Operational Use of Machine Learning and Numerical Weather Prediction to Predict Aircraft Turbulence and Icing

The author previously showed how machine learning could be used with the Global Forecast System (GFS) to train data from pilot reports (PIREPs) to predict aircraft icing, and on AMDAR Eddy Dissipation Rate (EDR) values to predict turbulence. Several machine learning techniques were examined, including "black box" regressors. Last year, the "black box" appeared to be the better option, but the necessity of a less efficient, higher level language like Python to process the numerical weather model data reduced this method's operational advantage. Thus, the primary focus shifted to a "white box" approach, in order to allow the resulting linear regression coefficients and decision tree hierarchies to be hard-coded into a more efficient number-crunching c-based programming language. While the aircraft icing forecast used a fairly simple set of variables, the EDR forecast required more input variables, careful feature selection, and testing, especially as the goal was to split the results into three different types of turbulence: Clear Air Turbulence (CAT), Convective Induced Turbulence (CIT), and Mountain Wave Turbulence (MWT). The machine learning model output was tested and verified against the observed EDR data, and implemented operationally to produce a forecast product. This presentation details the advantages and drawbacks of the different machine learning techniques and the operational implementation of this aircraft icing and EDR forecast values.