An Intention-Aware Aircraft Trajectory Prediction Approach with Dynamic Caching Long- and Short-Term Flight State Representations

Abstract

Aircraft trajectory prediction aims to forecast the future states of dynamic agents within a three-dimensional airspace, given their current and historical flight states. Intuitively, the historical flight states of these agents reveal the intentions of their pilots, which directly influence their future trajectory. However, existing approaches struggle to handle the long-term dependencies present in flight states and overlook the fundamental connection between an agent's behavioral intention and its trajectory. To address these challenges, this paper proposes a hybrid cached attention encoder-decoder (CAED) framework for aircraft trajectory prediction, with behavioral intention recognition as an auxiliary task. Specifically, CAED incorporates a dynamic cache within the encoder-decoder structure to capture the long-term dependencies between the targets' historical and current flight state representations. During the training process, an additive selection mechanism is proposed to select relevant history states stored in each memory slot that contribute most to the prediction of future trajectories. Finally, we design a scoring function and implement multi-task learning to jointly train our model. The extensive experiments on two aircraft trajectory prediction datasets clearly validate the superiority of our proposed CAED compared with various existing baselines.