Evaluation of modeled precipitation intensity distributions and their application to short term forecasting

A short-term forecasting system called the Consolidated Storm Predication for Aviation (CoSPA), has been developed for aviation planning. The system blends forecasts of VIL and echo tops from MIT-LL's radar-based extrapolation with model forecasts produced by NOAA GSD's High Resolution Rapid Refresh (HRRR) over the continental United States. A critical component of this blending is the dynamic calibration of the model forecasts to better match the observed frequency distributions of VIL and echo tops. The new technique was run in realtime for the entire summer of 2010 as part of the FAA forecast evaluation project. It is found that the technique, described below, improves forecast skill in terms of CSI and bias for forecasts of over 6 hours at times, and aids in creating a smooth transition between the extrapolation and HRRR.

The dynamic calibration is performed using a frequency matching technique in which a calibration function, defined as the difference between the modeled and observed distributions of a discontinuous field, is used to adjust the model distribution to match that of the observed distribution. When applied to the CONUS, the forecast domain is divided into 54 overlapping tiles. The size of the tiles and overlap are adjustable parameters that were evaluated to optimize performance. The overlap was set to be 50% of the tile size to smoothly transition between correction functions and handle storms that lie on the tile boundaries. The calibration function is applied to forecast leadtimes valid in the future using a leadtime-dependent weighting function. The weighting function is determined by comparing observed and modeled distributions obtained from an archive of cases. These comparisons show that the modeled distributions of precipitation intensity tend to under/over-predict the higher intensity values in the morning/afternoon. The comparisons also show that the typical de-correlation period between calibration functions found for the longer leadtimes and those found at the shorter leadtimes is roughly 6 hours. Results from both the archive analysis model evaluation and the application of the dynamic calibration in realtime will be shown.