Optimizing Roll Motion Forecasting for Unmanned Surface Vessels Using PSO-Enhanced LSTM Networks

Unmanned surface vessels (USVs) navigating at sea experience severe roll motion due to waves, which can compromise their operational safety and efficiency. Accurate short-term prediction of roll motion is critical for decision-making in USV operations. The Long Short-Term Memory (LSTM) neural network model, with its capability to handle nonlinear problems, offers a potential solution for predicting the nonlinear roll motion of USVs. For the task of USV roll motion forecasting, determining the appropriate dimensionality of the input vector and the number of hidden layer neurons are points that need to be considered, especially when dealing with varying prediction horizons. Typically, these parameters are selected by exhaustively searching within a predefined range based on empirical experience. This approach, however, is computationally expensive and lacks adaptability in finding the optimal input vector dimensionality and number of hidden neurons. To address this issue, this paper introduces the Particle Swarm Optimization (PSO) algorithm to optimize both the input vector dimensionality and the number of hidden layer neurons. The results show that using PSO to optimize the LSTM model parameters for predicting USV roll motion is effective, providing a certain degree of model optimization and improving prediction performance. Moreover, the PSO-LSTM prediction model demonstrates generalization ability, making it not only suitable for roll motion prediction at specific forecast steps but also adaptable to different prediction horizons.