Estimation of skill and uncertainty in regional numerical models


  • The National Center for Atmospheric Research is sponsored by the National Science Foundation.


The results from 72-hour simulations and forecasts from the Penn State/NCAR limited-area model, in which a number of numerical and physical factors are varied, are analysed to understand the contribution to model error or uncertainty introduced by these factors. The factors include the initial conditions, horizontal resolution (80 km and 160 km) and domain size, lateral boundary conditions, and physical parametrizations. The results are compared with those of a global forecast model (the NCAR Community Climate Model). The simulations and forecasts are verified both for 12 individual cases and for the ensemble average of the 12 cases, using several objective measures of skill. The differences in these skill scores between different models are tested for their statistical significance.

The use of observed lateral boundary conditions (LBC) exerts a strong control on the growth of errors over a domain size 3600 × 4800 km2. Objective measures of error show little growth beyond about 36 hours, so that 72-hour errors are nearly as low as the 36-hour errors. On these time and space scales, the quality of the LBC is more important than any other factor tested in the temporal evolution of the model errors. These results show that the large-scale atmospheric motions have a major effect on the evolution of small-scale features in the model.

Small variations in initial conditions have little effect on the model skill beyond 12–24 hours. Of all the factors examined, small uncertainties or errors in the initial conditions have the least effect on model skill beyond 12 hours.

Latent heating effects associated with condensation and precipitation, and sensible and latent heat fluxes from the surface have a statistically significant effect on model skill. However, relatively simple parametrizations of these effects produce nearly the same skill as do more complex schemes for the cases studied here.

There is a large variation in model skill from case to case; some cases are easier to forecast than others. Also, there is much greater variation in different model performances for a single case than when averaged over all 12 cases.

The concept of climatological use and verification of regional models is introduced. The climatological skill of the control model version which includes all physical processes appears to be quite good. The skill scores of the ensemble average simulation are considerably better than the average scores for individual cases, the model has small bias errors, and the horizontal structure of the model-simulated atmosphere is similar to the observed structure for scales of motion resolved by the upper-air observational network over the United States.