Multi-horizon flight trajectory prediction enabled by time-frequency wavelet transform

Flight trajectory prediction is a fundamental task in air traffic control. While our previous WTFTP framework leveraged wavelet-based time-frequency analysis for trajectory prediction, its iterative single-horizon paradigm suffers from error accumulation in long-horizon forecasts. Here, the WTFTP+ framework is proposed to enhance multi-horizon prediction performance while further exploring the potential of time-frequency analysis in flight trajectory prediction tasks. An encoder-decoder neural architecture is designed to generate wavelet components of predicted trajectories, where a direct multi-horizon prediction paradigm is employed to mitigate the cumulative errors of the WTFTP. Additionally, a time-frequency bridging mechanism is proposed to explore intrinsic correlations among multi-scale flight patterns to enhance the learning ability of wavelet components. Experiments on real-world datasets demonstrate that WTFTP+ not only retains the superior single-horizon prediction performance of WTFTP but also significantly improves multi-horizon prediction accuracy, achieving over 40% mean deviation error reduction at the 5-minute horizon compared with WTFTP.