...
The methodology consists in identifying weaknesses in the model and test some physically-relevant parameters to infer sub-grid variability (and biases) in rainfall totals, and thereby forecast the probability of extreme point-rainfall.
What does sub-grid variability mean?
Table 1. Example of different scenarios that lead to different types of sub-grid variability in precipitation totals. Images show example radar-derived totals for cases that correspond to each of the example scenarios | The second column of this table shows the different types of sub-grid variability in precipitation totals (Images show example radar-derived totals for cases that correspond to each of the example scenarios). Based on clear-cut observational evidence from radar-derived totals, and from physical reasoning), one notes from the outset that within global ensemble member grid boxes (18km X 18km) very different geometries of sub-grid variability in precipitation totals can be observed. In principle we could have more (or indeed less) than three, though these serve as a useful illustration. Let us assume that the number of meaningful different scenarios is given by n, and that each one is indexed with label i. Scenario (i=) 1 could be said to be "zeroth order", i.e. point totals exhibit little sub-grid variability; scenario 2 is "first order", showing strong variability in one dimension; and scenario 3 is "second order" showing strong variability in two dimensions. Accordingly, within the grid box the distribution of point totals (i.e. the pdf, or probability density function) is as follows: for scenario 1 it is roughly Gaussian, with a sharply defined peak, and so "confident"; For scenario 3 it is roughly exponential, with a high probability of small or zero totals, tapering down to a small probability of very high totals, and so not confident at all; and for scenario 2 it lies somewhere in between.
|
Clearly to anticipate point totals, one must recognise and understand these types of sub-grid variability.
We define which parameters increase the sub-grid variability and biases in the total amounts of precipitations.
We post-process the ECMWF ensemble by using some parameters like type of precipitation (mainly convective or large-scale), wind speed, cape and solar radiation in order to asses the rainfall sub-grid variability and biases (for instance between the diurnal/nocturnal cycle). This means that you will see differences between the point-rainfall and the raw ensemble when the model grid values (i.e. the average of the infinite point values within the grid box) are not representative of the point-values.
Just an example.
Those depend on what the prevailing meteorological situation (i.e. weather type) is. Global numerical models include standard parameters in their output that allow what that situation is to be easily inferred, therefore, in a real-time forecasting mode one could then naturally anticipate the type of sub-grid variability that would be likely to occur. In this way one could automatically create a single forecast, covering all points in the gridbox that showed probabilities of different totals arising.
Let us now consider the physics of rainfall generation, and specifically what types of meteorological situation/model parameters would associate with sub-grid patterns similar to scenarios 1, 2, 3 in Table 1.
When mainly large scale precipitation is forecastWhen you have mainly large scale precipitation and high wind speeds, there is no much sub-grid variability in the precipitation and therefore, the model grid value is representative of the point-values and the raw ensemble and the point-rainfall plots are pretty similar. On the other hand, if there is mainly convective precipitation and light winds are forecast, this drives to a huge sub-grid variability in the total amounts of rainfall. Indeed, a plot of point-total amounts (let say a radar image) would show scenario 3 shows a cellular pattern in the totals showing several areas/points with zeros and just few with huge totals. Therefore, if the raw ensemble predict 10 mm/12h you know , it can be said that this is the an average of between lot of zeros and only few very high values of rainfall and won't be representative of the point-rainfall.
The above discussion use just simple scenarios but the model allows for any number of different parameters to be used and any number of scenarios to be defined. As well, we are not limited to atmospheric parameters, since fixed model parameters, like the sub-grid orography, could be also employed, as well as computed parameters from the model parameters like the "cell drift parameter" or the orographic modulation (not yet included in the current model, they are still in the research phase).
The current model take into account
- type of precipitation (mainly convective or large-scale);
- total precipitation;
- speed of steering winds (700 mbar);
- cape;
- solar radiation.
. Therefore, this are cases where you will see high differences between the raw ensemble and the point-rainfall plots because what we show in our maps is that even if is a small probability, there is a small chance to have very huge point total amount of rainfall which for the raw ensemble won't occur. Although, it is worth to notice the following main thing. We give the probabilities for extreme rainfall but we can't say where, within the grid-box, this extreme point amount will happen.
Point Rainfall Verification
...