Access to these pages is available only to ECMWF and authorised external users
POINT RAINFALL
The point-rainfall is a test product, developed at ECMWF. It is not part of ECMWF's operational suite. It is released on daily basis at 13 UTC but please note that it could be delayed in certain cases. In those cases, a note will appear in the section "NEWS" to let the users know about any delay or problem with the release of this test product. Apologises for the inconveniences that this may cause.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- NEWS ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
April, 1st: Failure disk. Point-rainfall forecast delayed.
April, 2nd: Failure of April, 1st solved. Point-rainfall system restored to produce forecasts as usual.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The point-rainfall consists in a statistical post-processing method, in order to define the probabilities for extreme point-rainfall, consists in the definition of "100 new and equally probable post-processed realisations" for each of ECMWF's 51 ensemble members (Control Run + 50 Perturbed Forecasts). This means that 5100 post-processed values are handled during the processing.
The point-rainfall probabilities are computed in two different ways:
- One hundred percentiles (1 to 100) are computed considering the whole set of 5100 values. Therefore, the dataset is composed of 100 values per grid box; These should define a "true" point probability, that can be verified as such.
- For each post-processed ensemble member (group of "100 realisations") the 95th and the 98th percentile are computed. So each new n'th percentile dataset comprises 51 (equi-probable) values per grid box. One can then compute the mean across the ensemble for a given percentile, as on plots B, D, F below, which gives another useful measure of potential for extremes (or alternatively plot distributions of these 51 values).
However, in order to post-process the total precipitation amounts derived from the IFS, we would need to know how its convective and large scale component behave separately. Since it is not possible to disentangle that information from the observations, it was considered appropriate to compute the post-processed rainfall in different ways and let verification decide which dataset gives better results.
In particular, three different datasets are computed:
a. the Point-Rainfall is defined on the basis of one decision tree for the CPr (convective precipitation ratio = convective rainfall/total precipitation);
b. the Point-Rainfall is defined on the basis of two decision trees (one for the convective component of the total precipitation and one for the large-scale component). Then, the convective and the large-scale component are added together to determine the total precipitation; in this case, one assumes that the convective and the large-scale component are dependent. Therefore, small values of the convective component are added to the small values of the large-scale component, and so on;
c. as in point b, but one assumes that the convective and the large-scale component are independent. Therefore, one randomizes the position of the values of the large-scale component and afterwards one adds them together.
3 month-verification (June, September and December 2016) has shown that the dataset that produces better results is the one computed on the basis of one decision tree (letter a). Therefore, now it is possible to compute the point-rainfall for longer lead-times. indeed, forecasts up to Day 5 are computed. Notice that a confluence page for the point-rainfall verification will be set up soon. There all the results for the verification will be shown.
The project aims to compute the 6 and the 12 hourly accumulations for the post-processed rainfall. At the moment just the 12 hourly accumulations are available. In early 2017 the 6 hourly accumulations will be computed as well.
Due to the fact that Europe has the biggest dataset of observations for future verification, in order to better take into account the diurnal and nocturnal cycle, the 12 hourly accumulations between 6-18 UTC and 18-6 UTC are considered (since those times define day and night time in Europe).
After this testing period, we aim to proceed with the computations until day 5 for the following intervals:
- for the 12 hourly accumulations, 00-12 UTC, 06-18 UTC, 12-00 UTC, 18-06 UTC;
- for the 6 hourly accumulations, 00-06 UTC, 06-12 UTC, 12-18 UTC, 18-24 UTC;
12 hourly accumulations - 95th percentile
Midnight Run (00 UTC)
Lead time: up to day 5
Notes for the Legend:
- The values of the Point Rainfall are in mm
- VT corresponds to "Verifying Time"
Lead - Time | ECMWF Ensemble | Point - Rainfall (TEST PRODUCT) |
---|---|---|
00UTC (t+06,t+18) | ||
00UTC (t+18,t+30) | ||
00UTC (t+30,t+42) | ||
00UTC (t+42,t+54) | ||
00UTC (t+54,t+66) | ||
00UTC (t+66,t+78) | ||
00UTC (t+78,t+90) | ||
00UTC (t+90,t+102) | ||
00UTC (t+102,t+114) | ||
00UTC (t+114,t+126) |