BiGRU-MHA-KAN based flight training trajectory prediction

基于BiGRU-MHA-KAN的飞行训练轨迹预测方法

¹ Engineering Technology Training Center, Civil Aviation University of China, Tianjin 300300, China
² School of Electronic Information and Automation, Civil Aviation University of China, Tianjin 300300, China

Received: 14 May 2025

Abstract

To improve the accuracy of trajectory prediction in flight training and enhance the reliability of prediction models, a deep hybrid neural network model named BiGRU-MHA-KAN is proposed, in which the bidirectional gated recurrent unit(BiGRU), multi-head attention(MHA) and Kolmogorov-Arnold networks(KAN) are integrated. The model strengthens the temporal feature extraction and nonlinear dynamic modeling through trajectory data preprocessing and reconstruction, combining with the bidirectional modeling, attention mechanisms and KAN networks. Simulated experiments systematically analyze the effect of the different parameter settings and historical data volumes on the model performance. The results demonstrate that, comparing with the other trajectory prediction models, the present method achieves an improvement in prediction accuracy by 4.81%-5.83%, while significantly reducing the mean squared error and root mean squared error, demonstrating the stronger temporal modeling capability and stability in flight training scenarios.

摘要

为提升飞行训练中轨迹预测的精确度, 优化预测模型的可靠性, 提出了一种融合双向门控循环单元(BiGRU)、多头注意力机制(MHA)与Kolmogorov-Arnold网络(KAN)的深度混合神经网络模型BiGRU-MHA-KAN。该模型通过航迹数据预处理与重构, 结合双向建模、注意力机制和KAN网络, 强化时序特征提取与非线性动态建模。仿真实验系统分析了不同参数设置及历史数据量对模型性能的影响。结果表明, 相较于其他轨迹预测模型, 所提方法将预测准确提高了4.81%~5.83%, 且均方误差与均方根误差显著降低, 在飞行训练场景下展现出更强的时序建模能力和稳定性。