An improved transformer-based model for long-term 4D trajectory prediction in civil aviation

Four-dimensional trajectory prediction is a crucial component of air traffic management, and its accuracy is closely related to the efficiency and safety of air transportation. Although long short-term memory (LSTM) or its variants have been widely used in recent studies, they may produce unacceptable results in long-term prediction due to the iterative output that accumulates error. To address this issue, a transformer-based long-term trajectory prediction model is proposed here, which utilizes the self-attention mechanism to extract time series features from historical trajectory data. For long-term prediction scenarios, we a trajectory stabilization module is introduced to ensure the stationarity of the time series for better predictability. Additionally, the transformer output strategy is improved to generate the prediction sequence by a single step instead of serial dynamic decoding, thus effectively enhancing the precision and inference speed. The proposed model is validated using real data obtained from China's Southwest Air Traffic Management Bureau. The experimental results demonstrate that this model outperforms the benchmark model. Further ablation experiments and visualizations are performed to analyze the impact of trajectory stabilization and one-step inference strategy.