Abstract: To tackle the challenges that complex and dynamic environments impose on motion prediction of marine vehicles， an intelligent prediction system integrating cascaded filter with error-triggered support vector regression （ETSVR） is proposed in this paper. Firstly， the system employs a moving average filter for preprocessing raw data， effectively eliminating outliers and high-frequency noise to establish a high-quality dataset for the following prediction. Subsequently， a second-order extended Kalman filter is incorporated for precise state estimation， further enhancing the data stability and reliability. Finally， the ETSVR algorithm works on the processed high-quality dataset to build an accurate motion prediction model for marine vehicles， with an error-triggered mechanism designed to improve the system's real-time performance and computational efficiency. Experimental results based on lake tests demonstrate that the proposed intelligent motion prediction system significantly outperforms traditional linear regression （LR） across various error metrics. For instance， in lateral velocity prediction， the mean squared error is reduced by approximately 53.2% compared to LR， while in yaw rate prediction， the maximum error decreases by around 58.2%. These results indicate that the proposed cascaded filter and ETSVR-based prediction system substantially improves the accuracy of motion prediction in complex unknown environments， highlighting its promising potential and significance.

Key words: marine vehicle, motion prediction, moving average filtering, second-order extended Kalman filter, error-triggered support vector regression