Institute of Transportation Studies at UC Berkeley

A go-around is an aborted landing event of an aircraft that is on final approach. Go-arounds are costly and detrimental to safety. Building upon our previous work in go-around detection and analysis of feature contributions, we investigate different learning models and prediction regimes for making sequential predictions of go-around probabilities based on realized trajectory data and environment factors as the aircraft proceeds on its approach. This paper develops and compares the performance of different learning algorithms and prediction strategies for the sequential go-around prediction problem. Applying these methods to a data set consisting of more than 100,000 flight approaches into JFK airport, we find that the Input Output Hidden Markov Model with multi-step prediction strategy, in general, outperforms other models due to its capability of capturing the inherent temporal structure of the entire flight sequence.