Designing Recurrent and Graph Neural Networks to Predict Airport and Air Traffic Network Delays

In this paper, we propose open machine learning models that can provide airport delay predictions in a network with an error of around or less than five minutes. Due to the complexity of different components of air traffic networks, traditional flight performance model-based predictions fall short when dealing with numerous flights and often are not able to deal with delays that propagate among airports in a network. In this study, we employ three different machine learning models to predict delays at three different scopes: individual flights, airports, and the network of airports. Consequently, we tested three approaches with different levels of complexity, including statistical regression models, recurrent neural networks, and spatial-temporal graph attention neural networks. We conduct experiments for all three types of models using the Eurocontrol research data archive. After training and testing with two years of data covering the top 50 European airports, our models produce prediction errors of around or less than 5 minutes with look-ahead time up to 3 hours. These metrics have shown a significant advancement compared to existing prediction models. We also openly share this model to support open science in aviation.