MODEL OUTPUT

 
 

 
 
 
 

INTRODUCTION

 Model output products include all products that use model data. The model forecast variables can be looked at directly, postprocessed into grids, plots, or station predictions, or can be used in combination with climatology and other data sources in statistical forecasts. Statistical procedures are used to generate specific quantities of interest at particular locations, such as the high temperature for a city, from the raw data. Collectively, they are an important part of the forecast process. Some of these statistical procedures can reduce biases which are inherent to the model.  The statistical procedures, gridded output, maps, and biases are described in more detail below.
 
 

STATISTICAL WEATHER FORECASTING

 Some sensible weather elements, such as visibility and thunderstorms, are not predicted by the model and cannot be derived directly from the model forecast variables. Other parameters, such as surface maximum temperature, are sensitive to model weaknesses and vary locally. Statistical techniques have been developed to predict weather elements at particular point locations by using direct and postprocessed model fields, and climatology. The data found from this information is then used to construct equations relating the variable (predictand) to other variables (predictors).   There are two types of statistical weather forecast products that exist today:

    1. Perfect-Prog
    2. MOS (Model Output Statistics)

Both approaches are analogous to what an experienced forecaster does in practice. A forecaster defines relationships between observations and forecast data, and then applies this to the interpretation of operational numerical model output, improving upon the raw model forecast. But while a forecaster carries the enhancement out subjectively, statistical forecasting accomplishes it through objective means.

Perfect-Prog and MOS use large multiple regression equations and have a high range of predictor variables.   The NGM, AVN, MRF (Medium Range Forecast model), ECMWF, and Eta use MOS, while the RUC uses Perfect-Prog.
 

Perfect-Prog:
 

Perfect-prog was the first statistical approach taken to numerical weather prediction.  Short for "Perfect Prognosis", this method lives up to its name by making no corrections for possible biases of the model. The Perfect-Prog statisical technique develops equations based on the relationship of co-existing observed weather elements (including climate data), which are then applied to raw  model output. Perfect-Prog equations may or may not account for time lag in their development (unlike persistence); instead they may relate simultaneous predictors and predictands values from same time frame, or they may relate a predictand to a predictor which was observed several hours earlier. Time lag within the Perfect-Prog approach is accounted for by applying these derived relationships to forecasts from the numerical model. For example, with the use of 100-500 hPa thickness to differentiate between frozen and liquid precipitation, Perfect-Prog develops relationship equations from observations of the thickness (predictor) and observed precipitation type (predictand). These relationship equations are then applied to model forecasts of the predictor (thickness) to produce a forecast of precipitation type. If a model's predictions are "correct", then a perfect-prog forecast will be accurate.

 
 

For example, suppose we want to predict tomorrow morning's minimum temperature at College Station, and the two best predictors are the 1000-500 mb thickness and the 700 mb temperature at 12Z.  In the PP method, an equation is developed from historical data by regressing observed values of these two parameters against the minimum temperatures.  To make a prediction for 12Z, we would use the values of the two parameters which are predicted by the numerical model as input to the regression equation.

 
 

Advantages
Disadvantages
  • Does not require a developmental data set of historical model data
  • Because historical model data are not used in the development of statistical equations, systematic model errors cannot be accounted for
  • Likely to improve with improvements to the raw model forecast
  • Cannot use important derived model parameters as predictors (e.g. model vertical velocity)
  • Multiple predictors can be used, resulting in a better fit to the predictand data, and more accurate guidance
  • Cannot account for the deterioration of model forecasts with lead time.

MOS
 

The most preferred method of statistical weather forecasting is MOS.  It involves "regression equations that include influences of specific characteristics of different parts of a model" (Wilks 201).  The MOS technique develops relationship equations between observed and model forecast weather elements and applies these relationships to raw model output (of the same or similar model) to produce statistical guidance. MOS is used frequently in the National Weather Service. Since MOS uses NWP output in both the development and implementation of the statistical equations, both time lag and systematic error from model biases can be accounted for. To develop MOS forecast equations, it is necessary to have a data set consisting of several years of historical records, which use predictands together with the model output on the same days the predictand was observed.

For example, forecasters unconsciously apply the MOS technique subjectively when using model forecasts of 70% RH to estimate cloud forecasts. This practice may work in a number of cases but a problem arises if the model has a dry or wet bias and lower/higher model RH values are not used to estimate clouds. In the case of a dry bias, if the forecaster does not account for the lower model RH values by using model RH values of less than 70%, clouds are likely to be underforecast by the model and the forecaster. Even if the forecaster does make some adjustment by using 60 or 65% RH, the selection of these values is somewhat subjective and not necessarily based on an established relationship. However, if MOS guidance is used to estimate the clouds, the statistical relationships developed for that model will already take into account such systematic errors as the dry bias and will forecast the clouds more accurately.

 
 

Advantages
Disadvantages
  • Can account for systematic model errors and deteriorating model accuracy
  • Requires a developmental dataset of historical model data that is used as the predictor.
  • Accounts for predictability of model variables by selecting those that provide more useful forecast information 
  • Equations are model dependent
  • Mulitple predictors can be used
  • Need to modify MOS when changes are made to the NWP model
  • Better skill for longer range forecasts
  • Precipitation requires regional equations

The following table summarizes the differences between Perfect-Prog and MOS:
 
 

  Perfect-Prog
      MOS 
Strong predictor-predictand relationships because only current observed data is used  Relationships weaken with time due to increasing model error variance 
Does not account for model bias;  model errors decrease accuracy   Accounts for model bias 
Large development sample possible  Generally small development samples 
Access to observed or analyzed variables  Access to model output variables that may not be observed 
 
 

GRIDDED DATA OUTPUT

Grid point and spectral models are two types of models that formulate and solve equations differently. While they are based on the same set of primitive equations, grid point models represent data on a fixed set of grid points and spectral models represent data by wave functions. A summary of the different types of models can be found here.

The primary type of model output is gridded data.  The Eta, NGM, and other regional models calculate various
meteorological variables in grid boxes that have varying resolutions.  The vertical coordinate of the grid boxes is normalized
pressure, or sigma, and spherical coordinates are used in the horizontal.  This is true for all models except the
RUC.   The National Center for Environmental Prediction distributes analyses and forecasts based on gridded data in
GEMPAK format.  GEMPAK is one of the primary programs used to graphically view the output of this data.

The following table summarizes the basic model fundamentals of some of the most commonly used models.
Model Fundamentals
ETA
AVN/MRF
RUC
AFWA
NOGAPS
Model FAQ Links
Eta FAQ
AVN/MRF Information
 
Model Structure & Dynamics

 

 

 

 

 
Model Type
Grid Point
Spectral
 Grid Point
  Grid Point
Spectral
Vertical Coordinate System
Eta
Sigma
Hybrid Isentropic-Sigma
 Non-hydrostatic Sigma
Hybrid Sigma/Pressure
Horizontal Resolution
12 km
T170
40 km
  45 km, 15 km, and 5 km
T159, Physics 83-km
Vertical Resolution
60 Layers
42 Layers
40 Layers
  42 Levels
24-Layer
Domain
Regional
Global
Regional
  Mesoscale
Global
Derived Products
   

 

 

 
Postprocessing/ Products
   
MAPS/RUC diagnosed variables
  Model Predicted & Derived Products
1x1 Grids
Statistical Guidance
None
  
None
  None
None
Model Assessment Tools
Eta Assessment Tools
AVN Assessment Tools

MRF Assessment Tools
RUC2 Assessment Tools
AFWA MM5 Diagnostic Issues & Links;
Other MM5 Assessment Tools

NOGAPS Assessment Tools


 
 
 

MAPS AND IMAGES

In addition to GEMPAK produced images, the National Weather Service produces model maps and charts. The surface data, soundings, profilers, etc. are all used to produce a forecast model.  Geographic maps of various pressure surfaces can be produced from this data.  These maps display various parameters such as temperature, humidity, vorticity, convergence/divergence, and vertical motion.

There are many web sites that make these model maps and images available.  ETA, NGM, AVN, MRF, RUC, GSM, and ECMWF all have various output images available.  The following is a list of web sites where they can be found:

        READY Page - featuring ETA, NGM, AVN, MRF, RUC, and RAMS

        Ohio State University - featuring ETA, NGM, AVN, MRF, ECMWF, MM5, and UKMET, RUC

        NCAR - featuring ETA, RUC, and GSM

        Texas A&M's NCEP Model Data - featuring ETA, NGM, and AVN

        Center for Ocean-Land-Atmosphere Studies (COLA) and Institute of Global Environment
        and Society (IGS) -  featuring ETA, AVN, MRF, and RUC

        NOAA's Climate Diagnostics Center (CDC) - various model maps

        ECMWF Images - images direct from the European model source

        National Weather Service - model maps from the NWS

To accompany these maps and images, MOS output is available for NGM, AVN, and MRF (courtesy of NOAA's Techniques Development Lab).
 
 
 

MODEL BIASES

Within the different models and the equations they use, many biases have been discovered. These biases are based on forecaster observations throughout the time the model has been in use. For example, up until recently the Eta tends to under-predict the monsoon convection over the Southwest U.S.  However, recent changes have been made to the ETA, AVN, and MRF models to correct for this systematic bias.  The new biases for each of  these models will only be discovered with time.
 

The following pages list some of the common model biases

      http://www.hpc.ncep.noaa.gov/mdlbias/biastext.html

      http://www.crh.noaa.gov/lmk/soo/docu/models.htm

      Interactive model biases page so you can view the biases for the past 5 or 10 days
 
 
 

          Click here to return to the main modeling page

 References

 
 
 
 

Wilks, Daniel S.  Statistical Methods in the Atmospheric Sciences.   San Diego:  Academic Press, 1995

Unidata, Inc., February 1, 1998:  GEMPAK Model Grid Reference, //www.unidata.ucar.edu/packages/gempak/examples/models/models.html

National Weather Service, April 28, 1998:  NWS Fax Charts, //weather.noaa.gov/fax/nwsfax.html

Techniques Development Lab, April 28, 1998:  Techniques Development Lab //tgsv5.nws.noaa.gov/tdl/

ECMWF, February 1, 1998:  ECMWF  //www.ecmwf.int

Texas A&M University, April 28, 1998:  Forecast Products from Texas A&M , //www.met.tamu.edu/weather/models.html

COLA/IGES, April 28, 1998:  COLA/IGES Weather Forecasts //grads.iges.org/pix/wx.html

NCEP, March 31, 2000: Model Biases//www.hpc.ncep.noaa.gov/mdlbias/modelbias.html

Air Resources Laboratory, February 28, 2000//www.arl.noaa.gov/ready/cmet.html

National Weather service:  Performance and Characteristics and Biases of the Operational Numerical Forecast Models//         http://www.crh.noaa.gov/lmk/soo/docu/models.htm
 

UCAR, February 14, 2002: Model Grid Output Characteristics // meted.ucar.edu/nwp/pcu2/index.htm

 

Page designed by:  Daniel Huckaby and James Lacy on May 4, 1998.
HTML:  metr452/mantis/modeloutput.html

Page updated by:  Jody Walls and John Strack on February 3, 1999.

Page updated by: Brad Armstrong, Christine Meleo, and Bruce Sherbon  February 28, 2000.

Page udpdated by: Lauri Horan, Abby Matlock, Ted Ryan on February 14, 2002.