A Bi-LSTM and AutoEncoder Based Framework for Multi-step Flight Trajectory Prediction

Abstract

Trajectory prediction (TP) is a key component in the route planning for civil aircraft. Most existing methods obtain multi-step TP via iterating the one-step TP model, which generally generates large cumulative error due to deviate from the original evolutionary pattern. To improve the situation, this paper proposes a multi-step TP framework with three modules: the Bi-directional Long Short-Term Memory Network (Bi-LSTM) based multi-step TP module, AutoEncoder based multi-step TP module, and voting fusion module. In the Bi-LSTM based multi-step TP method, to avoid the forgetting of evolutionary characteristics, the Bi-LSTM is designed to directly extract the mapping relationship between input of historical trajectory fragments and output of multi-step labels via data-driven method. In the AutoEncoder based multi-step TP module, the Bi-LSTM is deigned to learn mapping relationship between the input and core evolutionary features from output labels extracted via the encoder, and then the decoder is adopted to reconstruct predictions by outputs from Bi-LSTM. Third, the voting method was used to fuse the per-dimension predictions from the above two modules and further to refine multi-step predictions. The proposed multi-step TP framework is applied to real flight trajectory prediction of civil aircraft and outperforms multiple deep learning methods in the terms of RMSE and MAE.