A multi-USV trajectory and intent prediction model based on stacked temporal graph convolutional network and spatial-temporal Transformer

In modern naval warfare, trajectory prediction and intent recognition of adversarial unmanned surface vehicle (USV) swarms can provide critical support for combat command systems, thereby enabling strategic advantages in adversarial engagements.

This paper proposes a predictive model combining a stacked temporal graph convolutional network (STGCN) and a spatio-temporal Transformer (abbreviated as GST-Transformer) to forecast future trajectories and recognize the combat intent of adversarial USV swarms. In the GST-Transformer framework, the STGCN is designed to capture temporally correlated interaction features from historical USV trajectory data. The spatio-temporal Transformer employs a dual-channel encoder to extract and integrate spatio-temporal features from adversarial trajectories simultaneously. Subsequently, a generative trajectory encoder synthesizes representations for trajectory prediction by integrating interaction features and spatio-temporal characteristics. Finally, a multi-head decoder generates both the predicted trajectories and intent predictions for the adversarial USV swarms.

Experiments conducted on simulated adversarial USV data show that the proposed model achieves superior prediction accuracy in trajectory forecasting tasks. Compared to mainstream methods such as Informer and GRU, it reduces the average displacement error (ADE) by 25.72% and final displacement error (FDE) by 16.27%. Additionally, the model demonstrates high accuracy and robust performance in combat intent recognition tasks.

The proposed GST-Transformer model exhibits efficient trajectory prediction and intent recognition for USV swarms, offering novel technological support for modern naval warfare command systems. This advancement enhances situational awareness and improves decision-making precision in dynamic combat environments.