In low‑altitude urban flight, buildings reshape the ambient wind into a spatially varying 3D flow, making UAV energy consumption depend not only on path length but also on local wind exposure. Conventional computational fluid dynamics (CFD) can generate high‑fidelity wind fields, but each simulation is tied to a fixed inflow condition and takes hours to days, which is incompatible with typical UAV missions that last only minutes to tens of minutes.
GeoWind2Plan introduces a geometry‑to‑wind‑to‑planning pipeline that requires only a background wind vector, 3D building geometry, and a start‑goal pair to produce mission‑time wind predictions and energy‑optimal trajectories. The core steps are:
- Transform the building geometry into a reference‑wind frame to align all data with the wind direction;
- Apply a localized geometry‑conditioned neural operator to predict 3D wind patches along the mission corridor;
- Stitch the predicted patches into a queryable local wind field;
- Optimize a feasible 3D path and speed profile using a physics‑based UAV energy model.
Instead of aiming for CFD‑perfect reconstruction, GeoWind2Plan targets decision‑useful wind prediction: trajectories planned on the predicted wind are later evaluated on high‑fidelity CFD wind. Across unseen urban domains, various wind speeds, and mission wind‑angle regimes, the method achieves:
- Wind inference in about 3 seconds, compared with roughly 8 hours for CFD;
- Energy reductions of 6.9% (tailwind), 12.7% (headwind), and 4.5% (crosswind) relative to wind‑agnostic planning when evaluated with CFD;
- Recovery of 87.9% / 85.7% / 75.0% of the savings obtained by CFD‑reference planning.
These results demonstrate that fast, corridor‑localized 3D urban wind prediction makes wind‑aware UAV energy planning practical at mission time.
Review