Abstract
As a core piece of equipment for marine monitoring, search and rescue missions, and other applications, the motion state prediction accuracy of Unmanned Surface Vehicles (USVs) directly determines mission reliability and safety. However, existing methods fail to fully consider the motion characteristic differences in various vessel sizes and variable-speed navigation under complex sea conditions, and struggle to capture the spatiotemporal dynamic features of state variations. This paper proposes a hybrid prediction algorithm based on Random Forest-Long Short-Term Memory (RF-LSTM), which utilizes Random Forest for key feature selection while employing LSTM to excavate temporal correlations. An intelligent routing mechanism based on the dominant frequency energy ratio (Pd) is introduced to achieve adaptive prediction mode switching, enabling comprehensive characterization of state variations. Under the 20 kn high-speed condition of a 7.5 m USV, the proposed algorithm achieves a Circular RMSE for heading prediction that is 1.9 times lower than the Extended Kalman Filter (EKF) and 1.2 times lower than a standalone LSTM, with pitch and roll prediction RMSE reduced to 0.36° and 0.85°, respectively. On a 14.5 m-long USV at 23 kn, it maintains a heading prediction accuracy of 0.10°, verifying favorable scale generalization capability. Furthermore, the algorithm demonstrates strong robustness against Gaussian white noise and synthetic ocean noise. Experimental results indicate that RF-LSTM significantly outperforms traditional methods, effectively breaking through the application limitations of fixed-architecture models, substantially enhancing USV autonomy and adaptability in complex marine environments, and providing robust guarantees for mission reliability and safety.