Previsão de Turbulência de Céu Claro utilizando modelo hibrido sobre o sul do Brasil

Abstract

In this study, the application of machine learning (ML) algorithms is evaluated using simulations based on outputs from the Global Forecast System (GFS) and the Weather Research and Forecasting (WRF) models, with the objective of developing an optimized hybrid model for forecasting clear-air turbulence (CAT) over southern Brazil. Fourteen atmospheric predictors were used, extracted between the 500–200 hPa isobaric levels, initially derived from GFS forecast data and subsequently from WRF forecast data, with the latter driven by GFS analysis fields. These attributes were simulated at different spatial resolutions in order to train and validate 15 distinct machine learning models. The model development and comparison methodology involved stratified sampling, feature engineering, hyperparameter optimization, and robust validation of statistical metrics, including Probability of Detection of CAT events (PODy), Probability of Detection of non-events (PODn), True Skill Statistic (TSS), F-measure, Area Under the Receiver Operating Characteristic Curve (AUC), and the Relative Percentage Difference of AUC between WRF and GFS ((AUC_WRF–AUC_GFS)/AUC_GFS) × 100. Among the evaluated algorithms, the Bagging classifier showed the best overall performance when considering AUC values obtained from both GFS and WRF forecast datasets. Although Extra Trees model achieved a slightly higher AUC than Bagging when using WRF data, vi the AUC difference between the two models in the GFS simulations was marginally larger than that observed for WRF. Thus, Bagging stood out for its more consistent performance across the combined dataset, while Extra Trees exhibited marginally superior performance when applied specifically to WRF forecasts. The KNN classifier showed the greatest gain in Relative Percentage Difference of AUC in the comparison between GFS and WRF-based models. The predictor exclusion method confirmed that potential temperature (Θ) and vertical wind shear / wind speed difference indicators (VWS, DWS) were the attributes that best represented effective CAT detection, while turbulent kinetic energy (TKE) and turbulence indices did not achieve satisfactory performance. The results demonstrate the potential of regionally training ML techniques with numerical weather prediction models—emphasizing the use of higher spatial resolution physical models—to substantially improve CAT detection compared to traditional diagnostic methods. In the operational context, the Bagging and Extra Trees models showed more satisfactory performance for issuing more accurate clear-air turbulence (CAT) alerts in the upper air corridor between the cities of São Paulo and Porto Alegre, within the Curitiba Flight Information Region (FIR-CW). Consequently, reliance on the subjectivity of aircraft turbulence reports is reduced, enhancing flight safety and lowering operational costs. Future research should focus on using datasets with higher spatial and temporal resolution, exploring additional turbulence diagnostics, and selecting different study areas and seasonal periods to further improve CAT predictive capability for operational applications