A Hybrid State-Driven Trajectory Prediction Algorithm for Unmanned Aerial Vehicles

Unmanned aerial vehicles (UAVs), as a mainstream implementation of mobile agents, are often plagued by restricted access to the state information of peer formation members and compromised system stability during formation coordination missions, primarily due to intermittent communication link outages and inherent measurement noise. Meanwhile, prevailing deep learning-based trajectory prediction methods incur prohibitive computational burdens, making them ill-suited for deployment on UAVs with highly constrained onboard computing resources. To tackle the aforementioned challenges, this paper adopts a data correction-driven approach and puts forward a hybrid state-driven trajectory prediction algorithm tailored for UAV mobile agents, termed Linear Basis Model with Residual Long Short-Term Memory Correction (LBM-RLSTM), which fuses a linear kinematic motion model with the sequence prediction capability of deep neural networks. Through kinematic characteristic analysis of short-term UAV flight trajectories, we decouple the trajectory prediction task into two interdependent modules, which correspond to the fitting of linear dominant motion terms and the compensation of nonlinear residual errors respectively. Specifically, the proposed linear basis prediction model is adopted to generate initial trajectory prediction results with high computational efficiency and low complexity, ensuring reliable baseline prediction performance. Subsequently, a lightweight residual LSTM network is leveraged to perform compensation and correction on residual sequences, which incorporate unmodeled nonlinear motion components and noise-induced disturbances. This two-stage hybrid architecture eliminates the need for deploying large-scale neural networks required by conventional end-to-end full-trajectory fitting methods, and achieves a balance between prediction accuracy and the practical deployment feasibility on resource-constrained UAV onboard platforms. Both theoretical analysis and extensive simulation results validate that the proposed LBM-RLSTM method achieves superior prediction accuracy and stronger robustness under complex scenarios with mixed Gaussian and impulse noise interference as well as intermittent communication interruptions. Meanwhile, it significantly reduces computational overhead compared with mainstream deep learning-based benchmarks, and is more suitable for satisfying the strict real-time requirements of onboard UAV trajectory prediction in practical formation flight missions.