When can high-resolution NWP model forecasts be trusted? Using error characteristics of the initial conditions to evaluate the likelihood of accurate forecasts

Rigorous evaluation of limited-area high-resolution, convection-allowing (no cumulus parameterization) numerical weather prediction model (CAM) output (e.g., the 2005-2010 NOAA Hazardous Weather Testbed Spring Experiments) has revealed that on some days CAM forecasts are remarkably accurate and provide valuable guidance for severe weather forecasters, while on other days, CAM solutions are wildly incorrect. It has also been demonstrated qualitatively that CAM forecasts are very sensitive to their initial and lateral boundary conditions that are typically provided by a coarser-resolution, larger-domain “parent” model. Thus, it is likely that the high day-to-day variability of CAM forecast accuracy is modulated by the quality of the initial conditions (ICs) provided by the parent.

Inspired by the “hit or miss” nature of CAM forecasts, this study introduces simple quantitative methods to explore whether error characteristics of the ICs provided by the parent model can be utilized as a predictor of 12- to 36-hr CAM forecast accuracy. Convection-allowing 3-km WRF-ARW forecasts covering most of the United States produced at NCAR during the 2007 and 2010 spring seasons were used to investigate the relationship between the ICs and CAM output. WRF forecasts in 2007 were initialized from 40-km North American Mesoscale (NAM) model grids, while model fields from the 13-km Rapid Update Cycle (RUC) model provided ICs in 2010. For each initialization, the parent's fields were compared to radiosonde sites considered as “truth.” Then, the corresponding WRF forecast was also verified against soundings. Examining the relationship between the parent's errors at the initial time and WRF forecast errors provided insight regarding the predictability of the WRF forecasts. WRF precipitation forecasts were also compared to the parent's errors.

In the 2007 dataset, error characteristics of the ICs provided no indication regarding the likely accuracy of next-day WRF forecasts (i.e., “good” ICs did not necessarily mean good subsequent forecasts). However, results from 2010 indicated that the quality of the ICs was a useful predictor of next-day WRF forecast accuracy. Although there were many configuration differences between the 2007 and 2010 datasets, conflicting results between the two years may be largely attributable to changes in the parent's resolution and improved ICs in 2010 due to the RUC's hourly cyclic data assimilation.

Given the results from 2010, additional analyses using this method are planned and encouraged. It may be possible to ultimately define a near real-time metric based on the parent's errors that severe weather forecasters can use to help determine how much confidence to place in a particular CAM solution.