Regionalized probabilistic turbulence forecasting using machine learning with in-situ data

This paper describes an approach to using machine learning algorithms together with high spatial- and temporal resolution turbulence observations to provide regionalized, probabilistic turbulence forecasts for NextGen, an enhancement of the Graphical Turbulence Guidance (GTG) product. The high resolution and relatively plentiful supply of in-situ turbulence observation data provided by some commercial aircraft enables development of more sophisticated turbulence forecasting products in two ways: (1) they allow improved identification of a good set of input predictors (e.g., NWP-derived turbulence diagnostics), and (2) they permit development of specialized forecasts by region, both geographic areas and altitudes.

The forecasting performances of three machine learning algorithms - Support Vector Machines, Random Forests, and Logistic Regression - are evaluated in the context of a regionalized turbulence forecasting system. We show how each method can directly produce probabilistic forecasts without resort to a posteri calibration of deterministic forecasts, and compare statistical performance provided by these methods for a wide range of atmospheric conditions.