LGFG-Net: A Short-Term Trajectory Forecasting Approach Integrating Multi-Feature Extraction and GAN-GCN Framework

The existing deep learning methods struggle to effectively utilize the intrinsic correlation information in ADS-B track data, so the track prediction accuracy is limited. In this paper, a new method called LGFG-Net is proposed: light Gradient Boosting Machine (lightGBM) was used for preprocessing to extract the multi-feature information of the track, and the Generative Adversarial Network (GAN) was used as the backbone prediction model, then the random forest model was used to mitigate the overfitting phenomenon of GAN, and finally the Graph Convolutional Neural Network (GCN) was used to optimize the track prediction results of GAN. This model consists of four modules: lightGBM feature extraction, GAN, random forest, and correlation fusion using GCN. Firstly, the lightGBM algorithm is used for extracting the characteristics of the track, and the model will focus on the severe jitter points in the track data. Secondly, in the GAN module, the generator uses the transformer network, and the discriminator introduces the improved deep separable convolutional network for track data. Then, the random forest model was used to reduce GAN overfitting. Finally, GCN was used to fix the problem that the correlation of data would change after the random forest module. The experimental results show that compared with the track prediction results of the original GAN, the proposed method reduces the MAE of predicted longitude, latitude, and height by 91.01, 82.25, and 73.63%, respectively. It has smaller errors and higher prediction accuracy in all aspects.