Mesh-based data-driven approach for optimization of tidal turbine blade shape

As the global demand for renewable energy rises, tidal energy has emerged as a reliable and predictable source. Yet, optimizing the design of horizontal axis tidal turbine (HATT) blades to maximize their efficiency presents a formidable challenge. This paper addresses this challenge by introducing TurbineNet, an innovative data-driven model designed to enhance the prediction of hydrodynamic performance for blade shape optimization. Specifically, TurbineNet leverages a universal mesh-based representation of the blade, incorporating advanced mesh convolution and fully connected neural network layers to extract and analyze spatial and structural features. This approach achieves remarkable prediction accuracy with an error margin of less than 2 %. By integrating TurbineNet with free form deformation (FFD) and differential evolution (DE) optimization algorithms, the study optimizes the NREL S814 blade model to maximize its power coefficient. The optimization results demonstrate a substantial improvement, with at least 20 % increase in power coefficient across various tip-speed ratios. Furthermore, incorporating complex blade surface deformations through FFD offers an additional improvement of nearly 1 % compared to traditional optimization methods that focus exclusively on variations in twist angle and chord length. These findings underscore the efficacy of utilizing neural networks for blade design recognition and optimization, significantly enhancing both the performance prediction and design space for HATT blades. This study represents a significant advancement in the application of neural networks to tidal turbine optimization, paving the way for future research in addressing practical design considerations such as inflow fluctuations, structural strength, and refined experimental validation.