Incremental learning LSTM for short-term roll forecasting in USVs under varying wave encounter conditions

Unmanned surface vehicles (USVs) operating in open-sea environments experience significant roll motion due to wave-induced excitation. The dynamic roll response depends on various factors, including vessel speed, wave encounter angle, and sea state. This paper introduces an online learning method based on a long-short-term memory (LSTM) network that uses incremental learning for online parameter updates, enabling time-series forecasting of USV roll dynamics under varying conditions. The proposed method leverages the temporal modeling capabilities of the LSTM to capture the time dependence induced by the hydrodynamic memory effects of roll motion. Through incremental learning, the model continuously updates its network weights using new data, avoiding full retraining and enhancing computational efficiency. This study compares two online learning modes, incremental updating and retraining, for forecasting short-term roll motion under diverse operating conditions. Validation is conducted using both computational fluid dynamics (CFD) simulation data and sea trial measurements collected from a USV in the South China Sea. The evaluations focus on short-horizon roll prediction across varying sea states, wave encounter angles and encounter frequencies. Unlike offline models trained under fixed conditions, the proposed online learning framework adapts to changes in the statistical distribution. Notably, the incremental learning model achieves comparable accuracy to retraining while offering substantially higher update efficiency, especially in real sea conditions.