAO on winter near-surface wind speed (left column) and temperature (right column). Observed(ERA Interim) relationships are shown in the top row. Middle row shows ensemble member relationships in hindcasts. Bottom rowshows ensemble mean relationships in hindcasts. Mask (white areas) applied to correlations not significant at 10% level.

Environ. Res. Lett. 12 (2017) 024002

3

State of the stratospheric polar vortex (SPV) as a direct source of subseasonal predictability for European energy industry?

e.g., Baldwin & Dunkerton, 2001, SCI; Tripathi et al., 2015, ERL;

Charlton-Perez et al., 2018, QJRMS

NOAA EnergyWay NOAA

e.g., Clark et al., 2017, ERL; Brayshaw et al., 2011, RE

Motivation | Polar vortex – weather regimes – wind power

2 April 2019 | ECMWF

Institute of Meteorology and Climate Research (IMK-TRO) Dominik Büeler | [email protected]

Strength of SPV   (Δ Z@150hPa)60°-90°N from ERA-Interim   Daily, DJF, 1985 – 2014

  Wind power generation for every European country

Renewables.ninja dataset (Staffel & Pfenninger, 2016, ENE; www.renewables.ninja)   Daily month-ahead average, DJF, 1985 – 2014

3

Data | Statistical forecast

Beerli et al., 2017, QJRMS



Results | Simple 3-categorical statistical forecast 3034 R. Beerli et al.

in Figure 6(a)) show a significantly stronger than normal polarvortex from about 40 days before the event, but there are nosigns of gradual downward propagation of a stratospheric signaltowards the surface (consistent with the findings of Limpasuvanet al. (2005) in their figure 1(b)). In contrast, for weak polar-vortex events (Bin100

98 in Figure 6(e)), the composite mean φ′PC

shows the typical characteristics of sudden stratospheric warmings(SSWs: e.g. Baldwin and Dunkerton, 2001; Limpasuvan et al.,2004): Positive φ′

PC signals at 10 hPa 15 days prior to the eventsare propagating downwards and exhibit a surface signal that ispresent until 30 days after the events. The positive φ′

PC in thestratosphere from 25 to 20 days before the weak polar-vortexevents is also in line with a tropospheric precursor of SSWs,which has been documented by various previous studies (e.g.Baldwin and Dunkerton, 2001; Limpasuvan et al., 2004). Giventhe good agreement between the geopotential height anomalies ofweak polar-vortex events in Bin100

98 and the SSWs documented inthe existing literature, it is likely that these weakest polar-vortexevents are often part of the life cycle of SSWs.

So far, we have shown that the lower stratospheric circulationhas an impact on month-ahead wind electricity generation indifferent European countries when φ′

PC150 is far from its clima-tological mean. This relation between the lower stratosphere andmonth-ahead wind electricity generation is a result of long-livedperiods with NAO+ or NAO− conditions, which occur due tocoupled evolution of the troposphere and the stratosphere. Wenext investigate the implications of these results for the pre-dictability of month-ahead wind electricity generation in Europe.

3.3. Skill of month-ahead wind electricity forecasts based on thestate of the lower stratospheric circulation

We use simple statistical forecasts to investigate the predictive skillarising from the statistical relationship between the state of thelower stratosphere and month-ahead wind electricity generationin different countries of Europe. φ′

PC150 is used as a predictor forprobabilistic three-categorical forecasts of "CF31d. We predictprobabilities for the three climatological terciles of "CF31d(lower, middle and upper third of observed "CF31d) using thefollowing approach. For each daily value of φ′

PC150, we determineits percentile value within the overall φ′

PC150 distribution (Pinit).We then take all φ′

PC150 values that fall within +/− 10% of thispercentile (from Pinit− 10 to Pinit+10) as the basis to determine theprobabilities for the three terciles of "CF31d from all pairs ofφ′

PC150 and "CF31d within this range. The following example fora forecast of "CF31d for Sweden illustrates this approach. On 19January 1986, φ′

PC150 was − 80.8 m, which is the 18.02th percentileof the climatological distribution (Pinit = P18.02 = − 80.8 m).We then use all pairs of φ′

PC150 and "CF31d in the datasetfrom the 8.02th percentile (Pinit− 10 = P8.02 = − 119.4 m) to the28.02th percentile (Pinit+10 = P28.02 = − 53.7 m) to derive theprobabilistic forecast (days in the same season are left out toprevent artificial skill). Of all the φ′

PC150 –"CF31d pairs betweenthe 8.02th and 28.02th percentiles, 10.6% of the "CF31d values arein the lower tercile, 23.1% in the middle tercile and 66.3% in theupper tercile. Therefore the month-ahead forecast for "CF31dfrom 19 January 1986 is a 10.6%/23.1%/66.3% probability of"CF31d being in the lower/middle/upper tercile. If φ′

PC150 < P10,we use the φ′

PC150 –"CF31d pairs for φ′PC150 < Pinit+10 (and

similarly if φ′PC150 > P90). For each country, these forecasts are

made for each winter (DJF) calendar day from 1985 to 2014 ina ‘leave-the-current-season-out’ sense, i.e. only φ′

PC150 –"CF31dpairs of winter seasons different from the current season areconsidered, in order to prevent artificially skilful forecasts.

We perform the statistical forecast for each winter (DJF) dayin the dataset for the eight European countries with the highestinstalled capacity of wind power (Germany, Spain, UK, France,Italy, Sweden, Poland and Denmark) for lagged 31 day windows(lag 0 in Figure 7 shows the RPSS for wind-power forecastsaveraged over 0–30 days ahead, lag 1 the RPSS for 1–31 days ahead

Figure 7. The RPSS of three-categorical statistical forecasts of month-aheadaverage wind electricity generation as a function of lead time for eight Europeancountries. The lead time on the x-axis indicates the start of the forecast 30 dayperiod. For instance, the RPSS at a lead time of 15 days shows the skill offorecasts for wind electricity generation averaged over 15–44 days ahead. Theshaded colours show the confidence interval for the RPSS values derived by thebootstrapping approach described in the text for Sweden (blue), Germany (red)and Spain (yellow).

and so on). Additionally, we apply a bootstrapping approachin order to test the sampling sensitivity of the skill scores ofthese forecasts. In 200 repetitions, we randomly sample 80%of the winters and calculate the RPSS of each repetition. The10% and 90% percentiles among these 200 RPSS values are theconfidence intervals displayed in shaded colours in Figures 7and 8. Comparing the RPSS for the eight countries mentionedabove (Figure 7) reveals three groups of countries with similarlevels of predictability.

(1) High predictability of "CF31d: Sweden and Denmark (inblue in Figure 7), RPSS≈ 0.2 for lead time 0.

(2) Moderate predictability of "CF31d: Germany, UK andPoland (in red in Figure 7), RPSS≈ 0.1 for lead time 0.

(3) No predictability of "CF31d: Spain, France and Italy (inyellow in Figure 7), RPSS < 0 for lead time 0.

These groups are in line with the findings derived fromFigures 2 and 3. Sweden and Denmark are located in the centreof the high (low) wind corridor when φ′

PC150 is strongly negative(positive), which makes it very likely that these countries willindeed experience above (below) normal CF in the following30 days. The countries with moderate predictability (Germany,UK and Poland) are situated at the southern edge of these high(low) wind corridors, which again makes it likely for them toexperience above (below) normal CF, but this is less certain thanfor the Nordic countries –just a subtle change in the synopticset-up (which is not constrained by φ′

PC150) may change theCF outcome. Hence the skill of the statistical forecast for thesecountries is notably lower than for the Nordic countries. Forthe Southern European countries (Spain, Italy and France), theRPSS is even negative, which means that the predictions basedon the state of the lower stratosphere are worse than simplyforecasting the climatological distribution of "CF31d. Thesecountries are situated well outside the areas with significantlypositive or negative wind-speed anomalies shown in Figures 2and 3. Therefore, there is no sufficiently strong signal related to thestratospheric circulation, which could be exploited to issue skilfulforecasts. If the lead time for the forecasts of "CF31d is increased,the skill levels get gradually lower but, for both high and moderatepredictability countries, the statistical forecasts 15–45 days aheadare also better than the climatological reference forecast. The skillthat we find here for month-ahead wind electricity generation isslightly higher than the skill levels found by Karpechko (2015)

c⃝ 2017 Royal Meteorological Society Q. J. R. Meteorol. Soc. 143 : 3025–3036 (2017)

Beerli et al., 2017, QJRMS

Stratospheric Predictability for Wind Energy 3035

Figure 8. (a) The RPSS of three-categorical statistical forecasts for wind electricity generation averaged over 0–30 days ahead for forecasts initialized with a φ′PC150

value in the bins indicated on the x-axis. (b) Same as (a), but for forecasts 5–35 days ahead. (c) Same as (a), but for 10–40 days ahead. The shaded colours show theconfidence interval for the RPSS values derived by the bootstrapping approach described in the text.

(their figure 7) for month-ahead temperature forecasts for selectedcities of Europe. For instance, our RPSS for month-ahead windelectricity generation in Sweden is about 0.04 higher than for thetemperature forecast of Stockholm by Karpechko (2015).

The results presented so far suggest that a statistical forecastbased on φ′

PC150 might have different skill depending on whether itis initialized during normal or very anomalous φ′

PC150. Thereforewe recalculated the RPSS for four countries (Sweden, Germany,Spain and France) representative of different subregions ofEurope, stratified according to the 12 φ′

PC150 bins used earlierin this study (Figure 8(a)).

For many countries, the skill of "CF31d forecasts differsstrongly among the bins (Figure 8(a)). For Sweden, highlyanomalous states of the lower stratosphere (Bin2

0, Bin100 , Bin20

10,Bin90

80, Bin10090 and Bin100

98 ) exhibit high skill for month-ahead windelectricity generation (> 0.7 for Bin2

0). As expected, the RPSSdrops close to zero for the middle bins of φ′

PC150. Germany hasits peak predictability in Bin20

10, with a RPSS of just below 0.3,but predictability drops for bins with more negative φ′

PC150. Thisis consistent with a further northward position of the NAO+dipole and the high surface wind corridor associated with Bin2

0and Bin10

0 , which makes it uncertain to what degree Germanyis affected by the high surface winds connected to the NAO+configuration (Figures 2 and 3). There is a similar tendency forthe positive φ′

PC150 bins, but the skill is only reduced for Bin10098 .

Although there is an overall negative RPSS for Spain (Figure 7),forecasts are quite strongly skilful for extremely negative φ′

PC150,reaching above 0.4 for Bin2

0, when Spain is under the influenceof a strong ridge, leading to predominantly calm conditions. Thisridge is part of the strong NAO+ dipole located far north duringstrongly negative φ′

PC150. As a contrast, the skill for stronglypositive φ′

PC150 is clearly lower than that of the climatologicalforecast, almost reaching an RPSS of − 0.2, which is due tothe two distinct clusters of "CF31d in Bin100

98 discussed above(Figure 1(b)). It may be the case that these two clusters are aproduct of the small sample size of extremely positive φ′

PC150 andthis strongly negative RPSS may not reflect the fact that φ′

PC150in Bin100

98 indeed leads to a particularly unpredictable situationfor wind in Spain. A longer dataset would clearly be helpful toaddress this further. For France, the skill is close to climatologicalvalues and even falls well below for extremely positive/negativeφ′

PC150. France is located in between any significant surface windsignals associated with φ′

PC150 anomalies (see Figures 2– 5) andtherefore the statistical forecasts cannot exploit any clear signalsfor France. The overall shape of the forecast skill as a functionof φ′

PC150 is in all countries quite similar for "CF31d 5–35 daysand 10–40 days ahead (Figures 8(b) and (c)), but with graduallydecreasing skill values as the lead time increases. An interestingexception is Germany, where the RPSS is highest in Bin100

90 for10–40 days ahead.

Overall, the results presented in this section suggest thatanomalous flow conditions in the lower stratosphere offersubseasonal-range predictability for month-ahead wind electricitygeneration in Europe. This predictability is strongest in the Nordiccountries, but is also present in Germany/UK and to a lesser extentin Iberia. It is also highly dependent on the state of the lowerstratosphere. When the lower stratospheric circulation is far fromits climatological mean, the predictability is particularly high forselected countries (especially for the Nordics).

4. Summary and conclusions

We discussed the relationship between the lower stratospheric cir-culation and month-ahead wind electricity generation in differentcountries of Europe. As an indicator for the circulation in thelower stratosphere, we used φ′

PC150, i.e. the 150 hPa geopotentialheight anomaly averaged from 60◦N to the North Pole. The shapeof this relationship differs between countries. Scandinavian coun-tries generally show above normal month-ahead wind electricitygeneration for anomalously strong stratospheric circulation (i.e.negative φ′

PC150) and vice versa for a weak stratospheric circula-tion. Germany shows a similar dependence on φ′

PC150, but withdecreasing wind electricity generation for an extremely strongstratospheric circulation. In Iberia, wind electricity generationis reduced during anomalously strong stratospheric circulation,but there is no clear impact if the stratospheric flow is weak.

We showed that periods of coupling between the stratosphericand tropospheric circulation, which lead to persistent phases ofstrong NAO+ and NAO− , are the reason for these relationships.Such prolonged NAO phases lead to strong anomalies innear-surface wind, in particular in the northern half of Europe.The exact position of the 500 hPa geopotential height anomalydipole associated with these prolonged NAO periods is criticalfor wind electricity generation, in particular in Central Europe.Its position depends on the strength of the stratosphericgeopotential height anomalies, consistent with the mechanismproposed by Ambaum and Hoskins (2002), which involveslarge-scale adjustment to a stratospheric PV anomaly. Thismechanism favours a more northern position of the NAO+500 hPa geopotential height anomaly dipole for very strongstratospheric polar vortices, as observed in our study, which thenalso explains why wind electricity generation in Germany dropstowards normal levels for extremely negative φ′

PC150.For the reasons outlined above, the lower stratosphere is

a source of predictability for month-ahead wind electricitygeneration in many countries in Europe. The predictabilitygained from the lower stratosphere alone is highest in theNordic countries and weakest in Southern Europe. This hasbeen confirmed with simple statistical forecasts using φ′

PC150 aspredictor for categorical (terciles) month-ahead wind electricity

c⃝ 2017 Royal Meteorological Society Q. J. R. Meteorol. Soc. 143: 3025–3036 (2017)

0 – 30 days ahead

Stronger SPV

Weaker SPV

How does this mechanism influence the skill of subseasonal numerical weather models?



Subseasonal ECMWF model (www.s2sprediction.net)   2 reforecasts / week, DJF, 1995 – 2017   11 ensemble members

  Fields calculated for each reforecast

Data | Numerical forecast

  Strength of SPV = (Δ Z@100hPa) 60°-90°N   At forecast initial time

  (Δ 10m wind) European Countries   (Δ 2m temperature) European Countries   (Δ precipitation) European Countries   Average over 1 month lead time

5 2 April 2019 | ECMWF


Results | Regional model skill pattern 10m wind 2m temperature Precipitation

Anomalies after 10% strongest SPV states – anomalies after 10% weakest SPV states



Results | Model skill for 10m wind

Anomalies after 2% strongest SPV states Anomalies after 2% weakest SPV states

S2S ERA S2S ERA



Results | Model skill for 2m temperature

S2S ERA S2S ERA




Results | Model skill for 2m temperature

S2S ERA S2S ERA




  Strong spatial variability of statistical and numerical model skill for month-ahead prediction of wind power generation / surface weather in Europe

  Reason: anomalous SPV states at forecast initial time lead to persistent NAO-like anomaly patterns à model skill for countries located in affected regions tends to be enhanced

  However, model skill increase much more significant and robust after strongest SPV states than after weakest SPV states (~ SSWs), which even lead to significant skill reduction for certain countries (particularly T@2m)

  Implications

  Energy meteorology cannot rely on enhanced predictability after weakest (~ SSWs) but more after strongest SPV states

  Regional SSW response in S2S models needs to be improved

10

Conclusions 3034 R. Beerli et al.

in Figure 6(a)) show a significantly stronger than normal polarvortex from about 40 days before the event, but there are nosigns of gradual downward propagation of a stratospheric signaltowards the surface (consistent with the findings of Limpasuvanet al. (2005) in their figure 1(b)). In contrast, for weak polar-vortex events (Bin100

98 in Figure 6(e)), the composite mean φ′PC

shows the typical characteristics of sudden stratospheric warmings(SSWs: e.g. Baldwin and Dunkerton, 2001; Limpasuvan et al.,2004): Positive φ′

PC signals at 10 hPa 15 days prior to the eventsare propagating downwards and exhibit a surface signal that ispresent until 30 days after the events. The positive φ′

PC in thestratosphere from 25 to 20 days before the weak polar-vortexevents is also in line with a tropospheric precursor of SSWs,which has been documented by various previous studies (e.g.Baldwin and Dunkerton, 2001; Limpasuvan et al., 2004). Giventhe good agreement between the geopotential height anomalies ofweak polar-vortex events in Bin100

98 and the SSWs documented inthe existing literature, it is likely that these weakest polar-vortexevents are often part of the life cycle of SSWs.

So far, we have shown that the lower stratospheric circulationhas an impact on month-ahead wind electricity generation indifferent European countries when φ′

PC150 is far from its clima-tological mean. This relation between the lower stratosphere andmonth-ahead wind electricity generation is a result of long-livedperiods with NAO+ or NAO− conditions, which occur due tocoupled evolution of the troposphere and the stratosphere. Wenext investigate the implications of these results for the pre-dictability of month-ahead wind electricity generation in Europe.

3.3. Skill of month-ahead wind electricity forecasts based on thestate of the lower stratospheric circulation

We use simple statistical forecasts to investigate the predictive skillarising from the statistical relationship between the state of thelower stratosphere and month-ahead wind electricity generationin different countries of Europe. φ′

PC150 is used as a predictor forprobabilistic three-categorical forecasts of "CF31d. We predictprobabilities for the three climatological terciles of "CF31d(lower, middle and upper third of observed "CF31d) using thefollowing approach. For each daily value of φ′

PC150, we determineits percentile value within the overall φ′

PC150 distribution (Pinit).We then take all φ′

PC150 values that fall within +/− 10% of thispercentile (from Pinit− 10 to Pinit+10) as the basis to determine theprobabilities for the three terciles of "CF31d from all pairs ofφ′

PC150 and "CF31d within this range. The following example fora forecast of "CF31d for Sweden illustrates this approach. On 19January 1986, φ′

PC150 was − 80.8 m, which is the 18.02th percentileof the climatological distribution (Pinit = P18.02 = − 80.8 m).We then use all pairs of φ′

PC150 and "CF31d in the datasetfrom the 8.02th percentile (Pinit− 10 = P8.02 = − 119.4 m) to the28.02th percentile (Pinit+10 = P28.02 = − 53.7 m) to derive theprobabilistic forecast (days in the same season are left out toprevent artificial skill). Of all the φ′

PC150 –"CF31d pairs betweenthe 8.02th and 28.02th percentiles, 10.6% of the "CF31d values arein the lower tercile, 23.1% in the middle tercile and 66.3% in theupper tercile. Therefore the month-ahead forecast for "CF31dfrom 19 January 1986 is a 10.6%/23.1%/66.3% probability of"CF31d being in the lower/middle/upper tercile. If φ′

PC150 < P10,we use the φ′

PC150 –"CF31d pairs for φ′PC150 < Pinit+10 (and

similarly if φ′PC150 > P90). For each country, these forecasts are

made for each winter (DJF) calendar day from 1985 to 2014 ina ‘leave-the-current-season-out’ sense, i.e. only φ′

PC150 –"CF31dpairs of winter seasons different from the current season areconsidered, in order to prevent artificially skilful forecasts.

We perform the statistical forecast for each winter (DJF) dayin the dataset for the eight European countries with the highestinstalled capacity of wind power (Germany, Spain, UK, France,Italy, Sweden, Poland and Denmark) for lagged 31 day windows(lag 0 in Figure 7 shows the RPSS for wind-power forecastsaveraged over 0–30 days ahead, lag 1 the RPSS for 1–31 days ahead

Figure 7. The RPSS of three-categorical statistical forecasts of month-aheadaverage wind electricity generation as a function of lead time for eight Europeancountries. The lead time on the x-axis indicates the start of the forecast 30 dayperiod. For instance, the RPSS at a lead time of 15 days shows the skill offorecasts for wind electricity generation averaged over 15–44 days ahead. Theshaded colours show the confidence interval for the RPSS values derived by thebootstrapping approach described in the text for Sweden (blue), Germany (red)and Spain (yellow).

and so on). Additionally, we apply a bootstrapping approachin order to test the sampling sensitivity of the skill scores ofthese forecasts. In 200 repetitions, we randomly sample 80%of the winters and calculate the RPSS of each repetition. The10% and 90% percentiles among these 200 RPSS values are theconfidence intervals displayed in shaded colours in Figures 7and 8. Comparing the RPSS for the eight countries mentionedabove (Figure 7) reveals three groups of countries with similarlevels of predictability.

(1) High predictability of "CF31d: Sweden and Denmark (inblue in Figure 7), RPSS≈ 0.2 for lead time 0.

(2) Moderate predictability of "CF31d: Germany, UK andPoland (in red in Figure 7), RPSS≈ 0.1 for lead time 0.

(3) No predictability of "CF31d: Spain, France and Italy (inyellow in Figure 7), RPSS < 0 for lead time 0.

These groups are in line with the findings derived fromFigures 2 and 3. Sweden and Denmark are located in the centreof the high (low) wind corridor when φ′

PC150 is strongly negative(positive), which makes it very likely that these countries willindeed experience above (below) normal CF in the following30 days. The countries with moderate predictability (Germany,UK and Poland) are situated at the southern edge of these high(low) wind corridors, which again makes it likely for them toexperience above (below) normal CF, but this is less certain thanfor the Nordic countries –just a subtle change in the synopticset-up (which is not constrained by φ′

PC150) may change theCF outcome. Hence the skill of the statistical forecast for thesecountries is notably lower than for the Nordic countries. Forthe Southern European countries (Spain, Italy and France), theRPSS is even negative, which means that the predictions basedon the state of the lower stratosphere are worse than simplyforecasting the climatological distribution of "CF31d. Thesecountries are situated well outside the areas with significantlypositive or negative wind-speed anomalies shown in Figures 2and 3. Therefore, there is no sufficiently strong signal related to thestratospheric circulation, which could be exploited to issue skilfulforecasts. If the lead time for the forecasts of "CF31d is increased,the skill levels get gradually lower but, for both high and moderatepredictability countries, the statistical forecasts 15–45 days aheadare also better than the climatological reference forecast. The skillthat we find here for month-ahead wind electricity generation isslightly higher than the skill levels found by Karpechko (2015)

c⃝ 2017 Royal Meteorological Society Q. J. R. Meteorol. Soc. 143 : 3025–3036 (2017)


Documents

Stratospheric influences on subseasonal predictability of