A robust deep learning framework for rapid hydrodynamic analysis and design of multi-Salter's duck wave energy farms

Abstract

Ocean wave energy offers substantial potential for sustainable power generation, but designing efficient wave energy converter (WEC) farms faces high computational costs from complex hydrodynamic simulations. This study presents a novel deep learning-based surrogate modeling framework to accelerate the analysis and design of multi-Salter's duck WEC arrays. It decomposes hydrodynamic response prediction into two stages: first, estimating individual multi-Salter's duck unit dynamics, then predicting interaction effects between neighboring units. Specialized deep learning architectures for each stage were optimized via neural architecture search (NAS) and trained on a comprehensive dataset spanning diverse design parameters, array configurations, and environmental conditions. Against nine tuned benchmarks, our framework achieved a 19.3 % reduction in mean absolute error (MAE) for individual unit dynamics and a 62.5 % MAE reduction for interaction effects compared to the best-performing benchmarks. A case study using realistic wave climates from southeast Australia validated the model's accuracy in predicting annual average power absorption and interaction factors, while drastically reducing computational time relative to traditional numerical methods. This efficient and accurate surrogate model enables optimized co-design and advances the feasibility of large-scale wave energy deployment.