GeoWind2Plan: Mission-Time 3D Urban Wind Prediction for Energy-Efficient UAV Planning

Shaoxiang Qin, Yucheng Zhao, Fuyuan Lyu, Di Zhou, Jiachen Yao, Xue Liu, Anima Anandkumar, Liangzhu Leon Wang, Xiongye XiaoarXiv 2026 · arXiv

TL;DR

GeoWind2Plan is a new framework that brings high-fidelity, building-resolved wind awareness to UAV mission planning in real-time. By replacing hours-long CFD simulations with a 3-second neural operator inference, it allows drones to exploit complex urban wind structures for energy savings of up to 12.7%.

What Problem Does This Paper Address?

Planning energy-efficient routes for Unmanned Aerial Vehicles (UAVs) in cities is notoriously difficult because buildings drastically reshape ambient wind. A single background wind vector can turn into a complex 3D field of "street canyons" with high-speed channeling, sheltering wakes, and vertical updrafts.

To fly efficiently, a drone needs to know this local wind field. The gold standard for generating such data is Computational Fluid Dynamics (CFD), but a single high-resolution urban simulation can take 8 hours or more. Since UAV missions typically last only minutes and weather conditions change constantly, relying on traditional CFD for mission-time planning is impossible. Most current planners settle for "wind-agnostic" (zero wind) or simple height-based wind profiles, which miss the energy-saving opportunities hidden in building-induced flow structures.

Methods

The authors propose GeoWind2Plan, a "geometry-to-wind-to-planning" pipeline that operates in seconds. The framework consists of three main components:

  1. Localized Neural Wind Prediction: Instead of a full-city simulation, the system uses a localized geometry-conditioned neural operator (based on Fourier Neural Operators). It takes 3D building geometry—encoded via Signed Distance Functions (SDF) and Multi-Directional Distance Features (MDDF)—and predicts 3D wind patches.
  2. Physics-Based Adaptation: To handle varying wind conditions without retraining, the authors leverage two physical principles:
  3. Yaw Covariance: The city geometry is rotated into a reference-wind frame for inference and then the predicted wind vectors are rotated back.
  4. Mean-Flow Similarity: Predicted wind speeds are scaled linearly based on the background wind speed, a valid approximation for high-Reynolds-number urban flows.
  5. Wind-Aware Trajectory Optimization: The system extracts a 3D corridor between the start and goal, predicts the wind only within that corridor, and then uses a nonlinear optimizer. This optimizer jointly tunes the 3D path and the speed profile using a physically grounded multirotor energy model that accounts for air-relative velocity, drag, and rotor power.
Figure 1: Overview of the GeoWind2Plan framework. Given a start-goal pair, 3D building geometry, and a background wind vector, GeoWind2Plan extracts a mission-relevant corridor, rotates the local geometry into a reference-wind frame, predicts corridor-relevant wind patches with a localized geometry-conditioned neural operator, stitches them into a queryable local 3D wind field, rescales the field from the reference training speed to the given background wind speed, and rotates the predicted wind field back to the city frame. The resulting local wind field is used by an RRT-initialized continuous optimizer to produce an energy-efficient 3D UAV path and speed profile.

Datasets

The model was trained and evaluated using high-fidelity CFD labels generated by CityFFD, a GPU-accelerated urban airflow solver.

  • Training: Approximately twenty 1.2 km × 1.2 km urban CFD cases.
  • Evaluation: Four held-out 1.2 km × 1.2 km standard urban blocks (A–D) and one larger 3 km × 3 km block (E) to test scalability.
  • Missions: 256 feasible start-goal pairs were sampled per evaluation block across different wind-angle regimes (tailwind, headwind, crosswind).

Key Insights and Ideas

  • "Decision-Useful" vs. "CFD-Perfect": A major insight is that a wind predictor doesn't need to be perfectly accurate at every point to be valuable. It just needs to preserve the structures (like wakes and channels) that allow a planner to make better energy-saving decisions.
  • Corridor-Localized Inference: By only predicting wind in a task-relevant 3D corridor (e.g., 20% of the total area), the framework reduces inference time from 15 seconds to just 3 seconds without sacrificing planning quality.
  • Joint Path and Speed Optimization: The paper highlights that energy efficiency isn't just about the route; it's about speed. For example, in a headwind, the optimal strategy is often to fly lower and slower to reduce drag-related power, even if it slightly increases total flight time.

Key Results

GeoWind2Plan demonstrates that fast, localized 3D wind prediction is a practical "mission-time" signal for UAVs:

  • Speedup: Inference takes roughly 3 seconds, compared to 8 hours for traditional CFD—a massive reduction in computational latency.
  • Energy Savings: Compared to wind-agnostic planning, trajectories optimized with GeoWind2Plan reduced energy consumption by 6.9% in tailwinds, 12.7% in headwinds, and 4.5% in crosswinds.
  • Oracle Performance: The framework recovered between 75% and 88% of the energy savings achieved by an "oracle" planner using ground-truth CFD data.
Figure 2: Qualitative trajectory-planning comparison in urban block D. Panels (a)–(d) show four randomly selected start–goal missions with different relative wind angles. Each pair compares the trajectory planned with the offline CFD reference wind field (Ground truth) and the trajectory planned with GeoWind2Plan’s predicted 3D wind field (Ours). Wind-speed slices and horizontal vectors at 75 m are shown as context, with optimized UAV trajectories overlaid.

Limitations and Future Work

  • Steady-State Assumption: The current model uses time-averaged velocity fields and assumes background wind remains fixed during a mission. It does not yet account for short-term gusts or turbulent fluctuations.
  • Safety Risks: The planner focuses on energy efficiency but does not explicitly model turbulence-related safety risks (e.g., high turbulent kinetic energy areas).
  • Future Directions: The authors aim to extend the predictor to include turbulence quantities and support dynamic replanning under time-varying inflow conditions.

Summary updated

Abstract

In urban low-altitude flight, buildings reshape ambient wind into spatially varying 3D flow, making unmanned aerial vehicle (UAV) energy depend on local wind exposure as well as path length. However, building-resolved wind information is rarely available when a mission must be planned. Computational fluid dynamics (CFD) can produce high-fidelity urban flow fields, but each simulation is tied to a fixed inflow boundary condition and can take hours to days, which is incompatible with urban UAV missions that typically last minutes to tens of minutes. We present GeoWind2Plan, a geometry-to-wind-to-planning framework for mission-time 3D urban wind prediction and energy-efficient UAV planning. Given only a background wind vector, 3D building geometry, and a start-goal pair, GeoWind2Plan transforms the building geometry into a reference-wind frame, predicts mission-relevant 3D wind patches with a localized geometry-conditioned neural operator, stitches them into a queryable local wind field, and optimizes a feasible 3D path and speed profile using a physically grounded UAV energy model. Rather than pursuing CFD-perfect reconstruction, GeoWind2Plan targets decision-useful wind prediction: trajectories are planned with predicted wind and evaluated under high-fidelity CFD wind. Across held-out urban domains, wind speeds, and mission wind-angle regimes, GeoWind2Plan performs corridor-localized wind inference in about 3 seconds, compared with roughly 8 hours for CFD. Under CFD evaluation, trajectories planned with GeoWind2Plan reduce energy by 6.9%, 12.7%, and 4.5% in tailwind, headwind, and crosswind missions relative to wind-agnostic planning, recovering 87.9%, 85.7%, and 75.0% of CFD-reference savings. These results show that fast, corridor-localized 3D urban wind prediction can make wind-aware UAV energy planning practical at mission time.
BibTeX
@misc{qin2026geowind,
  author = {Shaoxiang Qin and Yucheng Zhao and Fuyuan Lyu and Di Zhou and Jiachen Yao and Xue Liu and Anima Anandkumar and Liangzhu Leon Wang and Xiongye Xiao},
  title = {{GeoWind2Plan: Mission-Time 3D Urban Wind Prediction for Energy-Efficient UAV Planning}},
  year = {2026},
  eprint = {2609.36056},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.36056},
}

Loading related content...

Figure