Improving the spatial-temporal generalization of flight block time prediction: A development of stacking models

Flight block time (BT) is the time between gate departure and gate arrival. BT is difficult to predict because it depends on many latent variables such as airport layout, taxiing procedures, airspace status, and weather events. There is some increasing interest among aviation practitioners in machine learning techniques to help predict complex and nonlinear relationships. Among them, several machine learning methods have proved their significant advantages. However, recent research showed that these methods may not be able to generalize at different origin-destination (OD) pairs and periods (i.e. spatial-temporal scales), which challenges their further application in practice. The main goal of this paper is to improve the spatial-temporal generalization of the BT prediction models using the stacking method. We select four busy OD pairs in the National Airspace System as the cases and model BT prediction on individual OD pairs using data from two periods. Following seven typical machine learning methods, we first investigate their performance in different OD pairs and different periods. The results demonstrate that they also fail to generalize at spatial-temporal scales in the BT prediction. Then, we propose a stacking model to predict flight BT. Compared with the previous methods, our stacking method is able to achieve promising generalization at various spatial-temporal instances. In addition, we analyze the feature importance on each OD pair and find that they may also vary with periods. In order to improve the computational efficiency, we also develop lightweight BT prediction models that are trained with fewer but more important features. Although they show promising prospects in the reduction of computational costs, analysts should be cautious since the feature importance may vary with periods and the neglected features with lower importance may play a key role in the real world.