1 WMO SWFDP Macau 8 April 2013 Anders Persson Introduction to ensemble forecasting
Jan 17, 2016
1 WMO SWFDP Macau 8 April
2013 Anders Persson
Introduction to ensemble forecasting
2 WMO SWFDP Macau 8 April
2013 Anders Persson23-04-21
Computer models
Forecaster
Customer/Public
Atmosphere
Scientists
The meteorological science in the service of MankindThe meteorological science in the service of Mankind
…is investigated and explored byScientists
…who summarize their finding into mathematicalComputer models
…which are used as an important tools by Forecasters
…whose final work is used as a basis for decision making by Customers/Public
But are they still needed?
In 1966 I was told that “in 5-10 years time there will be no need for human weather forecasters”
3 WMO SWFDP Macau 8 April
2013 Anders Persson
The arrival of the computer meant
increasing forecast skill and efficiency
but also new educational needs.
Irony: In agriculture nobody said:“ -With the
introduction of the tractor in 5-10 years there will be
no need of farmers”
The progress of weather forecasting
The human weather forecaster before the scientific age: simple rules and no complicated machinery
The arrival of the scientific method meant increasing forecast skill and efficiency but also an increased burden with thousands of observations, complex rules and more stressful work
4 WMO SWFDP Macau 8 April
2013 Anders Persson21/04/2321/04/23 4
On the contrary:There are perhaps more weather forecasters today than ever, even in – or particularly in – the commercial sector
Training Course at Meteo Group, Wageningen, NL
But what are they doing?
5 WMO SWFDP Macau 8 April
2013 Anders Persson
How can human weather forecasters compete with the super computers?
• Humans should not try to compete with them
• Instead they should play an entirely other “game”!
• The key word is not “skilful”, but “useful”
– How to best serve the people!
6 WMO SWFDP Macau 8 April
2013 Anders Persson23-04-21
Computer models
Forecaster
Customer/Public
Atmosphere
Scientists
The forecaster misled me!
The NWP misled me!
Erroneous observations misled the NWP!
The atmosphere is ”chaotic”!
Some don’t and engage in the Blame GameSome don’t and engage in the Blame Game
7 WMO SWFDP Macau 8 April
2013 Anders Persson23-04-21
Computer models
Forecaster
Customer/Public
Atmosphere
Scientists
Now I make better decisions!
I will take the uncertainty into
account!
Erroneous observations may mislead the NWP!
The atmosphere is chaotic!
Most meteorologists surely do this!Most meteorologists surely do this!
8 WMO SWFDP Macau 8 April
2013 Anders Persson
The main reason why we need ensemble forecasting: We want to estimate the uncertainties, in
particular the risks of extreme or high-impact weather
-But I do not want any risks, or probabilities or uncertainties – I
want to KNOW!
OK, let’s take your words seriously
9 WMO SWFDP Macau 8 April
2013 Anders Persson
Come with me to nice friendly Scandinavia
You venture out in the forest. . and who might turn up there?
Although the risk of meeting a wolf is small you would have liked to be warned
10 WMO SWFDP Macau 8 April
2013 Anders Persson
-12 0 12 24 36 48 60 72 84 h
ψ
Computer based “accurate-looking” forecast
Dangerous threshold
No risk? No problems? Should we go ahead?
Computer made
weather forecast
(NWP)
11 WMO SWFDP Macau 8 April
2013 Anders Persson
-12 0 12 24 36 48 60 72 84 h
ψ
Computer based “accurate-looking” forecasts are far from perfect
Dangerous threshold Computer made
weather forecast
(NWP)●
●●
● ●
●●●
●
●
●●
obs ●● ●
1. Forecast doesn’t verify “now”
2. Good timing but systematically too low
3. Forecast out of phase
12 WMO SWFDP Macau 8 April
2013 Anders Persson
1st problem: -Is the forecast correct
“now”?
13 WMO SWFDP Macau 8 April
2013 Anders Persson21/04/23 13
0 12 24 36 48 60 72 84 96 h
ψ
obs● ●●
The forecast does not verify “now”
It did not even verify at initial time (t=0)
Most NWP models do not analyse the weather!
The forecasters nudge the forecast towards the observation
The problem with very short computer forecasts
14 WMO SWFDP Macau 8 April
2013 Anders Persson
Most state-of-the-art NWP models do not assimilate weather observations, only:
1. Upper air temperature, wind, relative humidity and winds from radio sondes
2. Radiances from satellites to be converted to temperature and humidity
3. Upper air winds from drifting clouds
4. Surface winds from satellites, ocean based ships and buoys
5. Surface or MSL pressure from land and sea platforms
They do NOT assimilate 10 m winds, 2 m temperatures or dew points, clouds and weather
These are (pretty well) calculated from the other parameters!
15 WMO SWFDP Macau 8 April
2013 Anders Persson
2nd problem: -Are the NWP systematically wrong?
16 WMO SWFDP Macau 8 April
2013 Anders Persson
ψ
obs - Ψ= corr
●●
●
●
●●
●
●
● ●
●
●●
●●corr = AΨ + B
Statistical interpretation (archived data)
17 WMO SWFDP Macau 8 April
2013 Anders Persson21/04/23 17
0 12 24 36 48 60 72 84 96 h
●●
Statistical correction or “calibration”
●●
●
●● ●
ψ
From experience (verification or statistical interpretation) we know that the NWP model underestimates high forecast values, which can be corrected for
The solution to the problem of systematically misleading computer forecasts
18 WMO SWFDP Macau 8 April
2013 Anders Persson
3rd problem: -Is the forecast “jumpy”?
19 WMO SWFDP Macau 8 April
2013 Anders Persson
-12 0 12 24 36 48 60 72 84 h
ψ
Computer based “accurate” forecast can not only be wrong but also “jumpy”
Dangerous threshold
Today’s forecast
yesterday’s forecast
Tomorrow’s forecast
20 WMO SWFDP Macau 8 April
2013 Anders Persson
L
HL
LH
L
HL
HL
HL
LH HH
LL
Downstream development of influence
Day 2
Day 0
Day 4
Energy propagation
21 WMO SWFDP Macau 8 April
2013 Anders Persson
L
But the influence can also be in the opposite direction
Persson-Petersen WMO workshop 1996
Extra-tropical influence → Tropics
22 WMO SWFDP Macau 8 April
2013 Anders Persson
At the same time as we try to improve the initial analysis by
1. Increasing the number of observations2. Improving their quality3. Improving our analysis methods
…. we also do the opposite:
We “tickle” the analysis by imposing perturbations (possible errors) to fins out how it affects the NWP
23 WMO SWFDP Macau 8 April
2013 Anders Persson
Where and how are the atmospheric analyses perturbed?
24 WMO SWFDP Macau 8 April
2013 Anders Persson
Stochasticphysics
everywhere
Singular vectors
Singular vectors
Tropical singular vectors (when
a cyclonic feature is formed)
EDA Singular vectors
EDA Singular vectors
25 WMO SWFDP Macau 8 April
2013 Anders Persson
EDA in action – typhoon Aere over northern Philippines
The first guess is fairly reliable to the SW of the typhoon, but not to the NE of the typhoon
26 WMO SWFDP Macau 8 April
2013 Anders Persson
00 UTC 03 UTC 06 UTC 09 UTC 12 UTC 15 UTC 18 UTC 21 UTC Time
Surface pressure
10 (from June this year 25) EDA short range forecasts are constantly running in parallel randomly perturbed by stochastic physics and varying SST
Now we want to make a new analysis for the 12 UTC forecast
27 WMO SWFDP Macau 8 April
2013 Anders Persson
00 UTC 03 UTC 06 UTC 09 UTC 12 UTC 15 UTC 18 UTC 21 UTC Time
Surface pressure
To arrive at the best possible analysis for 12 UTC we consider all the forecasts 09-21 UTC as 12-hour first guesses in anew assimilation cycle
Assimilation window
To launch a 10 day forecast from here
28 WMO SWFDP Macau 8 April
2013 Anders Persson
09 UTC 12 UTC 15 UTC 18 UTC 21 UTC Time Assimilation window
●●
●
10 forecasts (first guesses)
Observations perturbed within their error estimates
Surface pressure
These 10 forecasts, 12-hour first guesses, are confronted with observation, perturbed to account for observation errors and representativeness
29 WMO SWFDP Macau 8 April
2013 Anders Persson
09 UTC 12 UTC 15 UTC 18 UTC 21 UTC Time Assimilation window
●●
●
Su
rfac
e p
ress
ure
4 DVAR trajectories
Influenced by these observations the 10 first guesses are modified
30 WMO SWFDP Macau 8 April
2013 Anders Persson
09 UTC 12 UTC 15 UTC 18 UTC 21 UTC Time Assimilation window
Su
rfac
e p
ress
ure
4 DVAR trajectories
Influenced by these observations the 10 first guesses are modified
Odd member 3 is perturbed by SV 6 times to
produce members 3, 4, 23, 24, 43 and 44
Even member 8 is perturbed by SV 4 times to produce members 17,
18, 37 and 38
31 WMO SWFDP Macau 8 April
2013 Anders Persson
1 1 2 21 22 41 422 11 12 31 323 3 4 23 24 43 444 13 14 33 345 5 6 25 26 45 466 15 16 35 367 7 8 27 28 47 488 17 18 37 389 9 10 29 30 49 5010 19 20 39 40
EDA memberCorresponding EPS members 1-50
32 WMO SWFDP Macau 8 April
2013 Anders Persson
09 UTC 12 UTC 15 UTC 18 UTC 21 UTC Time Assimilation window Forecast
Su
rfac
e p
ress
ure
Ensemble forecast 50 members perturbed by singular vectors and stochastic physics
EDA forecast 10 members perturbed by stochastic physics, varying SST and perturbed observations
Formally starting from 12 UTC
33 WMO SWFDP Macau 8 April
2013 Anders Persson
-12 0 +12 +24 +36 +48 +60 +72 +84 h
ψ
Exchanging the “accurate” forecast with a more “honest” one
Dangerous threshold
Today’s NWP forecast
Today’s EPS forecast
34 WMO SWFDP Macau 8 April
2013 Anders Persson21/04/23 34
ψ
Correction for systematic errors
-12 0 +12 +24 +36 +48 +60 +72 +84 h
35 WMO SWFDP Macau 8 April
2013 Anders Persson21/04/23 35
ψ
●●
●
● ●
●●●
●
●
●●
obs ●● ●
The final ensemble forecast – with verification
70% 50%
-12 0 +12 +24 +36 +48 +60 +72 +84 h
36 WMO SWFDP Macau 8 April
2013 Anders Persson
Prob(> 15 m/s) 20 March 2013 12 UTC + 156h
Prob(> 15 m/s) 22 March 2013 12 UTC + 108h
Prob(> 15 m/s) 24 March 2013 12 UTC + 60h
Probability maps of the 10 m wind exceeding 15 m/s
+12 h forecast (verification) →
37 WMO SWFDP Macau 8 April
2013 Anders Persson
Probability maps of more than 20 mm rain in 24r h
Prob(> 20 mm/d) 24 March 2013 12 UTC + 60h
Prob(> 20 mm/d) 21 March 2013 00 UTC + 144h
Prob(> 20 mm/d) 22 March 2013 12 UTC + 108h
Prob(> 20 mm/d) 26 March 2013 00 UTC + 24h
38 WMO SWFDP Macau 8 April
2013 Anders Persson
Storm
Tropical storms genesis map 2 March 12 UTC VT 3-5 March
39 WMO SWFDP Macau 8 April
2013 Anders Persson
Tropical cyclones genesis map 2 March 12 UTC VT 3-5 March
40 WMO SWFDP Macau 8 April
2013 Anders Persson
Tropical cyclones genesis map 3 March 00 UTC VT 4-6 March
41 WMO SWFDP Macau 8 April
2013 Anders Persson
The TC was born on the 6 March!
6 March 00 UTC ensemble plume
7 March 12 UTC ensemble plume
42 WMO SWFDP Macau 8 April
2013 Anders Persson
9 March 12 UTC ensemble plume
11 March 12 UTC ensemble plume
43 WMO SWFDP Macau 8 April
2013 Anders Persson
Summary:
Ensemble forecasts help us
1. To judge the (un)certainty of the weather situation
2. To acquire probability estimates of anomalous events (extreme or high impact)
3. To get the most accurate and least “jumpy” deterministic forecast value
44 WMO SWFDP Macau 8 April
2013 Anders Persson
Other advantages with ensemble forecasting
45 WMO SWFDP Macau 8 April
2013 Anders Persson
Southern Sweden
Central Sweden
The jumpiness is decreased by 50%-75%
24 hour ”jumpiness” of 2 m temperature forecasts
ECMWF
Ens Mean
Ens Mean
ECMWF
46 WMO SWFDP Macau 8 April
2013 Anders Persson
Error decreased after lagging
Jumpiness decreased
even more
Lagging reduced the MA error by 20% but the jumpiness by 70%
47 WMO SWFDP Macau 8 April
2013 Anders Persson
1.0
a
f
g
h ●
●●
Averaging will decrease error by 13%
87.05.01 2
…and jumpiness
by 50%
Why does an ensemble technique affect the jumpiness more than the error?? Look at a small ensemble of consecutive forecasts
f
gh
gh
a
0.50
error = f-a g-a
h-a
Mean of g and herror
48 WMO SWFDP Macau 8 April
2013 Anders Persson21/04/23 48
The perturbations on average make the analysis
worse
On average 35% of the perturbed analyses are better
49 WMO SWFDP Macau 8 April
2013 Anders Persson21/04/23 49
The perturbed forecasts are individually on average 1-1½ day worse than the unperturbed forecast
50 WMO SWFDP Macau 8 April
2013 Anders Persson
L
HL
LH
L
HL
HL
HL
LH HH
LL
Downstream development of influence
Day 2
Day 0
Day 4
Analysis perturbed
Response
Response
51 WMO SWFDP Macau 8 April
2013 Anders Persson
L
HL
LH
L
HL
HL
HL
LH HH
LL
Downstream development influence
Day 2
Day 0
Day 4
Response
Response
Analysis perturbed