Application of Machine Learning to a Low-level Aviation Turbulence Forecasting System

The physical mechanisms of turbulence in the lower troposphere are much more complex than and significantly different from those in the mid- and upper troposphere, which leads to relatively poor performance of aviation turbulence forecasting in the lower level (below 10,000ft) compared with that in the upper levels. Recent progress of machine-learning techniques is encouraging to apply them to aviation turbulence forecasting, especially, in the lower troposphere. In this study, we developed a Random Forest (RF) model for predicting low-level turbulence, and its performance is evaluated by comparing observations as well as results obtained from the global Korean aviation Turbulence Guidance (G-KTG) system, the current operational model of aviation turbulence at Korea Meteorological Administration (KMA). The RF model was trained by 334,120 observation-prediction pairs, utilizing 78 turbulence diagnostics derived from the Global Data Assimilation and Prediction System (GDAPS) of KMA for one year (2022.1-2022.12). The observational data used for the training and validation are the in-situ flight eddy dissipation rate (EDR) provided from the International Air Transport Association (IATA). The RF model exhibited exceptional performance across all evaluated metrics, notably achieving a significant correlation coefficient and an area under the ROC curve (AUC) exceeding 0.9. The performance of RF model is much higher than that of G-KTG, at least in 2022. The model's strengths were evident in its ability to effectively predict moderate-to-severe (MOG) turbulence and reduce overestimation of null intensity turbulence, which have been a long challenge for turbulence forecasting systems including G-KTG system. For optimization, the number of input variables was reduced from 78 to 13, sustaining its original performance, showcasing its practical effectiveness. The RF models’ reliability was further validated by their consistent turbulence prediction patterns, which aligned with known physical phenomena such as orography and convective activity. Additionally, we examined the RF model's time-dependent performance, showing the lowest accuracy during the June-August period. Details in performance of RF model and its comparison with G-KTG will be presented in the conference.