A graph convolutional network-based surrogate model with enhanced graph embedding for real-time prediction of wind turbine mainframe stress distribution

Graph Convolutional Neural Networks have been extensively applied in predicting attributes of mesh-based simulations, including Finite Element analysis. However, when the mesh of finite element models is non-uniform and the boundary conditions vary abruptly, such as in the finite element model for stress assessment of wind turbine mainframes, the accuracy of Graph Networks is compromised. In response to these limitations and to facilitate real-time stress field prediction for rapid design iteration of wind turbine mainframes, this paper proposes a surrogate model based on Graph Convolutional Networks with enhanced graph embedding. By implementing an additional master vertex and global vertex connectivity, the Graph Convolutional Network model leverages message-passing mechanisms to learn relationships between node attributes, external loading conditions, and stress distribution in an effective way. The proposed architecture includes an encoder–decoder framework with three message-passing layers. Numerical experiments demonstrate that this Graph Convolutional Network-based model achieves high precision (mean absolute percentage error < 9 % compared to finite element results) and strong generalization ability in predicting von Mises stress distributions under varying geometries and large-range boundary conditions, outperforming Graph Convolutional Network models without enhanced graph embedding. Furthermore, the model reduces computation time by orders of magnitude compared to traditional finite element solvers with less hardware usage, making it suitable for iterative design processes. • Application of Graph Networks in Industrial-Level Dense and Complex Meshes of Wind Turbine Design. • Enhanced Graph Embedding for Adaptive Meshing and Broadly Varying Boundary Conditions. • Orders-of-Magnitude Faster Prediction Time Relative to Finite Element Simulation.