A Short-term Aircraft Trajectory Prediction Framework Using Conditional Generative Adversarial Network

Short-term aircraft trajectory prediction (TP) plays an important role in current air traffic control systems. However, existing works usually perform the multi-horizon TP task in an iterated manner which easily suffers from error accumulation problems. In this work, a novel short-term aircraft TP framework, called TPGAN, is proposed which predicts the multi-horizon trajectory in a single step using the conditional generative adversarial network (CGAN). Compared with the conventional approaches, the TP task is formulated to the probability distribution estimate problems by CGAN architecture. In this framework, the generator is employed to output the predictions while the discriminative features between ground truth and predictions are learned by the discriminator. The generative adversarial training strategy is applied to optimize the proposed framework. Moreover, to validate the generality and effectiveness of the proposed framework, three neural network architectures are designed to develop the proposed framework, including Conv1D-TPGAN, Conv2D-TPGAN, and LSTM-TPGAN. In addition, a dataset collected from real-world ATC systems is used to construct the experiments. The experimental results demonstrated that the proposed framework achieves significantly performance improvements than baseline.