...
The raw ENS forecasts provided good guidance here from 4 days in advance, albeit with a large underestimation versus the point values. That said there is some jumpiness in the locations forecast to be at greatest risk. And also in the time sequence of probabilitiesthere is also some temporal jumpiness probability magnitudes, with for example a big jump to lower probabilities apparent between the last two forecasts issued.
...
Graph shows the maximum of the 99th percentile ecPoint field in forecasts leading up to the event (over Spain, but always, in practice, close to Turis). The map shows a 200-year return period estimate for 24h gauge-based rainfall totals from ERA5-ecPoint (data from 2000-2019).
It has been noted that the standard 99th percentile value from ecPoint forecasts might be construed as too low a probability to justify assistance with warning issue (even if forecasts are consistent). This motivated a re-investigation of the ecPoint output, which can be presented in different ways. An alternative presentation option is shown below:
Comparison of two ecPoint-related products (see text) from the same data time (12UTC 25th October 2024), valid for 12h rainfall totals on day 4 (06-18UTC 29th October 2024)
For this we take the 99th percentile point rainfall value from each of the 51 ensemble members, just after the first stage of post-processing where members will have been dealt with individually. Via the assumptions underlying ecPoint post-processing these values are taken to represent the likely maximum value, within a given gridbox (here 18km scale), were that member's synoptic pattern to verify, and if one had 100 evenly spaced measurements in that gridbox. These 51 member 99th percentile values are then ranked, and from this ranking we can select and plot the median, for example, and then state that "there is a 50-50 chance of this value being exceeded somewhere within the gridbox". Such an approach has parallels with other neighbourhood-type post-processing approaches applied to limited area ensemble output, used for example at the Met Office in the UK to give the maximum likely rainfall value within a region. The day 4 forecast example plot above shows a general reduction in values with the new representation, as expected, although notably the extremes are largely "preserved" in the Valencia region. This will be due to relatively high inter-member consistency in that region.
We can then do the same as this for all relevant lead time and then adapt the ecPoint graph plot shown above to include also profiles for this new output, and indeed another related version where, instead of taking the median, we take the 10th percentile, to give a higher confidence forecast of exceedance somewhere in a gridbox. The revised ecPoint output (pink line) has comparable values to the pre-existing method (blue), albeit with a slightly steeper rise with lead time overall. Meanwhile the smaller quantile (10%) of the ranked 99th percentiles suggests high confidence of localised, dangerously high (if not catastrophic) values, say 200mm/12h, from 25th onwards (4 days in advance).
Different ways of representing ecPoint forecasts' post-processed output, for localised extreme rainfall, versus the 12h rainfall observation from Turis
3.4 Monthly forecasts
The plots below show forecasts of weekly precipitation anomalies for 28 October - 3 November.
...