Verification of CAPS Storm-scale Ensemble Forecasts from the 2017 HMT FFaIR Experiment and a Patchwise Localized Probability Matched Mean for Ensemble Precipitation Forecasts

For approximately four weeks during the summer of 2017, the Center for Analysis and Prediction of Storms (CAPS) ran a real-time ensemble forecast in support of the Hydrometeorological Testbed (HMT) Flash Flood and Intense Rainfall (FFaIR) experiment. The forecast ensemble included 10 WRF-ARW members and 1 FV3 member and covered the contiguous United States at 3 km horizontal grid spacing. The ARW members included diversity in microphysics and planetary boundary layer (PBL) schemes. This ensemble was used to produce quantitative precipitation forecasts and support experimental flash flood forecasting during the 2017 FFaIR.

Probability matching is often applied to produce ensemble mean precipitation forecasts with greater skill than a simple ensemble mean, but the probability-matched mean often suffers from a lack of small-scale structure, and its output is sensitive to the size of the model domain. To address these shortcomings in the probability-matched mean, a patchwise algorithm for producing a localized probability-matched ensemble mean (LPM) was developed, and was applied during FFaIR to produce forecasts of accumulated precipitation. To calculate the LPM, the domain is divided into a set of rectangular local patches, with each patch centered within a larger, rectangular calculation area. The patches do not overlap, but the calculation areas of adjacent or nearby patches do overlap; in other words, the LPM mean at all the grid points within a given local patch uses the same set of precipitation data, taken from the larger calculation area.

Precipitation forecasts from the CAPS storm-scale ensemble will be compared to operational forecasts produced using HRRR and HREFv2, as well as verified objectively. The LPM will be compared to other ensemble mean products, and to forecasts of individual ensemble members. Compared to neighborhood-based localized PM mean algorithms where a unique PM calculation is performed at each model grid point, the patchwise LPM produces comparable results at much lower computational expense. Compared to the simple ensemble mean and the conventional PM mean, LPM mean exhibits improved retention of small-scale structures, evident in both 2D forecast fields and variance spectra.