Transcript Ensemble Clustering

A seamless system for probabilistic
forecasts of energy demand:
days to seasons
Judith Curry
James Belanger
Mark Jelinek
Violeta Toma
Peter Webster
Tropical Cyclones
10-32 Day Temperatures
The 7th Annual Earth
Networks Energy Weather
Seminar:
Winter Outlook 2012-13
1-15 Day Temperatures
Seasonal Outlook
ECMWF Integrated
Forecasting System
(IFS)
High Resolution
Medium-Range
Extended-Range
Long-Range
Day 0-10
1 member
16 km
Day 0-10
51 ensemble mem
32 km
Day 10-32
51 ensemble mem
64 km
Month 0-13
51 ensemble mem
80 km
Statistical Post-Processing
Basis:
 Reforecasts/hindcasts
 Recent model performance
Statistical methods:
 Bayesian bias correction
 Quantile-to-Quantile
distribution calibration
 Model Output Statistics
(MOS)
IFS allows for consistent and integrated statistical post processing
Comparison of two different post processing schemes
forF 1-15 day U.S. temperature forecasts (6/12 – 9/12)
o
7
RMSE
6
ECMWF raw ens mean
Method 1 (operational)
Method 2 (test)
5
4
3
2
7
RMSE
6
5
4
3
2
1 2 3 4 5 6 7 8 9 10 11 12 13 14
U.S. Daily Temperature Forecasts
Input: ECMWF
Variable Ensemble
Prediction System
Q-to-Q Mapping
Developed from
Hindcast Products
Variable Averaging
Bias Correction
Using Recent
Forecast Skill
Output:
Deterministic &
Probabilistic Daily
Max & Min Temp
Deterministic:
Daily Max/Min
Temperature Forecasts for
105 U.S. Cities Based on
Energy Trading Regions
Probabilistic:
Daily Max/Min
Temperature
Interpercentile Plumes
for Each City
4/12/2012
6
Ensemble distribution of seasonal forecasts
Hurricane Sandy Forecast: 10/23 12Z
(Oct 30 landfall)
ECMWF raw tracks
ECMWF bias corrected tracks
Bias corrected tracks gave 2 days advantage for landfall forecast
Tropical Cyclones: Monthly Outlooks
Input: ECMWF
Monthly Forecast
and Hindcasts
Bias-track
adjustment for TCs
forming in the
eastern Atlantic
ECMWF Forecast - Climatology
Determine prob.
bias-correction from
model and obs.
climate
Output:
Bias-corrected track
density probabilities
and anomalies
Hindcast Calibrated Forecast
Observed Tropical Cyclones in Black
• Contours show bias-calibrated probability of a tropical cyclone for specified
forecast period and shading denotes anomaly relative to climate
• Forecast confidence assigned based on phase and amplitude of the
4/12/2012
Madden-Julian Oscillation
9
U.S. Monthly Temperature Forecasts
Input: ECMWF
Monthly Forecast
and Hindcasts
Theoretical Extreme
Value Distribution from
Hindcast Products
Output: Probabilistic Extreme
Temperature and Heat/Cold
Wave Forecast
Heat/Cold Wave
Probability:
Weekly Departures from
Normal and Probability of
Exceedances
Output: Regional
Temperature Outlook
w/Forecast Confidence
Regional &
Averaged
Outlooks:
4/12/2012
10
Objective Forecast Confidence Assessment
•
•
•
•
Historical predictability analyses
Recent prediction verification statistics
Phase and amplitude of the MJO and ENSO
Spread of the forecast ensemble members and
intercorrelation of ensemble members
• Relationship between ensemble
spread and forecast error
conditioned on teleconnection
regimes.
Ensemble Clustering:
Grouping Members of Forecast Ensembles
Clustering strategies:
 Self clustering
 Regimes
 Initial verification
 HRES forecast
 Subsequent shorter
term forecasts
Cluster
Ensemble Mean
Seasonal
Forecast
Clustering
TC Track Cluster: Ophelia (2011)
ECMWF Ensembles and HRES
Cluster
VarEPS
Deterministic
Mean VarEPS
Observations
VarEPS Cluster
Mean VarEPS Cluster
Observations
•
Cluster: Top five ensemble members whose correlation coefficient with the
ECMWF HRES track is largest during the first 72 hr
•
Working Hypothesis: When the ensemble spread is large, the cluster is more
likely to align closer to observations than either the HRES or ensemble mean
Conclusions
ECMWF Integrated Forecast System enables:
• Internally consistent postprocessing across time scales
• Internally consistent and hierarchical predictability and forecast
confidence assessment
• Hierarchical ensemble clustering strategies
Postprocessed IFS forecasts are competitive with multi-model
Ensembles (better for extreme events)
Addressing distributional errors is essential for
extreme event forecasts
There is untapped prediction skill in ensemble
Interpretation through clustering