Online Learning-Based Model Predictive Control for Autonomous Surface Vehicles by Improved Long Short-Term Memory Network

Traditional model predictive control (MPC) methods rely heavily on accurate system models. However, obtaining precise models for real-world systems is often challenging, which may lead to performance degradation or even closed-loop instability. To address this issue, this article proposes an online learning-based MPC framework for trajectory tracking control of autonomous surface vehicles (ASVs) by exploiting the representation capability of deep learning. First, a selective updated strategy-based long short-term memory (SUS-LSTM) network is developed to learn the residual dynamics between the nominal and actual ASV models using a simplified gating mechanism and feature selection. Then, an independent extended Kalman filter (IEKF)-based online learning algorithm is employed to dynamically update the network parameters, thereby improving adaptability and prediction accuracy under dynamic operating conditions. By integrating the learned residual model into the MPC framework, an online learning-based MPC scheme is constructed that effectively addresses state and input constraints while maintaining satisfactory trajectory tracking performance. Moreover, rigorous theoretical analyses are provided, including the convergence of SUS-LSTM prediction error and the closed-loop stability of the proposed MPC scheme. Simulation studies and hardware experiments validate the effectiveness and advantages of the proposed method, achieving an MSE reduction of approximately 35% compared with representative first-order learning-based MPC schemes.