Here, we demonstrate that we can indeed match observed second-by-second winds within an LES run by using a machine learning algorithm (known as a diffusion model) and creating observation-informed boundary conditions. To achieve this, we run an ensemble of synthetic field campaigns within LES that match the layout of RAAW. In these synthetic campaigns, we simulate a Doppler lidar that scans upwind of the turbine, and we gather concurrent measurements of the inflow plane 375 m upwind of the turbine (far enough upstream to be unaffected by the wind turbine and close enough to be representative of the flow field it will encounter in a few seconds). We train the diffusion model to generate a plausible inflow plane when given inflow lidar measurements, and then we feed real-world lidar measurements into the diffusion model to generate plausible inflow planes. Crucially, the algorithm can generate many plausible realizations of inflow when given just one set of lidar observations, enabling us to probabilistically reconstruct flow in the turbulent atmospheric boundary layer.
Furthermore, we validate our reconstructed flow against real-world observations, finding good accuracy. RAAW has a second lidar that faces downwind of the turbine. We feed our observation-informed boundary conditions into LES, and we demonstrate that LES then accurately reconstructs downwind flow. With this initial atmospheric work done, future work will examine the second-by-second power and structural dynamics of the turbine.

