Real-time trajectory prediction of drifting unmanned surface vessels using an adaptive extended Kalman filter with dynamic forgetting factor

Unmanned surface vessels (USVs) can reduce self-noise during low-noise marine data acquisition and sea-trial support missions, including underwater radiated-noise measurements, by operating in a drifting state with the propulsion system turned off. However, limited position control during drifting operation makes it difficult to maintain the measurement position and predict the trajectory as drifting motion changes under combined external influences. Predicting the drift trajectory in advance is therefore important for measurement and sea-trial planning. This study proposes an adaptive extended Kalman filter (AEKF) incorporating a curvature-deviation-based adaptive mechanism for real-time drift-trajectory prediction that reflects variations in drifting motion observed in marine environments. While Global Navigation Satellite System (GNSS) measurements are available, the proposed method uses curvature deviation calculated from the estimated states as an adaptive indicator to dynamically adjust the Sage–Husa forgetting factor and the relative weighting in the process noise covariance update. When GNSS measurements are unavailable, the future drift trajectory is predicted using the system model starting from the last state estimate and retaining the last updated process noise covariance. Performance was evaluated using numerical simulations and real-sea drift data. In the real-sea experiment, GNSS position data were used for real-time trajectory prediction, while recorded data were used for quantitative evaluation after the experiment. Comparative simulations showed that the proposed AEKF produced lower long-term prediction errors than comparison filters under changes in drift direction and curvature. In the real-sea experiment, it achieved the lowest M A E 2 D among the three methods in Cases 1 and 2. Sliding-window analysis over the entire experimental period showed a mean prediction error comparable to that of the augmented-model EKF and the lowest error standard deviation among the three methods.