1
Systematic and Random Errors in Operational Forecasts by the UK
Met Office Global Model
Tim Hewson
Met Office
Exeter, England
Currently at SUNY, Albany (until Feb 2005)
Utility of different model forecasts
A multi-model (poor man’s) ensemble can provide the best forecast guidance
Operationally, can use be made of different models ?
Requires appropriate tools, and a detailed knowledge of typical model performance:
– Relative Errors, Seasonal and Regional differences [A]
– Individual Model Characteristics (systematic and random errors) [B]
A and B will be discussed here, focusing on the UK Met Office global model (~60km resolution, 38 levels)
A
Global Model Intercomparison:
Net, Seasonal and Regional differences
Northern Hemisphere RMS Mslp errors vs Lead Time
1 5 days 10
RANK Best - EC UK FR US JAP GER CAN… -Worst
Seasonal differences (NH mslp, RMS at T+72)
EC UK FR US JAP GER CAN
EC Best throughout; then UKMET, but NCEP consistently better in summer
Regional Performance – Europe, vs Lead Time Europe-based models perform better in forecasting for Europe
EC UK FR US JAP GER CAN
EC UK FR US JAP GER CAN
Regional Performance – N America, vs Lead Time Relative to performance over Europe: UKMET does worse over US/Canada, GFS better
B
UK Global Model Characteristics -
Systematic and Random Errors
Precipitation (net / orographic)
Low level Winds (Land / Ocean / Severe cyclonic storms)
Handling of Cyclones (Cyclone spectra / Regional / Random errors)
Global Precipitation
Precipitation ~ 30% overestimate globally
EnhancedResolution60km30L
3DVar & ATOVS
New DynamicsHadAM4 physics
c/o Sean MiltonMet Office, Exeter
Precipitation errors mainly oceanic – tropics and extra-tropical storm tracks
Largely ‘balanced’ by too much evaporation – boundary layer locally too dry
Soil moisture is one global weakness being addressed – led to under-prediction of daytime temperatures during 2003 European heatwave (UK bias -4C)
Orographic Precipitation
ODNDMTNS
A B C D E F G
New Model Old Model Orographic precipitation
Smoothed orography (in new model = “New Dynamics”) reduces upslope rainfall, and similarly reduces the rain shadow
Older model better (even if for the ‘wrong’ reason!)
Magnitude of impact is proportional to flow strength
Important for QPF
ODNDMTNS
A B C D E F G
B
A
C
DE F
B
A
C
DE F
GG
NE Region
Model orography peaks much lower than reality
Many key features missing – eg Hudson Valley
Expect similar ppn problems to those found in Europe – eg insufficient upslope rain in flow from SE quadrant (factor of 2?)
‘European’ higher resolution (20km) model may help
Convective Precipitation
Diurnal cycle in convection
A significant problem area (especially tropics, but also mid latitudes)
Decay can be too rapid towards dusk
Surface Winds over land
Example – Oct 2004
15kt winds in GFS model
(mslp v similar)
UK Global Model Effective Roughness Lengths
Account for roughness due to missing orography + …
Slows down low level winds considerably
10m winds especially poor in Albany: ~50% of reality
GFS model seems much better
Changes to be implemented in ~1 year
~50% reductionIn 10m winds
Surface Winds over Oceans
GFS model
Peak winds 55kts on S flank of deep, mature low
UKMET model
Peak 10m winds only 45kts
Gradients and low depth the same as GFS
Complex interface with ocean
GFS seemed to validate better in this case (and may well be better generally)
Surface winds in Extreme Storms
L
High resolution required (90 levels?) to model sting jet
Mslp may be OK but winds not
38 Levels(operational)
90 Levels
Greater strengthalong downwardtrajectory
Severe windstorms
c/o Pete ClarkJCMM, Reading
Cyclone Spectra
Cyclone Database - Snapshot
(a) standard frontal wave
(c) standard potential wave
(b) ‘barotropic’ low
(d) weak frontal wave
(e) weak potential wave
GM cyclone spectra for year 2000, categorised by ‘max wind speed within 300km radius of centre’
North Atlantic Domain
Geographical biases in cyclone forecasts, based on trends in total numbers T+0 to T+144
Under-predictionOver-prediction
Random Errors in Cyclogenesis
November 2003 Example
15Z
18Z
18Z
Intense cyclonic storm missed at short range – random error
Perhaps 3 similar poorly forecast events per year around UK
Expect similar problems elsewhere. High Impact.
Summary Met Office global model’s broadscale evolution is on average second only
to ECMWF (NH)– Performance over Europe better than over N America
– Performance in the 3 summer months lags behind GFS
Despite this a number of significant problem areas exist– Precipitation over-forecast globally by 30%
– Some significant errors around orography and in convection
– Low Level winds under-forecast over land with unresolved orography
– Some under-prediction of stronger winds over oceans?
– Wind maxima under-forecast in extreme storms (resolution limitation)
– No systematic drift with lead time in the number of intense storms
– Fewer modest cyclones predicted at longer lead times (main bias regions include Great lakes, Gulf stream wall)
– Significant random errors still occur occasionally, even at short leads
Many of the above noted through active forecaster-NWP liaison Most are now being addressed within NWP division at Met Office HQ