4D Trajectory Prediction for Inbound Flights

Weizhen TangJie Dai^*

• Civil Aviation Flight University of China, Guanghan, China

Addressing the key challenges in current 4D trajectory prediction, including cumulative prediction errors, insufficient modeling capabilities for complex spatiotemporal features, and limitations in computational efficiency and generalization ability, this paper aims to propose a high-precision, robust prediction solution. To this end, this study innovatively constructs a hybrid model SVMD-DBO-RCBAM that deeply integrates sequential variational modal decomposition (SVMD), the dung beetle optimization algorithm (DBO), and the ResNet-CBAM (RCBAM) network. Its core innovations are as follows: 1) Frequency-domain feature decoupling: SVMD adaptively decomposes the original track signal into multi-scale sub-modes at the frequency domain level, effectively separating noise interference and extracting essential features; 2) Dynamic parameter optimization: The DBO algorithm is introduced, leveraging its global-local collaborative search mechanism to efficiently optimize the key hyperparameters of the RCBAM network (learning rate, number of filters, residual block configuration, etc.), significantly improving model convergence speed and generalization ability; 3) Enhanced spatio-temporal feature focusing: The RCBAM network is designed to integrate ResNet's deep residual learning capabilities to overcome gradient vanishing/explosion issues, and embeds CBAM's channel-spatial dual attention mechanism to dynamically assign higher weights to key feature dimensions (such as longitude, latitude, altitude) and key time windows, precisely capturing the spatio-temporal evolution patterns of flight paths. Experimental validation using real ADS-B flight path data from Zhuhai Jinwan Airport demonstrates that the proposed model achieves a low longitude MAE of 0.0377 in single-step prediction, representing a significant 38.5% reduction compared to the baseline RCBAM model; in multi-step prediction, its longitude R² reaches 0.9844, with a 72.9% reduction in cumulative error rate, and the interquartile range (IQR) of prediction errors is reduced to less than 10% of traditional models, demonstrating exceptional prediction accuracy, stability, and resistance to cumulative errors. This study provides a strong theoretical basis and technical solutions for the construction of an intelligent air traffic control system based on dynamic and refined airspace management, real-time flight conflict warning, and trajectory-based operations (TBO).