Accelerating wind farm structural load estimation through flexible surrogate modeling

Offshore wind capacity will grow substantially in the coming years. The farms of the future will have
tighter fatigue life designs and see control strategies that operate them more aggressively from a
fatigue point of view. Therefore, we need to estimate fatigue rates at each critical location for every
turbine of the farm. Currently, structural loads can either be numerically modelled, which is often too
slow for real-time fatigue-aware optimization or design, or surrogated by data-driven methods with
real-world monitoring data, which are restricted to their training space. Furthermore, there are still
relevant hurdles on virtual sensing. Thus, we propose a flexible surrogate model based on graph
neural networks (GNN). We train GNNs on farm and turbine numerical simulations to understand
fundamental physics (e.g. wake or structural dynamics), learn to generalize and extrapolate turbine
response for unseen scenarios. We specifically use GNNs to perform fatigue load forecasting in wind
energy systems by (i) use GNNs to serve as surrogates of a population of wind farms, (ii) create a
GNN structural surrogate of several turbine types based on numerical models, (iii) combining GNNs
with transfer learning to adapt to new (real) farms/turbines. By enabling wind farm and turbine
fatigue load forecasting for unseen conditions and structures, my proposal can alter wind farm
development by accelerating scenario testing and wind farm operation by leading to fatigue-aware
farm control.