An 18 member ensemble of the 10 km simulations was created by varying the convective parameterization (Betts-Miller-Janjic, Kain-Fritsch, no scheme, both schemes alternating), the convective time step, and by adjusting the initial conditions to better represent mesoscale features such as cold pools, outflow boundaries, or moist tongues, through the use of surface mesonetwork observations or radar data. In addition, more extensive variation of initial conditions was investigated by creating some ensemble members whose initial perturbations from a mean state varied by a small amount from the control runs.
The 10 km ensemble results indicate that ensemble spread is rather small for adjustments to better represent mesoscale conditions. Changes in moist physics result in much larger spread. For these difficult-to-forecast convective systems which are often nocturnal and elevated, the ensemble guidance may be of limited value because the verifying scenario can often differ substantially from all members of the ensemble. Reliability diagrams show some tendency for higher probabilities of occurrence when forecast probabilities are high, but the curves hover near the no skill line, as has been found in other studies for warm season precipitation forecasts using coarser grid mesh ensemble members. An overprediction of probability of precipitation generally occurs for all precipitation thresholds. Although other measures of skill such as the area under the relative operating characteristic curve and Brier scores for this 10 km ensemble also imply only modest predictive ability, the probabilistic quantitative precipitation forecasts from the ensemble nonetheless appear to be more skillful than any single 10 km deterministic run.