Physics-guided spatial-temporal graph neural network for finite element surrogate modelling of a floating offshore wind turbine with wind-wave load directionality effect

Digital Twin (DT) modelling is a crucial asset to provide structural health monitoring (SHM) for Floating Offshore Wind Turbine (FOWT), which aims to enhance the structural integrity and minimize the operation and maintenance expenditure (OPEX) under severe ocean environment. DT models are required to capture the important dynamics and run in real-time. Data-driven DT models are beneficial to produce highly nonlinear dynamic models in real-time and the existing models are mostly based on traditional machine learning techniques, which generate the surrogate models in a “black box” manner, the input and output data are blindly fed. This hinders the generalizability and adaptability to unknown scenarios.

To address the SHM practical issues and overcome the limitation of existing data-driven DT models, we present the first Explainable Artificial Intelligence (XAI) framework for structural dynamics of FOWT with the expressive power of the novel Heterogeneous Spatial Temporal Graph Neural Network. The aero-hydro-servo coupled simulation results are generated from the software QBlade. The semi-submersible OC4 DeepCwind FOWT is analysed with the circumstances of above rated wind speed with different wind directions, rough sea state and sea current condition. This novel XAI framework can interpret the complicated aero-hydrodynamic coupling of FOWT, the imbalance loads created from the misalignment of wind inflow directions against the wave and current direction are investigated. Remarkably this is the first XAI illustrating the wind-wave load directionality with latent chaotic structural dynamics interaction in time domain. Time domain analysis is more prominent to dynamic internal stress analysis and nonlinear effect for fatigue analysis than frequency domain analysis, especially this new XAI model can include the second-order hydrodynamics in real-time and provide more precise remaining useful life predictions for high-cycle fatigue in the subsequent stages.