About the ECMWF db

The database use to sample the large scale fields is based on the ECMWF ensemble reforecasts. Twice a week, the ECMWF reruns their current operational model for a number of past days. In particular, on January 16th 2017, the model was rerun for all January 16th's from 1997 to 2016. Instead of a full ensemble, only 5 members are run.

Download the database here.

The database has the following variables:

Variable	units
2m temperature	K
Precipitation	mm
Sea level pressure	Pa
Cloud area fraction	0-1
Zonal wind speed	m/s
Meridional wind speed	m/s
Incoming solar radiation	W/m2

These are all daily averages, except precipitation, which is a daily total.

Drift correction

For some forecast parameters, the model has a climatology that varies with leadtime. That is, the first day of the forecast could be dryer or wetter on average than the other forecast days. In addition, the first day may have greater or less spread of possible states than the other days.

This can be problematic for two reasons:

1. The "truth" dataset (i.e. the first day of the forecasts) will have a different climatology than the simulation
2. The end states of a 10-day trajectory can end up in a state that is rare for the beginning state, thereby incorrectly sampling the distribution.

To fix this, a quantile mapping correction has been applied to each leadtime. That is, leadtimes 1-9 have been adjusted such that they have the same frequency of values as leadtime 0. This proceedure is done separately on every gridpoint. Currently, we have not done this separately for each season.

Estimation of solar radiation

The raw data in the database does not include incoming solar radiation. We diagnose this value [more from Cristian here].