Research on springback prediction of ship hull plates based on the improved WOA-optimized CNN-GRU-attention model

• A springback dataset was constructed from Abaqus simulations of the full stamping–unloading process under multiple plate thicknesses and loading conditions. • A hybrid CNN–GRU–Attention network was developed to extract local sequence features and learn temporal dependencies for springback prediction. • An improved whale optimization algorithm (IWOA) integrating Tent chaotic mapping, Cauchy perturbation, and greedy selection was proposed and used for hyperparameter tuning to enhance search stability and avoid premature convergence. • Comparative studies against GRU/LSTM and multiple hybrid/WOA-optimized baselines demonstrate consistently better accuracy (lower MAE/MSE and higher R²). • Physical equipment-level verification was conducted on an SKWB-2500 3D CNC plate bending machine, using 0.1 mm-accuracy 3D laser scanning on sail-shaped and saddle-shaped Q235 plates; one-step compensation achieves MAE = 0.68 mm / 0.38 mm and RMSE = 0.7301 mm / 0.4141 mm, respectively, confirming practical feasibility beyond FEM benchmarking. The 3D numerical control plate bending machine for ships faces challenges in achieving springback accuracy during hull plate forming. Traditional methods suffer from limited efficiency and precision due to significant springback prediction errors and simplistic compensation strategies. To address this, a springback dataset is first constructed using finite element simulations. Subsequently, a hybrid prediction model (termed IWOA-CNN-GRU-Attention) is proposed, which integrates convolutional neural network (CNN), gated recurrent unit (GRU), and attention mechanism modules. The model's hyperparameters are optimized by an improved whale optimization algorithm (IWOA), which enhances the standard WOA by incorporating Tent chaotic mapping, Cauchy perturbation, and a greedy strategy to improve search capability. Within the hybrid model, the CNN extracts local features along the sequence axis, the GRU captures temporal dependencies, and the attention mechanism dynamically weights the GRU outputs to focus on critical information. A fully connected layer then generates the springback prediction. Comparative experiments demonstrate that the proposed model achieves superior predictive performance for ship hull plate springback, yielding a more accurate and robust methodological framework for this field.