Aircraft taxi time prediction: Feature importance and their implications. (March 2021)

• Record Type: Journal Article
• Authors: Wang, Xinwei; Brownlee, Alexander E.I.; Woodward, John R.; Weiszer, Michal; Mahfouf, Mahdi; Chen, Jun
• Is Part Of: Transportation research. Volume 124(2021)
• Publication Date: 2021-03
• Subjects: Air traffic management -- Feature importance -- Machine learning -- Prediction -- Taxi time; Transportation -- Periodicals; Transportation -- Technological innovations -- Periodicals; 388.011
• DOI: 10.1016/j.trc.2020.102892
• Languages: English
• ISSNs: 0968-090X

Abstract

Highlights: Multiple machine learning approaches are applied to aircraft taxi time prediction. Models are constructed for Hong Kong, Manchester, and Zurich airports. Features used include route and aircraft information, traffic levels, and weather. Features required for a high accuracy model are identified. These are: depart/arrival; distance; departure traffic; recent traffic average speed. Abstract: Taxiing remains a major bottleneck at many airports. Recently, several approaches to allocating efficient routes for taxiing aircraft have been proposed. The routing algorithms underpinning these approaches rely on accurate prediction of the time taken to traverse each segment of the taxiways. Many features impact on taxi time, including the route taken, aircraft category, operational mode of the airport, traffic congestion information, and local weather conditions. Working with real-world data for several international airports, we compare multiple prediction models and investigate the impact of these features, drawing conclusions on the most important features for accurately modelling taxi times. We show that high accuracy can be achieved with a small subset of the features consisting of those generally important across all airports (departure/arrival, distance, total turns, average speed and numbers of recent aircraft), and a small number of features specific to particular target airports. Moving from all features to this small subset results in less than a 1 percentage-point drop in movements correctly predicted within 1, 3 and 5 min.