AI framework for automated terminal aerodrome forecasting

Abstract

Accurate Terminal Aerodrome Forecasts (TAFs) are essential for aviation safety and operational efficiency worldwide. This study develops an AI framework for automated TAF generation, including data preprocessing, model development, and evaluation. Using GFS and ECMWF datasets from 2020–2023 and real TAF forecasts from Brno International Airport the study explores the effectiveness of ML approaches for wind speed and visibility prediction. Principal Component Analysis (PCA) efficiently reduced dimensionality for wind speed predictors but proved less effective for visibility, highlighting its complex nature. Feature importance analysis identified initial observations and seasonal patterns as dominant predictors, underscoring the influence of data quality. Regression models for wind speed met ICAO standards. While Gradient Boosting (GB) classification outperformed human forecasts in raw accuracy, it suffered from poor probability calibration due to dataset imbalance. A critical evaluation of accuracy metrics – such as log-loss and F1-score – revealed their advantages and limitations, particularly in handling imbalanced datasets and probabilistic forecasting. Beyond its empirical findings, the study provides a theoretical foundation for integrating machine learning (ML) into TAF generation, discussing methodological considerations and the interaction between model performance and forecast interpretability. Future research is recommended to focus on the local models, explore advanced models, and expand the framework to diverse climatic conditions.