Investigating the Viability of Machine Learning for the Prediction of Icing Occurrences at Airports

Abstract

Ensuring the functionality of airport operations amidst changing weather conditions is crucial for maintaining operational efficiency and ensuring safety. This thesis investigates using Machine Learning (ML) techniques to predict icing weather events at airports, aiming to enhance aviation safety through improved forecasting. Leveraging input data from Numerical Weather Prediction (NWP) models and real-time observations extracted from Meteorological Aerodrome Reports (METAR), probabilistic classifiers are developed and further assessed through their efficacy in providing reliable predictions for freezing weather occurrences. Due to the rarity of these events, the resulting dataset is inherently imbalanced, necessitating heavy downsampling to facilitate efficient model training. Despite the small dataset size, the developed models demonstrate promising capabilities, exhibiting notable improvements in reliability and accuracy, particularly within the temporal models. Notably, even with minimal training data, the models accurately predict freezing weather occurrences up to five timesteps ahead, each representing one hour.

Hence, this study emphasizes the importance of collaboration between ML experts and domain specialists in aviation meteorology to gain deeper insights and refine the models. The results serve as a solid foundation for reflection and offer valuable suggestions for future research directions to enhance ML models' predictive capabilities in this domain.