Generative Modeling of Microweather Wind Velocities for Urban Air Mobility

TL;DR

This work addresses the challenge of predicting localized microweather wind velocities for urban air mobility by learning a probabilistic mapping from regional macroweather forecasts to local wind profiles. It compares Gaussian Mixture Models, Denoising Diffusion Probabilistic Models, and Flow Matching as conditional generative approaches, demonstrating that DGMs provide superior conditional sampling and generalization to unseen macroweather combinations, while GMMs can underperform in conditional scenarios. The proof-of-concept using SoDAR wind data and nearby forecasts shows that a temporary measurement campaign can yield realistic, stochastic wind samples without permanent sensors or heavy CFD simulations, with potential applicability to UAM safety and reliability. Limitations include the lack of temporal dynamics and spatial extrapolation, motivating future work on space-time field generation, continuous conditioning, and larger NASA datasets to enable broader applicability.

Abstract

Motivated by the pursuit of safe, reliable, and weather-tolerant urban air mobility (UAM) solutions, this work proposes a generative modeling approach for characterizing microweather wind velocities. Microweather, or the weather conditions in highly localized areas, is particularly complex in urban environments owing to the chaotic and turbulent nature of wind flows. Furthermore, traditional means of assessing local wind fields are not generally viable solutions for UAM applications: 1) field measurements that would rely on permanent wind profiling systems in operational air space are not practical, 2) physics-based models that simulate fluid dynamics at a sufficiently high resolution are not computationally tractable, and 3) data-driven modeling approaches that are largely deterministic ignore the inherent variability in turbulent flows that dictates UAM reliability. Thus, advancements in predictive capabilities are needed to help mitigate the unique operational safety risks that microweather winds pose for smaller, lighter weight UAM aircraft. This work aims to model microweather wind velocities in a manner that is computationally-efficient, captures random variability, and would only require a temporary, rather than permanent, field measurement campaign. Inspired by recent breakthroughs in conditional generative AI such as text-to-image generation, the proposed approach learns a probabilistic macro-to-microweather mapping between regional weather forecasts and measured local wind velocities using generative modeling (denoising diffusion probabilistic models, flow matching, and Gaussian mixture models). A simple proof of concept was implemented using a dataset comprised of local (micro) measurements from a Sonic Detection and Ranging (SoDAR) wind profiler along with (macro) forecast data from a nearby weather station over the same time period.

Figures (13)

• Figure 1: Our method for utilizing conditional deep generative models to produce distributions over wind profiles based on known macroweather conditions.
• Figure 2: Empirical distribution of SoDAR-measured velocity components averaged over altitudes. This plot represents all 6542 observations in the dataset.
• Figure 3: Microweather wind speed altitude profiles under different macroweather wind speed conditions where the mean wind speed over altitude is plotted in dark blue with plus and minus one standard deviation shown by the shaded blue regions. Each macroweather wind speed range contains an approximately equal number of samples. High macroweather wind speeds are associated with higher microweather wind speeds.
• Figure 4: Total variance explained for the PCA dimension reduction and BIC evaluation across the number of GMM components. Seven principal components explains $\approx 96.4\%$ of the variance while still achieving realistic results (see \ref{['sec:results']}. $21$ GMM components is roughly the lowest model complexity achieving a low BIC.
• Figure 5: Altitude-averaged: (left) Samples from the fitted GMM model along with kernel density estimations of the marginal $u$ and $v$ distributions. Both the joint and marginal distributions closely resemble those seen in \ref{['fig:real_bivariate']}, providing strong evidence that the GMM is fitting the observed data distribution well. (middle) Samples from the fitted DDPM along with kernel density estimations of the marginal $u$ and $v$ distributions. This fit misses some of the visually identifiable modes as seen in \ref{['fig:real_bivariate']}. (right) Samples from the fitted FM along with kernel density estimations of the marginal $u$ and $v$ distributions. Similar to those results in the DDPM bivariate plot, this fit misses some of the visually identifiable modes as seen in \ref{['fig:real_bivariate']}