Automated Decision Support Tools (DSTs) being developed for the Next Generation Air Transportation System (NextGen) require probabilistic forecast information that relates the likelihood of aviation-impacting convection with a descriptive set of corresponding characteristics (storm area coverage, orientation, and organization) that convey the severity of the event. The prediction of these features is inherently uncertain and thus should be provided in terms of forecast probabilities. Forecast uncertainty can be quantified using a number of techniques including post-processing of model ensemble data or statistical techniques.
12 hour likelihood forecast for a large-scale convective storm based on a HRRR time-lagged ensemble (left) compared to the observed storm at valid time (right).
NCAR/RAL has been developing new concepts of how probabilistic weather forecasts can be tailored for aviation needs. The focus of this research is on developing tools that can distill large amounts of observational and ensemble model data into a probabilistic prediction of convective storms with attributes that are likely to impact air traffic flows across the National Airspace System (NAS). In a preliminary proof-of-concept effort, the impacting hazard has been defined as contiguous areas of convection with VIP 3+ exceeding 100 km in length (but allowing for small gaps) and lasting longer than 1 hour. The definition of the convective hazards for aviation can be refined to include echo tops, lightning flash rates, and minimum area constraints, as deemed appropriate.
Storm attributes are obtained using TITAN applied to both model ensemble data and the observations. The thresholds used to identify convection in the model are adjusted as a function of lead time, time of day and season using the Iterative Optimization of Thresholds Algorithm (IOTA) developed by RAL. Use of these optimized thresholds helps to remove model biases and to generate scale-specific probabilistic forecasts of convective hazards that are very reliable. Further, data mining techniques have also proven to be valuable for combining model and observational datasets for very short-range probabilistic forecasts of large-scale storm initiation.
Using an object-based verification technique that keys off the radar-retrieved vertically integrated liquid (VIL), it was found that the High-Resolution Rapid Refresh (HRRR) model is able to reasonably well reproduce the first (i.e., counts) and second moments (size distribution) of observed mesoscale convective systems (MCSs) over the eastern United States, as well as their aspect ratio, orientation, and diurnal variations. In a recent paper, Pinto et al. (2015) discuss regional variations in the skill of HRRR at predicting convective storms with a maximum dimension exceeding 100 km, otherwise known as Mesoscale Convective Systems (MCSs). They also discuss how skill varies from year to year in response to enhancements made to the model physics and data assimilation techniques. An ongoing object-based assessment of the operational HRRR running at NCEP is available here.
Using a data mining and statistical learning method known as a random forest (RF), Ahijevych et al. (2016) have been able to demonstrate remarkable success in detecting initiation of large-scale convective storms two hours ahead. The RF technique employed used decision trees to relate radar, satellite and model analysis datasets to storm initiation. The skill of the RF forecasts was found to increase with the number of trees and the fraction of positive events used in the training set. The skill of the RF was also highly dependent on the types of predictor fields included in the training set and was notably better when a more recent training period was used.
Current research is underway to use scale-dependent ensemble-based storm forecast uncertainty information to improve the heuristic treatment of storm initiation and growth used in the blending algorithm of CoSPA.