Abstract:To validate the second-by-second dynamics of turbines in field experiments, it is necessary to accurately reconstruct the winds going into the turbine. Current time-resolved inflow reconstruction techniques estimate wind behavior in unobserved regions using relatively simple spectral-based models of the atmosphere. Here, we develop a technique for time-resolved inflow reconstruction that is rooted in a large-eddy simulation model of the atmosphere. Our "large-eddy reconstruction" technique blends observations and atmospheric model information through a diffusion model machine learning algorithm, allowing us to generate probabilistic ensembles of reconstructions for a single 10-min observational period. Our generated inflows can be used directly by aeroelastic codes or as inflow boundary conditions in a large-eddy simulation. We verify the second-by-second reconstruction capability of our technique in three synthetic field campaigns, finding positive Pearson correlation coefficient values (0.20>r>0.85) between ground-truth and reconstructed streamwise velocity, as well as smaller positive correlation coefficient values for unobserved fields (spanwise velocity, vertical velocity, and temperature). We validate our technique in three real-world case studies by driving large-eddy simulations with reconstructed inflows and comparing to independent inflow measurements. The reconstructions are visually similar to measurements, follow desired power spectra properties, and track second-by-second behavior (0.25 > r > 0.75).
Abstract:In order to accurately reconstruct the time history of the atmospheric state, ensemble-based data assimilation algorithms need to be initialized appropriately. At present, there is no standard approach to initializing large-eddy simulation codes for microscale data assimilation. Here, given synthetic observations, we generate ensembles of plausible initial conditions using a latent diffusion model. We modify the original, two-dimensional latent diffusion model code to work on three-dimensional turbulent fields. The algorithm produces realistic and diverse samples that successfully run when inserted into a large-eddy simulation code. The samples have physically plausible turbulent structures on large and moderate spatial scales in the context of our simulations. The generated ensembles show a lower spread in the vicinity of observations while having higher variability further from the observations, matching expected behavior. Ensembles demonstrate near-zero bias relative to ground truth in the vicinity of observations, but rank histogram analysis suggests that ensembles have too little member-to-member variability when compared to an ideal ensemble. Given the success of the latent diffusion model, the generated ensembles will be tested in their ability to recreate a time history of the atmosphere when coupled to an ensemble-based data assimilation algorithm in upcoming work. We find that diffusion models show promise and potential for other applications within the geosciences.