| dc.description.abstract | This dissertation research was undertaken to better understand the optimal design of convection-permitting storm-scale ensemble forecast (SSEF) systems that resolve features ranging from synoptic to convective scales. The focus of the research is on the data assimilation (DA) and initial condition (IC) perturbation methods, both of which are uniquely affected by the multi-scale interactions inherent in an SSEF system. There are four components to this research. First, a GSI-based DA system is implemented in the multi-scale scenario with observations ranging from synoptic scale rawinsonde to convective scale radar observations. The GSI-based 3DVar and EnKF techniques are also compared to each other in this multi-scale context. Second, the systematic sensitivities of convection forecasts to different simple methods of IC perturbation are evaluated. Third, Observation System Simulation Experiments (OSSES) are conducted using ensemble analyses generated with the GSI-based EnKF to understand the impacts of different methods of generating more complex flow-dependent multi-scale IC perturbations. Fourth, the impacts of inconsistencies between the initial and lateral boundary condition (LBC) perturbations are evaluated as well as the impacts of model and physics errors in non-OSSE real-data experiments.
In the first part of this research, the multi-scale GSI-based EnKF and 3DVar techniques are systematically compared to each other to better understand the impacts of their differences on the analyses at multiple scales and the subsequent convective scale probabilistic forecasts. Averaged over ten diverse cases, 8h forecasts of hourly accumulated precipitation initialized using GSI-based EnKF are more skillful than those initialized using GSI-based 3DVar, both with and without storm-scale radar DA. The advantage from radar DA persists for ~5h using EnKF, but only ~1h using 3DVar. A case study of an upscale growing MCS is also examined. The better EnKF-initialized forecast is attributed to more accurate analyses of both the mesoscale environment and the storm scale features. The mesoscale location and structure of a warm front is more accurately analyzed using EnKF than 3DVar. Furthermore, storms in the EnKF multi-scale analysis are maintained during the subsequent forecast period. However, storms in the 3DVar multi-scale analysis are not maintained and generate excessive cold pools. Therefore, while the EnKF forecast with radar DA remains better than the forecast without radar DA throughout the forecast period, the 3DVar forecast quality is degraded by radar DA after the first hour. Diagnostics revealed that the inferior analysis at meso- and storm-scales for the 3DVar is primarily due to the lack of flow-dependence and coherent cross-variable correlation, respectively, in the 3DVar static background error covariance.
In the second part of this research, multi-scale precipitation forecast sensitivities are examined for two events and systematically over 34 events out to 30-h lead time using Haar Wavelet decomposition of hourly accumulated precipitation. The impacts of two small scale IC perturbation methods are compared to the larger scale IC and physics perturbations included in an experimental convection-allowing ensemble. For an event where the forecast precipitation is driven primarily by a synoptic scale baroclinic disturbance, small scale IC perturbations result in little precipitation forecast perturbation energy on medium and large scales, compared to larger scale IC and physics (LGPH) perturbations after the first few forecast hours. However, for an event where forecast convection at the initial time grows upscale into a Mesoscale Convective System (MCS), small scale IC and LGPH perturbations result in similar forecast perturbation energy on all scales after about 12 hours. Averaged over 34 forecasts, the small scale IC perturbations have little impact on large forecast scales while LGPH accounts for about half of the error energy on such scales. The impact of small scale IC perturbations is also less than, but comparable to, the impact of LGPH perturbations on medium scales. On small scales, the impact of small scale IC perturbations is at least as large as the LGPH perturbations. The spatial structure of small scale IC perturbations also affects the evolution of forecast perturbations, especially at medium scales. For these random homogeneous small scale IC perturbations, there is little additional impact of the small scale IC perturbations when added to LGPH. Additional study of more realistic flow-dependent IC perturbations, and their impacts on ensemble forecast skill in addition to deterministic forecast sensitivity, are therefore motivated.
In the third part of this research, the impacts of multi-scale flow-dependent IC perturbations (MULTI) for SSEFs are investigated using perfect model OSSEs. The MULTI perturbations are compared to downscaled IC perturbations from a larger scale ensemble (LARGE). Forecasts initialized at different stages of the upscale growth of an MCS case study are first used to qualitatively understand the impacts of the IC perturbation methods. Scale-dependence of the results is assessed by evaluating two-hour storm-scale reflectivity forecasts in 0-48km neighborhoods separately from hourly accumulated precipitation forecasts in mesoscale neighborhoods with a 48-km radius. For the reflectivity forecasts over small neighborhood radii (0-8km), the small scales of IC perturbation, resolved in MULTI but not LARGE, are advantageous for about 1h. For reflectivity forecasts at larger radii and for mesoscale precipitation forecasts, the differences in IC perturbations on scales resolved by both MULTI and LARGE dominate the forecast skill. The MULTI IC perturbations are more consistent with the analysis uncertainty than the LARGE IC perturbations in the vicinity of the developing MCS. However, an area of spurious convection away from the observed MCS contains unrealistically large mid-level moisture perturbations for MULTI that can have the effect of enhancing the spurious convection. The relative importance of these differences between MULTI and LARGE, and their effects on forecast skill, depends on when during the MCS upscale growth process the forecasts are initialized.
The perfect-model OSSE case study is also extended to 11 diverse cases. The mesoscale precipitation forecasts from MULTI are systematically more skillful than LARGE at 1h and ~5-9h lead times. This is due to the smaller magnitude mesoscale IC perturbations near analyzed convective systems for MULTI that are more consistent with the analysis uncertainty than for LARGE. This difference also leads to systematically more skillful reflectivity forecasts for MULTI than LARGE using radii >4km. The reflectivity forecasts using radii of 0-4km are systematically more skillful for MULTI than LARGE during the first hour due to the presence of the small scale IC perturbations. The small scale IC perturbations also systematically contribute to further improving the MULTI mesoscale precipitation forecasts after ~5h.
In the final part of this research, two considerations for operational application of the multi-scale IC perturbation methods are investigated. First, the impact of inconsistencies between the multi-scale IC perturbations and mesoscale LBC perturbations is evaluated. Spurious pressure waves originating at the LBCs result from this inconsistency. However, unlike previous studies with a larger resolution difference between then inner and outer domains and with different DA methods on each domain, significant impacts on convective scale probabilistic forecast skill are not found with the multi-scale GSI-based DA system. Second, real-data experiments with model error are used to further understand the practical implications of the OSSE results. In real-data experiments, LARGE is generally more skillful than MULTI except for reflectivity forecasts at short lead times of ~30-90 minutes, depending on spatial scale. This is because the larger magnitude mesoscale IC perturbations in LARGE compensate for unrepresented model errors. The flow-dependent small-scale IC perturbations are even more important for storm-scale reflectivity forecasts in the ensemble with unrepresented model error than in the perfect-model OSSEs. | en_US |
