Status: Ongoing analysis Finalized Material from: Linus, Ervin, Tim H., Ivan, Fernando, Mark
Discussed in the following Daily reports:
...
http://intra.ecmwf.int/daily/d/dreport/2013/10/31/sc/
This case is documented in ECMWF Newletter 139 (Hewson et al): http://www.ecmwf.int/sites/default/files/Newsletter_139_smaller.pdf
1. Impact
On the 28 October a wind storm hit north-western Europe. In total 14 people were killed across Europe (http://www.bbc.co.uk/news/world-europe-24705734). The main affected countries where France, UK, The Netherlands, Germany, Denmark and Sweden.
In Denmark the highest ever wind gust (for Denmark) was measured on Kegnäs on Als (53 m/s).
The figures above shows the return period of the event from wind gust observations (left, limited number of stations) and HRES forecast compared to ERA Interim climatology (right). The two estimates agrees that the 10 year return period was exceeded along the North Sea coast.
On the 24 October, UKMO issued On the 24 October, UKMO issued an amber warning (see figure above) for southern England for wind speeds. They classifed the potential impact as the highest category and the likelihood of the event as 2 out of 4. In their warning matrix it resulted in an amber warning.
2. Description of the event
Sequence is the Met Office surface analysis, with fronts, covering the period when the maximum gusts occurred . It shows the storm to be a small and rather innocuous looking feature! Note also that some of the charts - eg when the low was over Sweden - show a Keyser-Shapiro type structure, with a bent back cold front rather than an occlusion - implying a warm core.
...
This figure shows the same as the figure above but zoomed in over Denmark.
3. Mesoscale Structure of the Storm
Animation - IR | Reading wind trace | |
Animation - Reading WCB SJ phase |
Animation - Reading SJ WCB phase
Animation - time of Reading max SJ gust
| |
Part of Sting Jet conceptual model | Damage Swathe conceptual model |
The panels above relate to the mesoscale structure of the storm as it passed over the UK (and beyond). We refer here to the warm conveyor belt (WCB), sting jet (SJ) and cold conveyor belt (CCB). Refer also to fronts on the Met Office analysis charts at the start of this page - the WCB is in the warm sector, the sting jet tends to be near to the tip of the occlusion / bent back front, and the CCB on the western and southern flank of this wraparound front.
...
For hour by hour plots of MSLP and wind gusts from SMHI surface analysis (MESAN), see http://www.smhi.se/klimatdata/meteorologi/vind/storm-okt-2013 .
4. Predictability (including 40r1 E-suite evaluation)
4.
...
0 Data assimilation
The figures above shows MSLP forecast valid 28 October 00 UTC. The black contours shows forecasts from 27 Oct 00 UTC (left), 27 Oct 12 UTC (mid) and 28 Oct 00 UTC (right). The red contours are forecasts initialised 12 hours earlier. The figures therefore illustrates the shift between two consecutive forecasts. (The black contours become red in consecutive plot.) For the 27 Oct 00 UTC, the cyclone was shifted to the east and had not developed as much as in the previous forecast. For the forecast from 12 UTC (mid), the cyclone is even more shifted to east. For the last plot, where we compare the analysis from 28 Oct 00 UTC with the forecast from 12 hours earlier, the cyclone is shifted somewhat back towards west, although not a deep as in 26 Oct 12 UTC (red contour in the right plot). These shifts in the forecast had a big impact on the forecasts for the storm (see Section 4.2).
The figures above shows the same plots but fore the e-suite. The there seems to be more consistency between the forecasts.
4.1 HRES
4.1.1 O-Suite
Sequence of forecasts below (left and second left columns) shows 48h maximum gusts from successive HRES forecasts from DT 00UTC 21st through to 00UTC 27th for VT period 06UTC 27th to 06UTC 29th (on 12up image left click, then right click -> view image to get full resolution). Note how the early forecasts jumped around, as one might expect. Later on very extreme gusts were forecast for areas SW of the UK, and as the event approached this zone of extreme gusts migrated its way eastwards, which is perhaps not what one would want/expect. End of period mslp is also shown.
By 'adding together' equivalent plots from two very short range forecasts we can provide a model-based pseudo-verification to compare the above with. These are below in rapid animation and two-up formats (covering the period 18Z 27th to 06Z 29th). No gusts of note were forecast for the 06Z-18Z 27th period.
Forecasts | Forecasts |
Model-based verification
| Model-based verification
|
---|
4.1.2 O-Suite vs E-suite
The figures below shows forecasts of MSLP valid for 28 Oct 12 UTC and maximum wind gust during the 28 October. The o-suite is plotted in the left column and the e-suite in the right.
The figure above shows forecasts initialised 28 Oct 0 UTC. For southern England, the e-suite has less intense wind gusts than o-suite. For souther Scandinavia, it seems like the band of the highest wind gusts are a little bit to far north for both suites. In the morning on the 28th, SMHI (Swedish met service) issue a red warning for the Gothenburg area but later they had to move the red warning to the Malmo area.
The figure above shows forecasts initialised 27 Oct 0 UTC (+36h).
The figure above shows forecasts initialised 26 Oct 0 UTC (+60h). Here a large area west of England had very high wind gusts, what did not happened. We also see that the cyclone centre is more to the west compared to later runs. Both these results is due to a development of the cyclone too far west.
The figure above shows forecasts initialised 25 Oct 0 UTC (+84h).
4.1.3 Mesoscale Structure
Mean wind speeds and wind gusts are not just dependant on low depth. The multiple possible scenarios provided by HRES and ENS, in O- and E-suites, have shown considerable sensitivity, in different parts of the storm - that make the difference between a normal windstorm, and a once-in-a-lifetime event. Pinning this down is the real challenge. The following plots illustrate some of the issues - they have been extracted from animations in Ervin's daily reports, linked to above.
Max 10m gusts (in preceding hour) - Obs on right, same colour scale | 10m mean speed - Obs on right, same colour scale (but different to gust scale) | |
03UTC O-suite at top E-suite below
| ||
---|---|---|
04UTC O-suite at top E-suite below | ||
05UTC O-suite at top E-suite below | ||
06UTC O-suite at top E-suite below | ||
07UTC O-suite at top E-suite below | ||
Gust scale: 20-green-24-26-28-dark blue-30-orange-32-red-34-magenta-37-40 (m/s) |
Several things to note:
In the warm sector (E and SE of low centre)
1. Huge variability in gusts due E of the low centre in both suites, yet only minor variability in mean speeds. Gust variability does not appear to relate to geostrophic or gradient winds, both of which look, to the eye, to be relatively unvarying with time (consistent with the steadier mean speeds).
2. Much lower gusts overall in the E-suite, despite what appear to be mean winds that are very similar to the O-suite.
3. Considerable jumpiness in gusts between successive times - eg in the E-suite we go from locally 30m/s over Kent 02-03UTC, to widely only 25m/s 04-05UTC. E-suite gusts over the sea also reduce markedly 02-03 to 03-04UTC.
4. Compared to OBS, both suites look to have overdone the gusts, but the E-suite seems much better (lighter).
Sting jet area (SW of low centre)
1. Starts to appear in O-suite, correctly, by 05UTC, but not in E-suite.
2. In the O-suite sting jet the gust increase nicely mirrors the mean speed increase (allowing for rapid translation eastwards during the 1h interval).
3. Compared to OBS the O-suite representation is good (though gusts slightly underdone perhaps). The E-suite gusts are much worse (due to no sting jet).
So we have identified very different behaviours according to synoptic location. In general terms the model behaviour also nicely reflects behaviour seen on the Reading observation graphs above. Note the isolated 25 m/s warm sector gust, unrelated to a change in the mean, as discussed above, and a sting jet gust max later that was related to the mean. One could imagine that the isolated warm sector gusts could have been hard to predict, and in this sense the volatility of the gust parametrisation in this regime in both suites may not have been that surprising. Conversely the sting jet gusts, which relate much more clearly to pressure gradients and mean winds, are on the one hand easier to predicte because of this, but on the other very hard to predict because those pressure gradient details themselves pose considerable difficulties for the models.
As regards the source of the large gust differences between E- and O-suites, these must relate to the gust parametrisation scheme, which contains 3 terms, a) the 10m mean wind, b) a turbulence factor, related to stability and c) a convective gust adjustment. The latter is only activated if the deep convection parametrisation is active (though this does not have to be surface triggered). It would seem that contributions from (c) or more likely (b) are making the E-suite gusts less. Vertical diffusion has changed in the E-suite and this may be relevant; this is now under investigation within Physical Aspects.
Four supplementary short range forecasts runs were performed to see if any conclusions can be drawn about the gust differences between the E-suite and O-suite: (i) a copy of the E-suite run from 00 UTC on the 28th of october, (ii) a copy of the O-suite run from 00 UTC on the 28th of october, (iii) a run with the E-suite branch initialized from the analysis at 00 UTC on the 28th of october produced by the O-suite, (iv) a run with the O-suite branch initialized from the analysis at 00 UTC on the 28th of october produced by the E-suite. The differences in wind gusts at a short range (steps 3 and 4) between these runs suggest that the differences between the E-suite and the O-suite discussed above are partly due to the differences in the initial conditions. It is therefore difficult to attribute the changes in wind gusts discussed above to the changes to one of the parameterizations or to the data assimilation system. Another aspect examined in these four runs was the contribution of each of the three terms entering the gust parameterization to the differences in wind gusts between the E-suite and the O-suite. It appears that the terms (a) and (b) have relatively similar contributions to these differences, which is expected given they both depend on the mean wind. The convective term contributes on average less than the other terms, but can locally contribute more (up to 5-6m/s in this case), for e.g. in regions where there was no convection in the O-suite and there is convection in the E-suite, or the other way around.
4.2 ENS
The sequence of figures above shows the strike probability of cyclone with maximum wind speed (mean) over 60 kt (31 m/s) at 1 km height. The plots are valid for the 28 October (24-hours). The first forecast is from 28 Oct 0 UTC and one day is added for each plot. With longer lead time the feature is delayed (too slow) in the model, but the path of the cyclone seems very consistent up to 6 days before the landfall.
VT 00Z 28th:
VT 12Z 28th:
The figures above show cyclones in the ensemble (dots), where the colour indicates the strength of the maximum wind speed at 1 km height connected to the feature. All the plots are valid 28 October 00 or 12 UTC. The contour shows the MSLP for the control forecast. The forecasts are from 0UTC runs starting from 21 Oct to 28 Oct. With shorter lead times, the spread of the features decreases. We also see that the centre of gravity for the red-orange dots moves eastward with decreasing lead time, illustrating the too westward development in the model.
The spread in the 00Z Sunday T+36h forecasts, reproduced below as a single frame (left), is especially striking, showing an ordinary but active autumnal frontal wave (green dots = 55-60kts) in a few members, through to a once-in-a-lifetime intense (red = 75-80kts) low in rather more members. Example snapshots from each of these two categories are shown to the right. This is a 'nightmare scenario' for forecasters, but at the same time probably gives a true reflection of uncertainty in what appears to have been a dynamically very sensitive situation. It would seem, perhaps, that uncertainty at other times was under-estimated?? The intensity range (as inferred from colour range) in forecasts from slightly earlier times, on the dalmatian animations above, seems, curiously, to have been less.
EFI for 10-metre maximum wind gusts can be seen here: from 38r2 (ps) and 40r1 (ps) suites. There is no significant difference between both managing to give a signal of anomalous wind even 7-days in advance but unluckily weakening the storm in Sunday's forecasts. That was the main concern expressed by the local forecasters during the member state visit to the Netherlands later that week. They presented at the meeting the following results from the HRES which show that a forecast from the Sunday 12UTC run just after they had issued a warning suddenly made the storm substantially weaker (right plot).
HRES MSLP, wind and wind gusts forecasts for different lead times started from longer ones (left to right).
Comparing the CDFs from three locations: one in East England, second in the Netherlands and third in Denmark, where hurricane-force wind gusts were observed, show that in the runs where the storm was weaker the 40r1 suite had bigger spread than 38r2.
EAST ENGLAND (OBSERVED WIND GUSTS > 35 M/S):
38R2 40r1
THE NETHERLANDS (OBSERVED WIND GUSTS > 35 M/S):
38R2 40r1
DENMARK (OBSERVED WIND GUSTS > 40M/S):
38R2 40r1
The figures above shows the probability for wind gusts over 33 m/s during the 28 October. The first figures is for the ensemble forecast issued 28 Oct 00 UTC and the following are forecasts with 12 hours in between. For the 2nd and 3rd ensembles, the signal for hurricane gusts are weaker than before and after, especially for the forecast from the 27 Oct 12 UTC. This was the problematic forecast for the Netherlands discussed above.
------------------------------------------------------------------------------------------------------------------------------------------------------------------
CDF for wind gusts (left) and maximum mean wind speed (right) for Reading valid on the 28 October. Different colour represents different initial times and the o-suite (38r2, solid) and e-suite (40r1, dotted). Comparing different initial times, the forecast from 26 October (green and black), were the most severe. For the forecasts from the 27 October the wind speed was less. This is partly because the development of the cyclone occurred later and therefore had not reached such a deep stage when it passed over Reading. Comparing e-suite and o-suite, we see that the winds (both mean and gusts) are less in the e-suite and o-suite.
With regard to ensemble spread, from a well-behaved ensemble, one would expect that, between consecutive runs, the cdf curves tend to become steeper, and gradually move less and less in the lateral direction as the event approaches. There are undoubtedly complicating factors for wind strength when one is close to the low track (which can for example give bimodal distributions = cdf jumps) but if one puts that aspect to one side it would seem that these EPS runs, in E and O-suites, are not behaving quite as they should. The inter-quartile range, which can be exactly inferred from these plots (and likewise the standard deviation which can be usefully estimated) appear not to be reducing with lead-time, and there are also notable lateral jumps in the profiles at shorter leads. A characteristic of a well-behaved ensemble is that measures of spread, for a given validity time, will reduce as the lead time reduces, i.e. for later and later forecasts.
4.1.1 O-Suite
Strip of forecasts below (left column) shows 48h maximum gusts from successive HRES forecasts from DT 00UTC 21st through to 00UTC 27th (left click, then right click -> view image to get full resolution), for VT period 06UTC 27th to 06UTC 29th. Note how the early forecasts jumped around, as one might expect. Later on very extreme gusts were forecast for areas SW of the UK, and as the event approached this zone of extreme gusts migrated its way eastwards, which is perhaps not what one would want/expect. End of period mslp is also shown.
By 'adding together' equivalent plots from two very short range forecasts we can provide a model-based pseudo-verification to compare the above with. These are below in rapid animation and two-up formats (covering the period 18Z 27th to 06Z 29th). No gusts of note were forecast for the 06Z-18Z 27th period.
Forecasts |
Model-based verification
| Model-based verification
|
---|
4.1.2 O-Suite vs E-suite
The figures below shows forecasts of MSLP valid for 28 Oct 12 UTC and maximum wind gust during the 28 October. The o-suite is plotted in the left column and the e-suite in the right.
The figure above shows forecasts initialised 28 Oct 0 UTC. For southern England, the e-suite has less intense wind gusts than o-suite. For souther Scandinavia, it seems like the band of the highest wind gusts are a little bit to far north for both suites. In the morning on the 28th, SMHI (Swedish met service) issue a red warning for the Gothenburg area but later they had to move the red warning to the Malmo area.
The figure above shows forecasts initialised 27 Oct 0 UTC (+36h).
The figure above shows forecasts initialised 26 Oct 0 UTC (+60h). Here a large area west of England had very high wind gusts, what did not happened. We also see that the cyclone centre is more to the west compared to later runs. Both these results is due to a development of the cyclone too far west.
The figure above shows forecasts initialised 25 Oct 0 UTC (+84h).
4.2 ENS
The sequence of figures above shows the strike probability of cyclone with maximum wind speed (mean) over 60 kt (31 m/s) at 1 km height. The plots are valid for the 28 October (24-hours). The first forecast is from 28 Oct 0 UTC and one day is added for each plot. With longer lead time the feature is delayed (too slow) in the model, but the path of the cyclone seems very consistent up to 6 days before the landfall.
The figures above show cyclones in the ensemble (dots), where the colour indicates the strength of the maximum wind speed at 1 km height connected to the feature. All the plots are valid 28 October 12 UTC. The contour shows the MSLP for the control forecast. The forecasts are from 0UTC runs starting from 28 Oct (top-left) to 23 Oct (bottom-right). With increasing lead times, the spread of the features over the North Sea increases. We also see that the centre of gravity for the red-orange dots moves westward with increasing lead time, illustrating the too westward development in the model.
ADD EFI PLOTS
CDF for wind gusts (left) and maximum mean wind speed (right) for Reading valid on the 28 October. Different colour represents different initial times and the o-suite (38r2, solid) and e-suite (40r1, dotted). Comparing different initial times, the forecast from 26 October (green and black), were the most severe. For the forecasts from the 27 October the wind speed was less. This is partly because the development of the cyclone occurred later and therefore had not reached such a deep stage when it passed over Reading. Comparing e-suite and o-suite, we see that the winds (both mean and gusts) are less in the e-suite and o-suite.
4.3 Monthly forecasts
We cannot expect to see a strong signal for extreme cyclones in the monthly forecast. However, we can investigate whether the environment was favourable for windy conditions. The figures above show the weekly MSLP anomaly for the week starting on the 28 October. The first figure shows the forecast from the 28 October (Monday of the verifying week), followed by 24 Oct, 21 Oct, 17 Oct, 14 Oct and 10 Oct. At least for the 5 first forecasts a positive NAO signal (negative MSLP anomaly in the northern Atlantic and positive further south) is present. A positive NAO is usually leading to stronger winds over western Europe.
4.4 Comparison with other centres
CDF for 10-metre mean wind for Reading (left) and a point (55N, 9E ) in western Denmark(right). The data is of the maximum for 4 time step valid 28 Oct 0, 6,12 ,18 UTC and 29 Oct 0 UTC. The different colours represents different centres in the TIGGE archive. Two different initial times are plotted, 24 Oct 00 UTC (dotted) and 26 Oct 00 UTC (solid). For Reading the maximum mean wind was 10 m/s and the point in Denmark lies within the area of 25-30 m/s. The results for ECMWF is not convincing, especially for the Danish point; the ECMWF forecast has the lowest wind speeds for both initial times. However, there could be a mixture of land points and sea point between the difference centres. The figure below shows therefore the mean wind on 850 hPa for the same point. Here, at least for the 26 October, ECMWF shows the highest wind speeds. For the 24 October it could be the case that the timing error led to that the cyclone had not reached this point within the window for the diagnostics.
...
The figures above show forecasts from 22 October 00 UTC (144 to 168 hour forecasts).
5. Experience from general performance/other cases
- The sting jet of the 4 January storm over Scotland 2012. See attached poster from Tim Hewson.
6. Good and bad aspects of the forecasts for the event
- The early signal of the storm in the forecasts (from ~6 days before the event).
...
- The storm developed too far west in the forecasts, and therefore over-forecast the intensity for western England.
- Also for the very short forecast, the path of the strongest winds where too far north over Denmark and Swedenand Sweden.
- Too weak storms in the forecast from the 27 October, especially 12 UTC, which the KNMI complained about.
- The difference in wind speeds between e-suite and o-suite. This has been further evaluated with the ENS reforecasts. Comparing the 99th percentile of the wind speeds and wind gusts for October (in total 1920 forecasts), we see some signs of lower extreme winds in the e-suite. However, the magnitude of the difference is small (about 1 m/s0 and it is hard to verify whether the change is good or bad.
7. Additional material
Tendency plot for a few analysis cycles can be found here: http://intra.ecmwf.int/plots/d/inspect/_dir_dir_38R2_STORMS/dir_38R2_STORMS/Analysis/analysis_tendencies!38R2_20131027-20131027!mean!u!StormLocation!850!/