Data-Driven Prediction of Runway Incursions with Uncertainty Quantification

Abstract

In 215 only, more than 1,5 runway incursions (RIs) occurred at US airports, which could result in serious runway collisions. Nonlinear interactions among many factors and complex data structures pose challenges to RI prevention, and reportedly, the annual RI occurrence is gradually increasing. This study seeks to offer a data-driven solution of advanced statistical learning and prediction by leveraging the generalized additive model (GAM). The GAM holds a powerful flexibility with little restriction to many variables over a broad range of modeling distributions. This study proposes a method to systematically obtain, parse, and transform various factors from diverse databases to give rise to interpretable datasets. It also presents high-performance computational procedures to automatically select out salient factors to achieve the best GAM with a strong predictive power. Practical applications to RI of US airports show promising performance. A combination of GAM and bootstrapping method to build confidence intervals is expounded upon as a means to quantify underlying uncertainties.