Trajectory prediction for a multi-feature underwater gliding unmanned system under complex constraints and high dynamics with limited samples

As a crucial force in marine scientific research, the underwater gliding unmanned system must perform tasks with high accuracy and efficiency. Due to the influence of rocket fuel, underwater gliders exhibit rapid response times, making it challenging to control their future attitude based solely on their current state. Therefore, accurate trajectory prediction is essential for ensuring effective and reliable control of the system. However, the complex underwater environment presents significant challenges for traditional dynamic model-based methods in accurately predicting the future trajectory of underwater gliders. Moreover, conventional prediction models struggle to maintain accuracy in the final stages of the trajectory, particularly in high-dynamic scenarios where fuel depletion leads to a rapid decline in speed. To address these issues, this paper proposes a Maformer(Marine-Informer) model for trajectory prediction in underwater gliding unmanned systems. Building upon the traditional Informer model, Maformer integrates velocity feature encoding and replaces the original convolutional layer with dilated convolution. Experimental results demonstrate that, compared to the traditional Informer model, Maformer achieves an average reduction of 52.82% in prediction error. The findings of this study offer a promising solution for trajectory prediction of multi-feature unmanned systems operating in high-speed, dynamically changing environments.