Predicting arrival delays in the terminal area five hours in advance with machine learning

This paper presents a study aiming at predicting the arrival delays occurring in the terminal area up to five hours in advance. The motivation for the participating airlines is to better take into account the impact of weather at destination on fuel planning. Due to the uncertainty at these time horizons, we decided to consider delay intervals (low <5 minutes, moderate 5-10 minutes, high >10 minutes) over 30 minutes periods. We selected four European airports occasionally or frequently subject to high arrival delays (London Heathrow, Dublin, Lisbon and Zurich). The problem was framed as a classification problem and different machine learning models were developed using arrival delay, traffic demand and weather historical data from 2013 to 2019. A random forest model beats the baseline although still below a perfect prediction. The performance indicator (macro F1 score ranging from 0 to 1) increases from 0.3 (baseline) to around 0.5. In terms of prediction error, compared to the baseline, the model has slightly lower performance for the low delays, similar for the moderate delays and better for high ones. Finally, a test case using airlines data illustrated the potential benefits. Indisputably, there should be a “performance barrier” due to the intrinsic uncertainty, essentially in terms of take-off times. Still, the future work should aim at determining whether the performance may be increased, by analyzing the prediction errors and the delay class overlaps.