GeoWind2Plan: Mission-Time 3D Urban Wind Prediction for Energy-Efficient UAV Planning
Abstract Overview
GeoWind2Plan is a mission-time framework that predicts building-resolved 3D urban wind along a UAV mission corridor and uses that prediction for energy-aware trajectory planning. The method rotates local geometry into a reference wind frame, applies a localized geometry-conditioned neural operator to infer mission-relevant wind patches, rescales predictions for different inflow speeds, and stitches them into a queryable local wind field. It then combines this field with a physically grounded multirotor energy model and an RRT-initialized continuous optimizer to jointly refine path and speed in 3D. The paper emphasizes decision-useful wind prediction rather than full CFD reconstruction, evaluating planned trajectories under high-fidelity CFD wind.
Novelty
The distinctive contribution is the integration of ambient-wind-conditioned, building-resolved 3D urban wind prediction with continuous UAV energy planning at mission time. The framework is also notable for its wind-frame geometry transformation and corridor-localized inference strategy, which reduce the need for full-city CFD while preserving route-relevant flow structure.
Results
Across held-out urban domains and wind-angle regimes, corridor-localized inference took about 3 seconds versus roughly 8 hours for CFD, while full-field neural prediction took about 15 seconds. Under CFD evaluation, trajectories planned with GeoWind2Plan reduced energy relative to wind-agnostic planning by 6.9% in tailwind, 12.7% in headwind, and 4.5% in crosswind missions, recovering 87.9%, 85.7%, and 75.0% of the savings achieved by CFD-reference planning. Corridor localization to 20% width reduced inference time from 15 seconds to 3 seconds with little change in CFD-evaluated energy.
Key Points
- GeoWind2Plan predicts mission-relevant 3D urban wind from building geometry and a background wind vector using a localized geometry-conditioned neural operator with wind-direction and wind-speed adaptation.
- The planner jointly optimizes continuous 3D path and speed using the predicted local wind field and a physically grounded UAV energy model, rather than assuming fixed speed or discrete routes.
- Empirically, the method remains closer to CFD-reference planning than wind-agnostic or height-only profile baselines across held-out blocks, wind regimes, and tested inflow speeds.