Workshop on Advances in Meso- and Micrometeorology
3-4 November 2014 | Donja Stubica, Croatia
Wind Speed Ensemble Predictions with an
Analog-based Method in Complex Terrain
Iris Odak Plenković1, Luca Delle Monache2,
Kristian Horvath1, Mario Hrastinski1, Alica Bajić1
1 Meteorological
and Hydrological Service (DHMZ), Zagreb,
Croatia
2Research Applications Laboratory, NCAR, Boulder, CO, USA
Outline
 Introduction and AE method basics
 Results:
• Deterministic AE forecasting
 General results
 Adjustment to complex terrain
 Different starting models
• Probabilistic AE forecasting (current work)
 Conclusion
Introduction
Analog – based method needs:
• Time series of measurements on location of interest
• Historical NWP on the same location and period (training + verification)
• Current NWP
How does this method work?
1.
For each lead time of a current
prediction it searches the most similar
past NWP in training period considering
several predictors (variables forcasted)
and 𝑡 time steps before/after:
𝑁𝑊𝑃𝑡 𝐴𝑡 ′ =
Current NWP
𝑤𝑖
𝑖=1 𝜎𝑓𝑖
𝑁𝐴
𝐹 – NWP
𝐴 – analog
𝑡 – time (now)
𝑡′ - time (in the past)
𝑡, 𝑗 – time frame
𝑁𝐴 , 𝑖 – predictors
𝑡
𝑗=−𝑡
(𝐹𝑖,𝑡+𝑗 − 𝐴𝑖,𝑡 ′ +𝑗 )2
Introduction
Current NWP
NWPs in training period
9 UTC
3.4
4.2
How does this method work?
2.
For N most similar past NWPs in training
period we choose corresponding
measurements called analogs and
they make ANALOG ENSEMBLE (AE)
3.7
5.0
3.4 3.7 4.2 5.0
AE for current 9 UTC
Deterministic forecast
Probabilistic forecast
Methods
2.5
3.2 3.5 3.9 4.2 4.3
5.0 5.4 6.3 7.0
AE for current 9 UTC
Deterministic forecast:
•
•
•
•
AE mean
Kalman filter of AE mean (AE mean KF)
AE median (AE med)
Kalman filter of sorted AE metrics (KFSM)
Probabilistic forecast
Deterministic AE forecasting
• Trainig period: year 2010 & 2011.
• Verification period: year 2012.
• Starting model: ALADIN regional model with 8 km grid
spacing, 3 h lead time step, up to +72 h, starts at 0 UTC
• 14 stations
• How many analogs to choose? ~15
RMSE
RCC
Bias
0.75
2.4
model
2.3
2.2
1
0.7
0.5
KF
2.1
0.65
2
0
0.6
1.9
1.8
AE med
-0.5
0.55
1.7
1.6
KFSM
AE mean KF
AE mean
5
10
15
20
Nb. of AnEn members
25
0.5
5
10
15
20
Nb. of AnEn members
25
-1
5
10
15
20
Nb. of AnEn members
25
Deterministic AE forecasting
Adjustment to complex terrain
III
I
II
o I:
• Coastal area
• Largest wind
speeds (bora)
o II:
• Higher altitude
• Mountain area
o III:
• Continental part
• Smallest wind
speeds
Deterministic AE forecasting
Group 1
Group 3
Group 2
4
RMSE
model
3
KF
2
KFSM
AE mean
AE mean KF
1
RCC
0.8
0.6
Bias
0.4
0.5
0
-0.5
-1
0
Time
Time [UTC]
[UTC]
+18
+36
+54
0
+18
+36
+54
0
+18
+36
+54
+72
Deterministic AE forecasting
• How does change in horizontal resolution affects AE methods?
→ at Group 1 locations for 09-24h UTC:
ALADIN 8 km: 37 levels; 240 x 216 grid points; 72-hourly forecast, 3 hours output; hydrostatic
ALADIN 2 km: 37 levels; 450 x 450 grid points; 24-hourly forecast; 1 hours output; nonhydrostatic
DADA 2 km: 15 levels; 450 x 450 grid points; 72-hourly forecast; 3 hours output; hydrostatic
0.8
0.7
0
0.6
0.4
-0.3
1.5
0.3
-0.4
-0.5
0.2
-0.6
0.5
0.1
-0.7
0
0
-0.8
KFSM
-0.2
AE mean
0.5
AE mean KF
-0.1
2
1
0.1
KF
2.5
A8
A2
DA
0.2
model
3.5
3
Bias
RCC
RMSE
Deterministic AE forecasting
Critical Success Indeks – Group I:
Category 2
CSI
Category 1
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
Category 3
0.6
model
0.5
KF
AE mean
AE mean KF
KFSM
0.4
0.3
0.2
0.1
0
A8
A2 DA A8
A2 DA A8 A2 DA A8 A2 DA A8 A2 DA
Group
Probabilistic AE forecasting
0.2
0.6
0
0
0.2
0.4
0.6
0.8
1
Reliability diagram:
1 1
0.4
Group II
1
0.8
0.8
Group I
0.6
0.2
0.6
LR
0.8
0.4
AE
0.4
0.2
0
0 0.6
0.2
0.4
0.6
0.8
1
0.2
0
0
0.2
0
0
0.4
0.4
0.2
0.6
0.4
0.8
0.6
1
0.8
1
Forecast probability
1
1
1
0.8
0.2
0.6
Group III
0.8
0.6
0.4
0.8
0.4
0
0
0.2
0.4
0.6
0.8
1
Forecast probability
0.6
0.2
0.2
0
0
0
0
0.2
0.4
0.2
0.6
0.4
0.8
0.6
1
0.8
1
Forecast probability
0.4
1
Conclusion
AE methods:
• Well adjusting to all sorts of terrain
(especially AE mean)
• Reduce RMSE and bias, while improving RCC
• In most cases starting model with 8-km
horizontal resolution produces the best results
• Using higher resolution improves accuracy for
high wind speed forecasting
• Reliably quantify uncertainty
THANK YOU!

E-PAPER AUGUST 2014.pmd