Post on 31-Dec-2015
description
The CPC Consolidation Forecast
David UngerDan Collins, Ed O’ Lenic,
Huug van den Dool
NOAA/NWS/NCEP/Climate Prediction Center
Overview
• A regression procedure designed for ensembles.
Derive a relationship between the BEST member of an N-member ensemble and the observation:
Y = a0 + a1fb + ε
Ensemble Regression
• Weights represent the probability of a given member being the best.
• If weights are known, coefficients can be calculated from the ensemble set.
(No need to explicitly identify the best member)
Ensemble Regression
Example ForecastCFS 1-month Lead Forecast
Nino 3.4 SST, May, 1992April Data June-August Mean SST’s
A series of forecasts
• Start with the ensemble mean• Gradually increase the ensemble spread
K = The fraction of the original model spread
Multi Model Consolidation
• At least 25 years of “hindcast” data• Standardize each model (means and standard
deviations)• Remove trend from models and observations• Weight the various models• Perform regression• Add trends onto the results
Nino 3.4 Consolidation
• CFS, CCA, CA, MKV
(Statistical and Dynamic models mixed)• Lead -2 and Lead -1 are a mix of observations
and the one and two-month forecast from the CFS
Skill May Initial TimeCalibrated CFS Vs. Consolidation
CRPS Skill Nino 3.4
0
0.5
1
-2 -1 0 1 2 3 4 5 6
Lead (Months)
CR
PS
S
CFS
CONS
U.S. Temperature and Precipitation Consolidation
• CFS• Canonical Correlation Analysis (CCA)• Screening Multiple Linear
Regression(SMLR)• OCN - Trends.
SON Consolidation Forecast
Performance
.046 .076 .191 -.147 63%
.067 .076 .162 -.334 59%
.063 .100 .215 -.268 73%
.074 .100 .199 -.203 62%
.023 .040 .098 -.858 38%
CCA+SMLR
CFS
CFS+CCA+SMLR, Wts.
All – Equal Wts.
Official
HSSCRPSS RPSS - 3 % CoverBias (C)
Future Work
• Add more tools and models• Improve weighting method• Trends are too strong• Improve method of mixing statistical and
dynamical tools
END
Recursive Regression
• Y = a0 + a1fi
a+ = (1-α) a + α Stats(F,Y)
Stats(F,Y) represents error statistic based on the most recent case
α = .05
a+ = .95 a + .05 Stats(F,Y)
SST Consolidation
• CFS – 42 members (29%)
• Constructed Analog
(CA) – 12 members (18%)
• CCA – 1 member (17%)
• MKV – 1 member (36%)
Advantages
• Ideally suited for dynamic models.• Uses information from the individual
members (Variable confidence, Clusters in solutions, etc.)
Disadvantages• Statistical forecasts are not true Solutions • Trends are double counted when they
accelerate• Weighting is not optimum (Bayesian seems
appropriate)