End-to-end Wind Turbine Wake Modelling with Deep Graph Representation Learning

As one of the cleanest and most sustainable sources of renewable energy, wind energy has been undergoing rapid and unabated expansion worldwide. As the capacity of wind turbine farms increases, through the potentially closer clustering of increasing numbers of larger turbines to most efficiently exploit the available wind energy resource, it is inevitable that downstream turbines will at some times be operating within the full or partial wakes of upstream turbines. This can lead to reduced power generation as well as increased structural loads. Consequently, wind turbine wake modelling has been widely considered as one of the most crucial aspects of the optimal design and operational control of wind farms, see [1] and the references therein. Wake models across different levels of fidelity have been thoroughly studied by researchers over the years. Analytical models including the Jensen model [2], the Larsen model [3] and the Gaussian wake model [4] are commonly implemented in industrial standard software such as FLORIS [5], thanks to their very rapid execution speed, however their accuracy is consequently limited. In comparison, higher fidelity models based on computational fluid dynamics (CFD) simulations, such as Reynolds-Averaged Navier-Stokes (RANS) or Large Eddy Simulation (LES), can provide more accurate flow field predictions but at significantly higher computational cost and execution time, hampering their value for rapid resource assessment, and as part of iterative design optimisation and control tools. For instance, the computing time required by RANS modelling for the simulation of a wind farm tends to be in the order of several CPU hours, whereas LES simulations could take days of distributed computation on hundreds of processors [6].