Abstract
This work proposes a physics-informed machine-learning framework for sensor-based wind flow estimation along drone trajectories in complex urban environments. The method uses a database of flow simulations under different wind conditions to estimate velocity fields and their associated uncertainty within a target domain, enabling the optimization of sensor placement to minimize uncertainty. The framework introduces several innovations: scalability with domain complexity through a physics-based domain decomposition, the ability to extrapolate beyond the training wind conditions, and flexible sensor positioning as a free input variable. The approach combines Reynolds-number-based scaling of flow variables, cluster-based representations of subdomain flows, entropy-based correlations between subdomains, and a multivariate probabilistic model linking sensor measurements to velocity estimates. The methodology is demonstrated on drone trajectories through a three-building configuration, illustrating its capability for urban wind estimation and sensor optimization. The framework is expected to scale to larger urban domains and incorporate weather information in future applications.
