Physics-Aware Residual Variational Autoencoder for 3D UAV Trajectory Prediction

Trajectory prediction is crucial for intelligent systems to enable environmental perception, path planning and interaction coordination, yet it faces challenges by multiple influencing factors, complex constraints and heterogeneous motion patterns. Pure data-driven models generalize well but may yield physically inconsistent predictions, while mechanism-based models ensure interpretability and physical plausibility but struggle in complex interactive scenes. To address these challenges, we propose a Physics-aware Residual Variational Autoencoder (i.e., PhysVAE) that couples mechanism models within data-driven learning for Unmanned Aerial Vehicles (UAVs) trajectory prediction. PhysVAE comprises three modules: multidimensional feature extraction modeling motion states, sensor data and environmental features; a kinematic mechanism module providing physically consistent preliminary predictions; a residual prediction network based on VAE that compensates for mechanism model biases and enhances cross-scenario generalization capability. Extensive experiments on the ADS-B and ATS datasets indicate that PhysVAE reduces MSE by 19.5% compared to the state-of-the-art RLSTM, while consistently delivering high prediction accuracy and robustness in complex wind fields and multi-disturbance conditions, validating its effectiveness for 3D UAV trajectory prediction.