A deep learning-based method for rapid prediction of transient loads on flexible regions of local flexible hydrofoils

Abstract

As an essential component of maritime renewable energy apparatus, the rational design of local flexible hydrofoils is crucial for enhancing energy capture efficiency. However, the efficacy of optimization process is limited by the inability of conventional preliminary design frameworks to rapidly acquire transient load distributions, which are necessary for precise subsequent design to establish the appropriate optimization threshold. To determine the relationship between geometric parameters, boundary conditions, and transient load, a hybrid neural network is established, which includes convolutional neural networks (CNN) for geometric feature extraction, bidirectional long short-term memory (BiLSTM) for dynamic information capture, and an attention mechanism to enhance critical features. The modified grey wolf optimizer, combining nonlinear adaptation and adaptive dynamic weight factors, enhances prediction performance through hyperparameter optimization. The CNN-BiLSTM-Attention model captures dynamic pressure load fluctuations with an accuracy of approximately 96 %, displaying strong generalization across varied geometric factors and boundary conditions. Nevertheless, the inherent unsteady spatiotemporal characteristics of unstable flows such as vortex shedding and transient separation increases the difficulty of accurate prediction. Compared to existing algorithms, the improved Grey Wolf Optimizer (IGWO) has higher robustness and convergence rates for complicated nonlinear problems. Furthermore, hyperparameter optimization enhances prediction accuracy and computing efficiency. These results demonstrate the effectiveness of this deep-learning approach for improving the design of localized flexible hydrofoils.