What Is EKDMOS?

Ensemble Kernel Density Model Output Statistics (EKDMOS) is a statistical postprocessing technique that generates probabilistic forecasts for sensible weather elements.  EKDMOS forecasts are generated from the North American Ensemble Forecast System (NAEFS), which is a combination of NCEP’s Global Ensemble Forecast System (GEFS) and the Canadian Meteorological Centre’s Ensemble (CMCE).  EKDMOS forecasts are currently available or being planned for the following elements and domains:

EKDMOS became an operational product on NCEP's Central Computing System on Tuesday, May 29, 2012. It included gridded guidance forecasts for the following:

EKDMOS Domains and Elements
Element CONUS Alaska Hawaii Puerto Rico Guam
Temperature Operational Operational Operational Operational v3.0
Dewpoint Operational Operational Operational Operational v3.0
Maximum Temperature Operational Operational Operational Operational v3.0
Minimum Temperature Operational Operational Operational Operational v3.0
Apparent Temperature v2.1 v2.1 v2.1 v2.1 v3.0
Wind Speed v2.1 v2.1 v2.1 v2.1 v3.0
PQPF06 v2.1 v2.1 v2.1 v2.1 v3.0
PQPF12 v2.1 v2.1 v2.1 v2.1 v3.0

v1.1 - Operational version from 12/02/2015 through 11/15/2016
v2.0 - Operational version as of 11/15/2016
v2.1 - Currently in development.  Forecasts will be made available through the EKDMOS webpage.
v3.0 - Currently in the planning phase.

EKDMOS Technique

Need to modify publication references.

Development

 

EKDMOS starts by applying forward-screening multiple linear regression to the ensemble mean of the GEFS and the ensemble mean of the CMCE separately, relating them to a set of observations maintained by MDL.  This yields a MOS equation and an error estimate for each ensemble forecast system (EFS) for each station, element, forecast projection and cycle run.  These equations are then applied to the 21 members of each EFS respectively in the development sample to produce 42 ensemble MOS forecasts.  This technique is currently being applied to roughly 3000 stations throughout the CONUS, Alaska, Hawaii, Puerto Rico, and Canada. See "MOS Uncertainty Estimates in an Ensemble Framework" by Glahn, et al 2009 for more information.

 

A second regression step is then performed to obtain the spread-skill equations.  First,the absolute error of the mean of the 42 ensemble MOS forecasts and the standard deviation of the 42 ensemble MOS forecasts are both transformed by taking the square root of each.  This transformation step has been shown to produce a more robust spread-skill relationship, which is obtained from a one-term regression step for each station, element, forecast projection, and cycle run.  See "Spread Calibration of Ensemble MOS Forecasts" by Veenhuis 2013 for more information.

 

Production

 

In production, the MOS equations are applied to the 21 raw ensemble members of each EFS to produce 42 independent ensemble MOS forecasts, each with its own error variance.  Kernel density fitting is used to combine the ensemble members into a probability density function (PDF).  The square root of the standard deviation of the ensemble members is plugged into the spread-skill equation.  A back-transformation gives us the expected error of the ensemble mean.  The spread of the PDF is adjusted to match the expected error.  The PDF is then converted to a cumulative distribution function (CDF).  The EKDMOS forecasts consists of 11 points on the CDF (each decile along with .05 and .95) and the ensemble mean.

 

The ensemble mean and each probability level are gridded for each domain using the BCDG technique.  See “The Gridding of MOS” by Glahn, et al 2009 for more information on the BCDG technique.  Checks are performed to ensure consistency between elements for the ensemble means.  Checks are also performed to ensure that the probability levels are monotonically increasing.  All grids are then converted to GRIB2 before being disseminated.

MOS