Ensemble Forecasting System
Leader of the project: Dr. Jen-Cheng Joseph Chang
One of the proposed tasks in the 5-year NWP Project Phase III is to extend the CWB's
weather service to the 2nd-week forecasts. The progress of this task is further devided
into 3 phases, which fully depends on the available resources from the Meteorolofical
Information Center (MIC) of the CWB.
Phases of Ensemble Forecast System:
Economical (or Simple) Ensemble
- Lagged-Average Forecasting
- Optimal Initial Perturbation
Phase I -- Economical (or Simple) Ensemble
Implementation: To be implemented by the end of 1997.
- Forecasting period: 7 days.
- Output variables: 500hPa Height (may be extended to SLP, 850hPa Temp, and 850/200hPa
- Resolution: 5 deg x 5 deg.
- Area covered: Covering the areas north of 20N (Between 20S-20N for the wind fields.)
- Data Source: Use the CWB's GFS outputs and those received daily from other major
operation centers (e.g., ECMWF, NCEP, JMA) to form a different-model ensemble. No
additional time-integration is required.
- Averaging process: The averaging process takes place in spectral space, not in grid
space. The weighting factors are determined by the error statistics of each model verified
with respect to the CWB's global analysis.
Phase II -- Lagged-Average Forecasting (Hoffman & Kalnay 1983)
Implementation: To be implemented in 1999 or 2000, depending on available resources.
- Forecast period: Up to 10 days. (Needs to at least extend to a 14.5-day time-integration
for each run. Currently, the daily major run (at 12Z) is up to 7 days, and the minor run
(at 00Z) is up to 3 days. The model resolution can be reduced, say T21L19, after the 7th
day of each run to save computing resources.)
- Output variables: 500hPa Height, SLP, 850hPa Temp, 850/200hPa Wind. (Can be extended to
include other variables if desired.)
- Resolution: 2.5 degree x 2.5 degree.
- Area covered: Covering the entire NH (or the whole globe).
- Averaging process: Use previous forecasts (up to 4.5 days earlier) of the CWB's GFS as
initial perturbed fields. The weighting factors are determined by the forecasting error
statistics of the first 4.5 days (12 hr apart). Fully loaded LAF (includes all variables
and without reducing resolution) requires about 8 times of current maximum storage for the
Phase III -- Optimal Initial Perturbations
Implementation: Still under research mode. (No plan to be implemented before the end
of the NWP Project Phase III.)
- Options: Two options to consider -- The Breeding Method (Toth & Kalnay 1993; NCEP's
EPS is a combination of this and the LAF) to find the Lyapunov Vectors of a complicated
PE-model, or the Adjoint Method (Buizza et al. 1993) to find the Singular Vectors of a
simplified PE-model. The Breeding Method is much cheaper than the Adjoint Method. However,
the Sigular Vectors form a better set of optimals than the Lyapunov Vectors.