Multivariate prediction on wake-affected wind turbines using graph neural networks

Modern wind turbines are large and slender dynamical structures with a fatigue loading profile of complex nature. The guarantee of their structural integrity is paramount for materializing cost efficient and more reliable wind energy. The measurement of the global dynamic response and loads of wind turbines is fundamental for achieving this goal. However, an industry-wide, cost-effective direct sensing framework is yet to arise. Moreover, deploying physical sensors and measurement systems on every structural component of interest of a wind turbine induces prohibitive costs in deployment, maintenance and data management. Considering that direct fluid-structure interaction simulations on a farm level are not computationally feasible, the preferred path for structural response estimation on wind farms has been surrogate modelling. Within this landscape, new model architectures have risen in recent years which are able to take into account graph structured data (i.e. non-euclidean data). Wind turbines positioned in a farm, where there is a layout- and topology-dependant interplay of aerodynamic wake affecting the loading profile and power production, lend themselves perfectly to this paradigm. Thus, in this contribution, we introduce the use of graph neural networks (GNN) for layout agnostic saptio-temporal joint modelling of fatigue loads effects, rotor-averaged wind speed and power production on individual turbines of wind farms. To this end, we generate stochastic dependent samples of inflow conditions for wind speed, wind direction, wind shear and nacelle yaw angles. Additionally, wind farm layouts are randomly generated based on different geometric shapes (rectangle, triangle, ellipse and sparse circles) with random parametrization (varying orientations, length/width ratio) for different numbers of turbines and minimal distance (based on the rotor diameter). Both the arbitrary layouts and the random inflow conditions are used as inputs for PyWake, a wind farm simulation tool capable of calculating wind farm flow fields, power and fatigue loads. In our analysis, we develop and compare the performance of the GENeralized Aggregation Networks (GEN), the Graph Attention Networks (GAT) and the Graph Isomorphism Network with Edges (GINE) in their accuracy and ability to generalize their joint predictions for unseen layouts, uncertain inflow conditions and fatigue load estimation on the blade root, tower top and tower base of any wind turbine in the farm. Our results indicate that the GEN layer yields the best performance, followed by GINE, while the GAT layer under-performs and is unable to differentiate between different wake conditions. We further observe that the GAT layer causes a latent space collapse, due to the coupled effect of the manner in which we initialise node features and the way in which its messages are computed.