SBIR-STTR Award

The accurate numerical prediction of hazardous airborne plumes requires two important capabilities. First, meteorological conditions at fine spatial scale both at the time of plume release as well as a few hours into the future, and second, quantification of this information in a statistically reliable probabilistic framework. The proposed study will use a combination of fine-scale ensemble re-forecasts, as well as historical surface observations, to achieve the goals of the solicitation. Because no unique method exists for doing so, we examine several and will pursue that which is most accurate, efficient and adaptable to future needs. The first approach, arguably the most simple and computationally efficient, involves using a 20-year history of surface observations to create regression equations that yield statistically reliable probabilistic point forecasts given current conditions. At the other extreme, fine-scale reforecasts will be generated from historical re-analyses and both linear and nonlinear regression approaches applied for calibration. Uniquely, we combine the 20-years of historical observations with this framework such that the final outcome is a combination of dynamical forecast and observation-based statistics. Finally, we expand the nearest-neighbor analog method using ensemble reforecasts alone or in combination with historical surface observations.

Fine-scale, well-calibrated probabilistic weather forecasts are increasingly in demand for such weather-critical applications as battlespace management, and accurate prediction of high-value airborne operations. Built upon the Phase I effort we have just completed, the principal objective of this proposed study (Phase II) is to develop a reliable, efficient, and easily deployable reforecast-calibration system prototype based on fine-scale ensemble reforecast datasets and use it to produce well-calibrated, fine-scale probabilistic weather forecast products. This study consists of three major components. The first component is to produce a basic 20-year fine-scale ensemble reforecast dataset over the CONUS domain using the WRF-ARW modeling system. The second component is to further develop and examine various MOS-based statistical models that apply to more variables and multiple classes and to refine the Phase I KNN calibration techniques, using the large reforecast sample base, and to compare the two techniques. The third component is to conduct sensitivity experiments to study data-sparse impact and the trade-offs between ensemble size and reforecast length, in order to assess the feasibility and effectiveness of the deployable reforecast-calibration system.

Keywords:
Meteorology, Bayesian Model Averaging, Linear Discriminant Analysis, K Nearest Neighbor, Ensemble Forecast Model, Fine-Scale Probabilisitic Weather Mo