Techniques for providing probabilistic forecasts of turbulence for NextGen

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Wednesday, 20 January 2010: 5:00 PM
B204 (GWCC)
Matthew J. Pocernich, NCAR, Boulder, CO; and R. D. Sharman and J. K. Williams

Probabilistic weather hazard forecasts are an integral part of the NextGen system. However, development of probabilistic forecasts of turbulence for the aviation community is made difficult by the relative scarcity of “moderate” and “severe” events coupled with the lack of unbiased random measurements of turbulence. Ultimately though, probabilistic turbulence forecasts are necessary given the random nature of turbulence severity within turbulence patches or zones and the need to give decision makers probabilistic information.

Here, two approaches to providing turbulence forecasts are outlined and preliminary results presented. The first approach uses a pseudo-ensemble of forecasts by developed by computing a suite of turbulence diagnostics, such as those currently computed within the Graphical Turbulence Guidance system (GTG) and treating each turbulence diagnostic as an ensemble member. Using a hold-out cross validation procedure, optimal thresholds are calculated for each index creating a yes/no forecast for turbulence. To create a probability forecast, the proportion of members indicating turbulence is calculated. The second approach uses regression trees to determine the optimal order of variables and thresholds to correctly predict turbulence. A Random Forest is a variation of this technique which calculates a series of random, non-optimal regression trees. The consensus of these predictions is then used to generate a forecast.