SRS4: A stacked residual deep neural network for heave motion continuous prediction of salvage barge

• An innovative stacked residual deep neural network for barge heave prediction. • Convolution calculation and dual-residual connection improve prediction accuracy. • The proposed neural network accurately predicts heave motions for 15 seconds. • 30% accuracy improvement over the widely used long short-term memory network. • Proving the applicability of state-space sequential model for predicting barge heave motion. In shipwreck salvage operations, vessels often experience induced motions due to complex wind and wave interactions, which impact the effectiveness of lifting operations. Heave compensation technology mitigates these effects by decoupling payload motion from vessel motion, which is crucial for ensuring engineering safety. However, system delays due to active hydraulic control and IMU signal filtering process resist the compensation performance. Predicting barge motion to mitigate system delays and achieve leading compensation is necessary for salvage operations. Many neural network prediction methods based on LSTM models have shown remarkable accuracy; however, LSTM models have problems such as long-term memory loss and low training efficiency. This paper proposes a deep-layer neural network based on the structured state-space sequence model for salvage barge heave prediction, named the stacked residual structured state-space sequence predictive neural network. The optimal architectures and hyperparameters of the proposed network are determined through sensitivity analysis. Comparative analyses with LSTM-based models demonstrate the efficacy of the proposed method in multi-run consistency, prediction accuracy, and smooth prediction under varying sea conditions.