Operational Ensemble Prediction at the National Meteorological Center: Practical Aspects

M. Steven Tracton Climate Analysis Center, National Meteorological Center, NWS/NOAA, Washington. D.C.

Search for other papers by M. Steven Tracton in
Current site
Google Scholar
PubMed
Close
and
Eugenia Kalnay Development Division, National Meteorological Center, NWS/NOAA, Washington, D.C

Search for other papers by Eugenia Kalnay in
Current site
Google Scholar
PubMed
Close
Full access

Abstract

On 7 December 1992 NMC began operational ensemble prediction. The ensemble configuration provides 14 independent forecasts every day, verifying on days 1 through 10. The ensemble members are generated through a combination of time lagging [Lagged-Average Forecasting] and a new method, Breeding of Growing Modes (Toth and Kalnay). In adopting the ensemble approach, NMC explicitly recognizes that forecasts are stochastic, not deterministic, in nature. There is no single solution, only an array of possibilities, and forecast ensembles provide a rational basis for assessing the range and likelihood of alternative scenarios.

Given the near saturation of computer resources at NMC, implementation of ensemble prediction required a trade-off between model resolution and multiple runs. Before 7 December 1992, NMC was producing a single global forecast through 10 days with the highest-resolution (T126) version possible of its medium-range forecast model. Now, based an experiments that showed no adverse impact upon the quality of forecasts, the T126 model run is truncated to T62 resolution beyond day 6. The computer savings are used to generate the balance of the ensemble members at the lower T62 resolution. While these complementary runs are, on the average, somewhat less skillful when considered individually, it is expected that ensemble averaging will increase skill levels. More importantly, we expect that ensemble prediction will enhance the utility of NWP by (a) providing a basis for the estimation of the reliability of forecasts, and (b) creating a quantitative foundation for probabilistic forecasting.

A major challenge of ensemble prediction is to condense the large amounts of information provided by ensembles into a user-friendly format that can be easily assimilated and used by forecasters. Some examples of output products relevant to operational forecast applications are illustrated. They include the display of each member of the ensemble, ensemble mean and dispersion fields, “clustering” of similar forecasts, and simple probability estimates.

While this implementation of ensemble prediction is relatively modest (ensembles of 14 members for the forecasts encompassing days 1 through 10), it does provide the basis for development of operational experience with ensemble forecasting, and for research directed toward maximizing the utility of NMC's numerical guidance.

Abstract

On 7 December 1992 NMC began operational ensemble prediction. The ensemble configuration provides 14 independent forecasts every day, verifying on days 1 through 10. The ensemble members are generated through a combination of time lagging [Lagged-Average Forecasting] and a new method, Breeding of Growing Modes (Toth and Kalnay). In adopting the ensemble approach, NMC explicitly recognizes that forecasts are stochastic, not deterministic, in nature. There is no single solution, only an array of possibilities, and forecast ensembles provide a rational basis for assessing the range and likelihood of alternative scenarios.

Given the near saturation of computer resources at NMC, implementation of ensemble prediction required a trade-off between model resolution and multiple runs. Before 7 December 1992, NMC was producing a single global forecast through 10 days with the highest-resolution (T126) version possible of its medium-range forecast model. Now, based an experiments that showed no adverse impact upon the quality of forecasts, the T126 model run is truncated to T62 resolution beyond day 6. The computer savings are used to generate the balance of the ensemble members at the lower T62 resolution. While these complementary runs are, on the average, somewhat less skillful when considered individually, it is expected that ensemble averaging will increase skill levels. More importantly, we expect that ensemble prediction will enhance the utility of NWP by (a) providing a basis for the estimation of the reliability of forecasts, and (b) creating a quantitative foundation for probabilistic forecasting.

A major challenge of ensemble prediction is to condense the large amounts of information provided by ensembles into a user-friendly format that can be easily assimilated and used by forecasters. Some examples of output products relevant to operational forecast applications are illustrated. They include the display of each member of the ensemble, ensemble mean and dispersion fields, “clustering” of similar forecasts, and simple probability estimates.

While this implementation of ensemble prediction is relatively modest (ensembles of 14 members for the forecasts encompassing days 1 through 10), it does provide the basis for development of operational experience with ensemble forecasting, and for research directed toward maximizing the utility of NMC's numerical guidance.

Save