NCEP Modelsand

Ensembles

ByRichard H. Grumm

National Weather Service State College PA 16803

and Robert Hart

The Pennsylvania State University

Introduction

• NCEPS Global Forecast System• NCEP Medium Range Ensemble

Forecast System (MREF)

• Why we need ensembles– one model many potential initial states!

• Ensemble Strategies– Lagged Averaged Forecasting and

d(Prog)/dt– True ensembles

Global Forecast System• GFS suite has a medium range forecast model

(MRF) and a global data assimilation system (GDAS). We will generally use GFS to refer to both.

• History – The first global spectral model was developed

experimentally during the late 1970s (Sela 1980) and implemented as the global forecast model at the National Meteorological Center (NMC) on 18 March 1981

– Original Model:• Triangular truncation at 30 waves with 12 levels

(T30L12)• T24L12 from 48 to 84 hours• T24L6 from 84 to 144 hours (TPBs 282A and 282B)

Global Forecast System• Early 1980s computer power allowed for the use of

spectral models. This model replaced the older grid-point base 7-level primitive equation model (PE) which had a 191 km grid. The PE model was implemented in the late1960s.

• 1985 after big changes model was renamed the Medium Range Forecast (MRF) Model.

• new physics packages, • an increase in the number of waves resolved to rhomboidal

truncation at 40 waves (R40), and an• increase in the number of equally spaced layers from 12 to

18. Currently, the GFS is run four times per day (00 UTC, 06 UTC, 12 UTC, and 18 UTC) out to 384 hours. The initial forecast resolution was changed on.

• Many changes have taken place since 1985.

GFS-2003• 29 October 2002 GFS upgraded to T254

– equivalent to about 55-km grid resolution– with 64 levels out to 3.5 days (84 hours).– Later forecast times,

• GFS has a resolution of T170 (equivalent to about 80-km resolution) with 42 levels from 3.5 to 7.5 days (180 hours),

• T126 (110-km resolution) with 28 levels beyond to day 16 (384 hours). The reduction in GFS model resolution at 84 hours is consistent with studies showing that the GFS benefits from higher resolution only out to about day 3-4.

• All GFS runs get their initial conditions from the Spectral Statistical Interpolation (SSI) global data assimilation system (GDAS), which is updated continuously throughout the day.

GFS SummaryGFS

Model Type Spectral

Vertical Coordinate Sigma 64 layers

Resolution T254 (1.6o or 55 km)

Domain Global

CPS Arakawa-Schubert

Runs 4 times daily(00,06,12, 18 UTC)

Resolution with time Varies with less resolution at longer times

MREF• NCEPS Global Forecast System• NCEP Medium Range Ensemble

Forecast System (MREF)

• Why we need ensembles– one model many potential initial states!

• Ensemble Strategies– Lagged Averaged Forecasting and

d(Prog)/dt– True ensembles

• Based on the GFS system and uses the core model of the GFS to make all predictions.

• Uses positive and negative perturbations as input to the model for each run.

• 0000 UTC– GFS run

• Full T170 GFS resolution to day 7• T62 thereafter to day 16

– 1 MREF control Member• T126 medium resolution control out to 7 days• truncated top T62 at day 7 and out to day 16

– 10 Perturbed Runs

• 1200 UTC– GFS (AVN)

• T170 to 126 hours• T60 to day 16

– 10 Perturbed runs

NCEP MREF SYSTEMMedium Range Ensemble Forecast system

•5 Negatively & 5 positively perturbed Members– 0000 and 1200 UTC with 10 total

members– each run at T126 to 84 hrs, then

truncated to T62 horizontal resolution. – Lower resolution provides speed

NCEP MREF SYSTEMPerturbations

A word about Perturbations

• Breeding method– goal to find most rapidly growing modes. So random

will generally not work.– Initial seed is random, get differences between control

and perturbed over several 12-hour forecasts cycles,by definition, the difference reflects only the growing modes. Each seed gives both 1 positive and 1 negative perturbation. Different initial seeds provide a set of 2 new growing modes.

– In SREF, 09Z 12-hour forecast referenced to control makes the 21Z forecast initial perturbed conditions. The 21Z 12-hour forecasts makes the next 09Z initial conditions etc.

Breeding

• Start with initial random seed. Very first forecast is truly random

• At 12-hours get difference and correct forecast from control. This is the next cycles perturbation– finish forecast to 63 hours

• Next cycle use perturbation and get next runs value 12-hours into the forecast

• Self breeding methodology

SREF BreedingN SEEDS GIVE 2*N PERTURBATIONS

Scaled + perturbation

Initial random seed

Opposite sign is negative perturbation

Adjust magnitude to typical analysis errors

12-h forecast12-h forecast

CONTROL-CTL

Complete cycle forecast

The 6 rule

• Deals with the number of grid points needed to resolve a wave

• As general rule – to fully resolve a wave you must have a grid

spacing of 1/6 the length/width of the feature– hence the 6 rule

• CSI band about ~30km wide 30/6=~5 km• The GFS and MREF have too coarse

resolution to resolve many important meteorological phenomena.

Why Ensembles• NCEPS Global Forecast System• NCEP Medium Range Ensemble

Forecast System (MREF)

• Why we need ensembles– one model many potential initial states!

• Ensemble Strategies– Lagged Averaged Forecasting and

d(Prog)/dt– True ensembles

*Ensembles help us

• Deal with uncertainties in data– the ability to properly resolve the feature

• Deal with uncertainties in data verse resolution of the model– 6 rule, we may under sample a system.

• Deal with uncertainties in physics

• Displaying uncertainties in forecasts– spaghetti plots– probability charts (the most likely outcome)– consensus forecast charts – to visualize these is to see limits of any single solution.

Ensemble Strategies• NCEPS Global Forecast System• NCEP Medium Range Ensemble

Forecast System (MREF)

• Why we need ensembles– one model many potential initial states!

• Ensemble Strategies– Lagged Averaged Forecasting and

d(Prog)/dt– True ensembles

Ensemble Strategies• Lagged Average Forecasts (LAF)

– dProg/dt is often used instead of true LAF– simple uses 1 model and different runs of that model– See trends in 0000,0600, 1200 and 1800 UTC model runs– See areas of disagreement/dispersion– last member considered most skillful

• Single model with perturbed members– all members should be of similar skill– all initialized at same time (better than LAF)– Our SREF and MREF data

• Ensemble of many models (The best!)– vary conditions and physics– “super ensemble”

Ensemble from LAFmay convey uncertainty in initial conditions

Forecast Length

Envelope of solutions

at single time

Solution

Forecasts Initialized at different Times

AVN dProg/dt at 18Z 30 Dec 2000

Derived LAF

Ensembles with different initial conditions

Forecast LengthForecasts Initialized at most recent data time

Envelope of solutions

at single time

Solution

Putting it all together

• Why we need to consider ensembles– Avoid the binary yes/no single model solution– The more probable ensemble solution

• Do you a hit the bulls-eye yes/no?

• Do you want to approximate the bulls-eyes most likely position?

Hit the bull's-eyeone arrow could be a near miss (the MRF)

Approximate the Bulls-eyeError analysis approximates the bull's-eye verse our one “red

arrow”MREF Forecasts

Real example 30 Dec 2000

bulls eye is GREEN…MRF is BLUE

A brief word about displays

• A good ensemble system requires 30-50 members– NCEPs system will grow in members!

– Displaying individual members is rapidly becoming problematic

– We need to move away from • thumb nails (next slide)• individual model diagnostics

• We need to move toward – Spaghetti and dispersion to see variations between

members – consensus and clustered means– probabilistic displays

Thumb Nail

Surface Low

00 UTC 7 Jan 2002

12 QPF Thumbnails

Valid06 UTC 6-12Z

7 Jan

imagine 45 of these!

MREF ForecastsArctic Outbreak-From this…

To…Arctic Air for 23 Jan 2003not too shabby for 10 days!

Cold air NW Flow Lake Effect 10 days out/near miss cyclone off coast?

Lake Effect 10 days out

23 Jan 2002 1200 UTC

Conclusions

• NCEPS Global Forecast System– MRF model– Global data assimilation system– About 55 km resolution

• NCEP Medium Range Ensemble Forecast System (MREF)– Run 2X daily with 11-12 members– GFS is the model behind the MREF– MREF uses perturbed initial states to get

the 11-12 different forecasts.

Ensemble Summary

• Ensembles are– collections of forecasts from

• different forecasters• different models and models with different conditions

– ensemble mean (consensus) is a high probability outcome

• Use ensembles to deal with uncertainty– in initial conditions– in model physic and parameterizations

• Operationally– view spaghetti plots: the range of solutions– ensemble mean and probability forecasts– Avoid the binary YES/NO (right/wrong forecast)

Ensembles Capabilities• Great promise on some big snowstorms

– Did well 3-4 Dec and 30 Dec 2000– Did will 5 December and 25 December 2002 – Did well 3-4 January 2003 – Did poorly 6-7 Jan 2002

• MREFsystem is not perfect (but getting better!)– more members being added– improved quality of models used in system– Eventually 4 runs per day as 06 and 18 UTC will be added

• NCEP MREF/SREF Data is a valuable forecast tool.

Consider this• Consider this: Would you want to shoot

one arrow at the bullseye or a quiver full of arrows at the bullseye. Or…pick the Eta and maybe miss the mark or use the ensembles and have a better chance of hitting something.