Evaluating probabilistic precipitation forecasts generated by deterministic convection-allowing NWP models
Like coarser resolution, convection-parameterizing models that are run operationally by the NWS, the NSSL-WRF configuration typically suffers from placement errors in predicting the location of heavy rainfall. Compared to the coarser resolution models, however, the NSSL-WRF is relatively accurate in predicting the frequency of heavy precipitation amounts. For example, its long-term frequency bias for the 1-inch/6-h threshold is about 1.1, compared to about 0.5 for the operational NAM model. Thus, we have been exploring post-processing strategies that allow us to take advantage of the demonstrated skill in predicting the occurrence of heavy rain events, while providing meaningful quantitative expressions of the uncertainty in predicting the specific location where these events will occur.
This work explores application of a simple post-processing algorithm that introduces an estimate of spatial uncertainty in the prediction of heavy rainfall. The algorithm uses a two-dimensional Gaussian probability distribution function to define a regional “neighborhood” of non-zero probabilities for an event predicted by the model. The resulting guidance products show good reliability and resolution in preliminary testing and calibration holds promise for even better predictive skill with this technique.
Quantitative precipitation forecasts from the NSSL-WRF will be compared to forecasts from two experimental convection-allowing models being run daily at NOAA/NWS/NCEP/Environmental Modeling Center and to observed precipitation fields. This initial comparison will be made using traditional verification metrics such as the bias and Gilbert Skill scores. A subsequent comparison will use probability forecasts generated from each output file, and the reliability and resolution of these forecasts will be determined and compared. The potential value of these forecasts as guidance for flash flood prediction will be discussed.