Ensemble Prediction System

• EPS Dissemination products in field form
• EPS Dissemination requirements
• GRIB defintions for EPS products
• Control, Perturbed and Calibration/Validation forecast products
• Cluster Mean and Cluster Representative products
• Old Cluster Mean and Cluster Standard Deviation products
• Ensemble Mean and Ensemble Standard Deviation products
• Event Probability products
• TUBE products
• Extreme forecast index, Extreme forecast index control data and Shift of tails index
• EPS Weather Parameter products
• TROPICAL CYCLONE trajectory forecast products
• Transmission of EPS products

The Ensemble Prediction System (EPS) has been run at ECMWF since December 1992. Initially it was run on an experimental basis as a 32 perturbed member ensemble. Forecasts were run at resolution T63, at 19 levels.

In December 1996, the operational Ensemble Prediction System was changed in two ways:

i) The model on which the EPS is run, was upgraded to TL159 truncation with 31 levels.

ii) The number of perturbed ensemble members was increased to 50.

In October 1999, the model on which the EPS is run, was upgraded to 40 vertical levels. A control run was also made at TL159L40.

In November 2000, the model on which the EPS is run, was upgraded to TL255L40. A control run was also made at TL255L40.

In September 2005, the model on which the EPS is run, was upgraded to TL399L62. The same upgrade was made to the control run.

In November 2006 ECMWF upgraded its Ensemble Prediction System (EPS). In particular, the forecast range was extended to 15 days using the EPS system with a resolution of T399 L62 for day 1 to day 10 (leg 1) and T255 L62 for T+246 to day 15 (leg 2). EPS also included two other constant-resolution forecasts for calibration and validation purposes ran one at T399 L62 and the other at T255 L62 for the full forecast time range day 1 to day 15.

In January 2010, the model on which the EPS is run, was upgraded to T639 for day 1 to day 10 (leg 1) and T319 for T+246 to day 15 (leg 2). The same upgrade was made to the control run.

The core of the EPS is the generation of perturbations to the initial analysis used to start the forecast. Perturbations are computed to generate 50 modified analysis, so the total number of forecasts is 51 (including a control forecast based on the unperturbed analysis). The perturbations are not generated at random (as would be the case for example with a pure Monte Carlo technique), instead, they are a combination of 25 modes which have the largest impact on the forecast for the Northern Hemisphere in the short range (detailed information on the technique can be found in ECMWF Technical Memorandum No 188, August 1992).

On June 2012, EDA perturbations redefined using the EDA ensemble mean instead of the EDA control as the reference: Up to cycle 37r3, the EDA initial perturbations are defined as differences between perturbed EDA members and the unperturbed EDA control member. From 38r1, the EDA initial perturbations are defined as differences between perturbed EDA members and the mean of the EDA. If the distribution sampled by the EDA is symmetric, the latter method retains the shape of this distribution. In addition, the variance of the singular vector initial perturbations is increased in 38r1 in order to compensate for the reduction of spread that is due to the change of the EDA perturbations.

For full definitions of ECMWF GRIB local extensions for EPS products see http://www.ecmwf.int/products/data/software/grib_api.html