Machine-learned flow estimation with sparse data—Exemplified for the rooftop of an unmanned aerial vehicle vertiport

Abstract

This work proposes a physics-informed data-driven framework for urban wind estimation using sparse sensor data. The approach incorporates Reynolds number independence across different working conditions, enabling extrapolation to wind regimes beyond those present in the training data. A machine-learned non-dimensionalized manifold is constructed from snapshot flow data and used to represent the velocity field through a double encoder–decoder architecture. The first encoder normalizes the data with respect to the incoming wind speed, while the second projects the normalized data onto an isometric feature-mapping manifold. Decoding is performed using a k-nearest-neighbor reconstruction followed by denormalization. Clustering is employed to coarse-grain the manifold representation and reduce computational costs. The sensor-based estimation combines the prediction of incoming wind speed with a mapping from sensor measurements to latent manifold variables. The framework is demonstrated for the flow above an unmanned aerial vehicle vertiport rooftop, showing good generalization capabilities for rare wind conditions not included in the training database.