Verification and Optimization of a Patchwise Localized Probability Matched Mean for Ensemble QPF

As part of research conducted by the Center for Analysis and Prediction of Storms (CAPS) for the NOAA Hydrometeorology Testbed (HMT) Flash Flood and Intense Rainfall (FFaIR) experiment, we have developed a patchwise, localized probability matched mean (LPM mean) algorithm for ensemble consensus quantitative precipitation forecasts (QPF). The LPM mean is designed to mitigate some of the flaws of the standard probability-matched mean (PM mean), such as overly-smoothed features arising from taking a domain-wide ensemble mean, and the commingling of precipitation values from very different weather regimes and locations (for example, mixing values from tropical convection and mid-latitude stratiform rain) in the calculation of PM mean rainfall at a given location.

To calculate the patchwise LPM mean, a large model domain (in this case, one covering the continental United States (CONUS)) is broken down into a series of rectangular patches, each associated with a larger LPM calculation domain centered on the patch. The LPM mean within each patch is calculated by taking the PM mean over the associated calculation domain; the patches are then stitched together to form the LPM mean field for the entire model domain. The LPM calculation domains overlap, helping to minimize the potential for discontinuities in the LPM mean field at patch boundaries. As a final step, a Gaussian smoother can optionally be applied to smooth grid-scale noise and further reduce the potential for discontinuities. The LPM mean algorithm is thus designed to retain local structures in the QPF field while being relatively efficient to compute (the patchwise LPM requires only a single PM calculation for each patch, as opposed to a PM calculation for every gridpoint, as is the case for previously-developed pointwise LPM mean algorithms). The ability to select the patch size, the LPM domain size, and the degree of smoothing used allows for considerable flexibility in configuring the LPM mean, allowing the user to select preferred length scales over which to consider data (via the LPM domain size), the computational efficiency (via patch size) and the degree to which grid-scale structures are damped (via the smoother setting).

The LPM mean was produced for the CAPS storm-scale ensemble forecast (SSEF), consisting of 13 WRF-ARW and 2 FV3 members at 3-km grid spacing, and provided in real-time to participants of the 2018 HMT FFaIR experiment. Results indicate that the LPM mean is successful in retaining local structures within the CONUS forecast domain. The LPM mean outperforms the PM mean and simple ensemble mean in terms of fractions skill score (FSS) and variance spectra, while remaining comparable to the PM mean in terms of other objective skill score metrics. Participants responded favorably to the LPM mean, giving it higher average subjective ratings than any of the ensemble or PM mean products evaluated. Based on this performance, the LPM mean was recommended for transition to operations and has been tested at EMC with the HREFv3 model, where it has been found to perform quite well, particularly with regard to precipitation forecast bias. Results from the 2018 HMT FFaIR experiment will be presented, and the sensitivity of the LPM mean forecast to choice of configuration parameters will be discussed.