Date post: | 05-Jan-2016 |
Category: |
Documents |
Upload: | rodger-owens |
View: | 217 times |
Download: | 2 times |
Gale Warning for Swiss Lakes and Regional Aerodromes based on Ensemble Genetic Programming
L. Mayoraz (1), J. Ambühl (1), R. Voisard (2), C. Voisard (1), M. Züger (2), H. Romang (1) (1) MeteoSwiss, Zurich, Switzerland, (2) University of Zurich, Switzerland
Federal Department of Home Affairs FDHAFederal Office of Meteorology and Climatology MeteoSwiss
Schweizerische EidgenossenschaftConfédération suisseConfederazione SvizzeraConfederazium svizra
Swiss Confederation
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
2
Goal of Project GenWarn
Development of a semi-automatic short-term warning system for gale on Swiss lakes and regional
aerodromes, sending warning proposals to forecasters, based on genetic programming.
14th EMS 07.10.2014
www.kweeper.comaviaswiss.xooit.com
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
3
Context / Current Situation
• Strong gusts (≥ 25 kt) = potential danger to aviation and maritime safety Gale warnings
• In Switzerland, gale warnings are issued for more than 50 lakes and aerodromes but not automated→ First wind gust frequently missed: Low hit rates!
• Benefit of GenWarn System: supports the forecasters in their ongoing weather surveillance and alerts them by proposing potential gale warnings
14th EMS 07.10.2014
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
4
Wind Gust ≥ 25 kt in the next 3 hours?
Method
14th EMS 07.10.2014
Development Phase (1X, with historical data)
Observations
COSMO-2 Forecasts
Evolutionary Algorithm
20 Java Methods
Optimal Probability
Threshold q* Verification
Predictor list (from a 2-year data set):- Observations from several observation stations- Forecasts from the COSMO-2 model
t0t0-1h t0+1h t0+2h t0+3h
Current timet0+0.5h
Observations Forecasts
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
5
Method
14th EMS 07.10.2014
Development Phase (1X, with historical data)
Observations
COSMO-2 Forecasts
Evolutionary Algorithm
20 Java Methods
Optimal Probability
Threshold q* Verification
Genetic ProgrammingMachine learning technique inspired by the evolution theory of species used for optimization problems.
1) Creation of a random population of computer programs from the predictor list. (= gen. 0)
2) Evaluation of the programs. Fitness function = Hit Rate * (1 - False Alarm Ratio) * 100
3) Selection of the best programs and application of crossing and mutation processes on the selection (= gen. 1)
4) Repetition of steps 2 to 3 until the maximum number of generations is reached.
20 X
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
6
Method
14th EMS 07.10.2014
Development Phase (1X, with historical data)
Observations
COSMO-2 Forecasts
Evolutionary Algorithm
20 Java Methods
Optimal Probability
Threshold q* Verification
Output of evolutionary algorithm 20 java methods forecasting the maximum wind gust in the next hours
plus.evaluate (max.evaluate (mmo, fxxs), sine.evaluate (minus.evaluate (dmo, max.evaluate (minus.evaluate (pow.evaluate (min.evaluate (qfdif, sine.evaluate (1.71)), sine.evaluate (multiply.evaluate (4.16, mmo))), pow.evaluate (min.evaluate (6.27, divide.evaluate (0.52, wshe)), log.evaluate (plus.evaluate (pow.evaluate (min.evaluate (6.27, minus.evaluate (dmo, max.evaluate (fxxs, 6.12))), log.evaluate (plus.evaluate(f00, ttt))), fxxs)))), mmo))))
Example of Java Method:
+
max
mmofxxs
sin
-
dmomax
-
^
log
+
^
min
-
max
fxx6.12
dmo
6.27
log
+
f00ttt
fxxs
min
6.27/
wshe0.53
min
qfdif^
sin
1.71
*
mmosin
4.16
mmo
Tree Representation
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
7
Method
14th EMS 07.10.2014
Development Phase (1X, with historical data)
Observations
COSMO-2 Forecasts
Evolutionary Algorithm
20 Java Methods
Optimal Probability
Threshold q* Verification
+max
mmo
fxxs
sin-d
mo
max-
^log+
^min-m
ax
fxx
6.12
dmo
6.27
log+f
00
ttt
fxxs
min
6.27
/wshe
0.53
min
qfdif
^sin
1.71
*mmo
sin
4.16
mmo
/ max / ca
pe
6.5
fxxs
* min m
ax -
m̂in 6.
27
/ wshe
0.53
min qf
dif
^ sin1.71
* mmo
sin4.16
mmo
ttt
fxxs
^
min
6.27/
wshe0.53
min
qfdif^
sin
1.71
*
mmosin
4.16
+ max
mmo
fxxs
sin- d
mo
max -
^ log+
^ min - m
ax
fxx
6.12
dmo
6.27
log+f
00
ttt
fxxs
min
6.27
/wshe
0.53
min
qfdif
^ sin
1.71
*mmo
sin
4.16
mmo
Output of evolutionary algorithm 20 java methods forecasting the maximum wind gust in the next hours
Herd of java methods ensemble forecast
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
8
Method
14th EMS 07.10.2014
Development Phase (1X, with historical data)
Observations
COSMO-2 Forecasts
Evolutionary Algorithm
20 Java Methods
Optimal Probability
Threshold q* Verification
Verification:“ROC-Curve”
False Alarm Ratio
Hit
Rate
Probability of occurrence in %
-Event-based-On a 2-year independent data set
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
9
False Alarm Ratio
Hit
Rate
Probability of occurrence in %
Method
14th EMS 07.10.2014
Development Phase (1X, with historical data)
Observations
COSMO-2 Forecasts
Evolutionary Algorithm
20 Java Methods
Optimal Probability
Threshold q* Verification
Verification:“ROC-Curve”
-Event-based-On a 2-year independent data set
Fitness Function = HR*(1-FAR)
q*: optimal probability threshold
q*: probability of occurrence above which an alarm proposal is sent
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
10
Method
14th EMS 07.10.2014
Development Phase (1X, with historical data)
Operational Routine
Observations
COSMO-2 Forecasts
Evolutionary Algorithm
20 Java Methods
Optimal Probability
Threshold q*
Observations
COSMO-2 Forecasts
Alarm Proposal
Probability P that wind
gust ≥ 25 ktIf P ≥ q*
Verification
For each warning object:
Classminus.evaluate(4.561881696354005, min.evaluate(divide.evaluate(9.9164395430037,
divide.evaluate(max.evaluate(pow.evaluat96354005, in.evaluate(divide.evaluate(9.9164395430037,)
divide.evaluate(max.evaluate(pow.evaluate(1.6373608239449522, fmo), tt40),
min.evaluate(divide.evaluate(9.9164395430037, ddd), f20)))
…
Method 1Method 2Method 3Method 4
every 10 min
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
11
ResultsVariation of Meteorological Threshold Q
14th EMS 07.10.2014
Maximum Fitness: - If Q = 25 kt : ~ 30- If Q = 12 kt : ~ 45
A clear performance limit is reached at this point. (HR ~95% , FAR ~70%)
Performance of system is higher for Q = 12 kt Storm events stronger than 25 kt are too rare for the system to detect
them correctly (tendency of detecting too many events)
Verification on the 2-year data set
Threshold Q = 25 knots Threshold Q = 25 knots Threshold Q = 12 knots
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
12
ResultsComparison with Forecasters Performance
14th EMS 07.10.2014
Overall increase in HR induced by GenWarn System Contribution of GenWarn System variable, object-dependent Role of forecaster: decrease the FAR
Typical ROC Curve GenWarn Vs. Forecasters Performance
Forecaster Experience
GenWarn System Typical Performance
Forecasters Performance per Warning Object
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
13
Conclusions
14th EMS 07.10.2014
• GenWarn gale warning system based on genetic programming shows so far the performance : Hit Rate ~95%, FAR ~70%
• General increase of hit rate when using the GenWarn System compared to the actual forecasters performance best solution: mix machine & forecaster to lower the FAR
• Outlook: – Try with additional predictors: radar data, INCA-
forecasts (nowcasting product)– Operationalization, in situ tests
Gale Warning for Swiss Lakes and Regional Aerodromes Lysiane Mayoraz
14
Thank You!
14th EMS 07.10.2014
Questions?
© Sebastien Marti/Scoopmobile