Hybrid clustering/HMM constrained-based learning for Aircraft Trajectory Prediction

Abstract:

Aircraft trajectory prediction (TP) is a challenging and inherently data-driven time-series modeling problem. Adding annotation parameters further increases the complexity of the search space, especially when ‘blind’ optimization algorithms are employed. In this paper, flight plans, localized weather and aircraft properties are introduced as trajectory annotations (or semantics), which enable modeling in a space higher than the typical 4-D spatio-temporal domain. A two-phase hybrid approach is employed for the core TP task: (a) clustering using properly designed semantic-aware similarity functions as distance metrics; and (b) a hidden Markov model (HMM) for each cluster, using non-uniform graph-based spatial grid and exploiting flight plans as constraints for a parametric probabilistic model for the emissions. The proposed method is applied in real radar tracks and weather data for a one-month dataset of flights in Spanish airspace. Using parametric Gaussians as the base for the emissions model and confidence interval estimations for the associated errors, the proposed method exhibits exceptionally low HMM complexity and per-waypoint prediction accuracy of a few hundred meters compared with submitted flight plans.