Estimation of Fatigue Damage of Offshore Wind Turbine Jackets Using Numerical Simulations and Artificial Neural Networks

Abstract

[Abstract] This thesis studies fatigue damage in welded joints of jacket substructures for offshore wind turbines (OWTs) and proposes a practical way to estimate junction-level damage at low computational cost during early design. The motivation is straightforward: jackets are attractive in the 40–60 m water-depth range, but the large number of welded tubular joints makes fatigue a design driver. Conventional time-domain simulations followed by joint-level fatigue checks are accurate but expensive when many jacket alternatives must be screened. The goal here is to keep the physics and standard practice in view while reducing computational cost to make broad design exploration feasible. The work is structured in three parts. First, a finite-element (FE) model is built for a reference OWT with a jacket substructure. The structure is represented as a 3D frame of circular steel tubes with Bernoulli beam elements. Environmental loading is parameterized compactly with four quantities: wind direction, annual mean wind speed, a wind–wave factor linking wind to wave height, and a wind non-uniformity factor that controls how short-interval signals compound to one year. Wave forces are computed with Airy kinematics and Morison’s equation. Damping is handled by small modal ratios appropriate for fixed steel structures. Modal analysis is performed, and the system is solved using the Method of Modal Superposition. The model is checked against the OC4 jacket; the first modes and frequencies agree well, which gives confidence in the structural modeling. A fatigue computation workflow is implemented. Nominal FE stresses are converted to hot-spot stresses using Efthymiou’s parametric equations for stress concentration factors in tubular joints. The hot-spot stress time series are rainflow-counted to obtain stress-range cycles, which are then evaluated using an appropriate S–N curve. Damage is accumulated by Miner’s rule. The analyses are repeated for numerous structural scenarios involving different geometrical and loading combinations. The resulting damage distribution exhibits a strong skew toward very small damages. A supervised artificial neural network (ANN) is trained as a metamodel to approximate the physics-based pipeline at junction level. Each junction is encoded by 30 features that summarize (i) global jacket geometry, (ii) global loading parameters, and (iii) local joint information. The model is a fully connected multilayer perceptron with ReLU activations and a linear output; training uses Adam and mean-squared error with Early Stopping. Hyperparameter grids are run over depth, width, and batch size.