In this study, we propose a novel path optimization methodology for UAV drones operating in complex urban wind environments. The method integrates high-fidelity wind field modeling, airspace grid discretization, and an improved A* algorithm that dynamically accounts for wind effects, search efficiency, and path smoothness. Our approach significantly enhances flight safety and efficiency in densely built urban areas.
Introduction
Urban air mobility has emerged as a promising solution to ground traffic congestion and last-mile delivery challenges. UAV drones offer autonomy and flexibility, but their low-altitude operations are severely impacted by complex wind fields induced by dense building clusters. Wind shear, strong crosswinds, and local turbulence can destabilize UAV drones, leading to safety risks and reduced efficiency. Existing path planning methods often overlook three-dimensional wind heterogeneity or fail to jointly optimize safety, search efficiency, and path quality. In this work, we develop a grid-based wind field representation and an improved A* planner that incorporates dynamic step size, line-of-sight reachability, and wind-aware cost functions. The proposed method is validated through simulations in a representative urban district in Shenzhen, demonstrating significant improvements in flight time and path smoothness.
Urban Wind Field Effects on UAV Drones
The wind field in urban canyons is highly non-uniform due to building geometries and street orientations. Dominant wind directions, identified from wind rose diagrams, interact with buildings to create high-speed channels, vortices, and sheltered zones. For our study area (Figure 1), the prevailing wind is from the north-northeast at speeds of 4–8 m/s, with gusts up to 12 m/s. Wind speeds exceeding the maximum tolerable wind speed of the chosen UAV drone (e.g., DJI M350, limit 12 m/s) define no-fly zones. Even below this threshold, headwinds increase energy consumption and flight time, while tailwinds reduce them. Thus, any path optimization must incorporate local wind vectors to minimize total flight time while avoiding hazardous regions.
Wind Field Modeling and Gridding
CFD Simulation
We construct a three-dimensional digital model of the urban area using high-resolution satellite imagery and elevation data. The model includes precise building footprints and heights. Computational Fluid Dynamics (CFD) simulations are performed using the RANS approach with boundary conditions derived from local wind rose data. The simulation domain is discretized into a hybrid mesh: fine cells (0.3 m) near building surfaces and ground, coarser cells (2 m) away from obstacles. The simulation outputs wind velocity vectors at all grid nodes. We then map these vectors onto a uniform cubic grid of 2 m resolution, which balances accuracy and computational cost. Each grid cell stores the center coordinates and the average wind velocity components Vw,i = (Vx,i, Vy,i, Vz,i). This grid serves as the environment representation for path planning.
Grid Sensitivity Analysis
We test grid sizes of 1 m, 1.5 m, 2 m, 2.5 m, and 3 m. The 2 m grid yields the best trade-off between flight time and number of turn points, as summarized in Table 1.
| Grid Size (m) | Flight Time (s) | Number of Turns |
|---|---|---|
| 1 | 45.79 | 30 |
| 1.5 | 42.31 | 18 |
| 2 | 42.27 | 7 |
| 2.5 | 45.63 | 23 |
| 3 | 43.46 | 25 |
Path Planning Model and Algorithm
Problem Formulation
Let the airspeed vector of the UAV drone be Va (constant magnitude 12 m/s), the wind vector Vw, and the ground velocity Vg. The relationship is:
$$ \vec{V}_g = \vec{V}_a + \vec{V}_w. $$
The magnitude of the ground speed on segment i is:
$$ V_{g,i} = V_a \cos\alpha + V_{w,i} \cos\beta, $$
where α is the angle between airspeed and ground speed, and β is the angle between wind and ground speed. The total flight time T is the objective:
$$ \min T = \sum_{i=1}^{n} \frac{dist(c_{i-1}, c_i)}{V_{g,i-1}}, $$
subject to constraints: no-fly zone avoidance (wind speed > max tolerable, building blocks plus 4 m buffer), node connectivity, altitude limits (80–150 m), path length ≥ grid size, and spatial bounds.
Improved A* Algorithm
Traditional A* uses a fixed step size, which fails in narrow urban corridors and generates excessive nodes. Our improved algorithm introduces three enhancements:
- Dynamic step size: The step size Si is adapted based on the distance to the goal and to the nearest obstacle:
$$ S_i = \min(k \times S_0 + \mu, S_{min}), $$
where k = Si,goal / Sstart,goal, S0 is initial step, μ is a small offset to avoid zero step.
- Line-of-sight check: When expanding a node, we check if a direct connection from the grandparent node to the current node is feasible (no obstacles or high-wind zones). If yes, we bypass the intermediate node, reducing turns.
- Wind-aware cost function: The total cost f(n) = g(n) + h(n), where g(n) is the accumulated flight time from start to node n, and h(n) is the heuristic estimate of flight time from n to goal using expected ground speed Vg,exp:
$$ g(n) = \sum_{i=1}^{n} \frac{dist(c_{i-1}, c_i)}{V_{g,i-1}}, $$
$$ h(n) = \sum_{i=n}^{k} \frac{dist(c_{i-1}, c_i)}{V_{g,exp}}. $$
Experimental Setup and Results
Scenario
We select a 1 km × 0.9 km area in Nanshan District, Shenzhen, with building density 62% and heights up to 239 m. The dominant wind (NNE) at 4, 6, and 8 m/s is simulated. A DJI M350 UAV drone (max wind tolerance: 12 m/s, cruise speed: 12 m/s) is used. Three delivery target points are defined from a single start. Additional sub-scenarios (left and right zones) and a non-stationary wind case (8 m/s → 6 m/s → 4 m/s at waypoints) are tested.
Comparative Tests
We compare our improved algorithm (IA*) against two baselines: (1) Traditional A* with Euclidean distance cost (ignoring wind); (2) Wind-aware A* without dynamic step or line-of-sight (denoted WA*). All experiments use the 8 m/s wind field for consistency. Results are summarized in Table 2.
| Algorithm | Path ID | Flight Distance (m) | Flight Time (s) | Number of Turns | Risk Zone Distance (m) |
|---|---|---|---|---|---|
| IA* (ours) | 1 | 668.13 | 42.27 | 7 | 0 |
| 2 | 237.46 | 12.19 | 8 | 0 | |
| 3 | 406.85 | 20.52 | 32 | 0 | |
| Traditional A* | 1 | 828.00 | 69.00 | 44 | 48.00 |
| 2 | 264.00 | 22.00 | 12 | 52.00 | |
| 3 | 464.00 | 38.67 | 35 | 28.00 | |
| WA* (wind-aware only) | 1 | 826.80 | 42.82 | 50 | 0 |
| 2 | 266.70 | 14.11 | 18 | 0 | |
| 3 | 1017.85 | 41.19 | 44 | 0 |
Compared to Traditional A*, our method eliminates all risk-zone penetration and reduces average flight time by 42.2% and turns by 48.5%. Compared to WA*, our method reduces average flight time by 23.6% and turns by 58.2%, while maintaining zero risk exposure. The dynamic step and line-of-sight checks are responsible for the efficiency gains.
Multi-Scenario and Non-Stationary Wind Tests
We test the left and right sub-areas (different building layouts). Our algorithm consistently produces safe, smooth paths with no risk zone crossings. Table 3 reports the performance.
| Scenario | Path ID | Flight Distance (m) | Flight Time (s) | Number of Turns | Risk Zone Distance (m) |
|---|---|---|---|---|---|
| Left zone | 1 | 475.50 | 27.83 | 19 | 0 |
| Right zone | 2 | 400.63 | 29.92 | 2 | 0 |
In the non-stationary wind simulation (8→6→4 m/s), the algorithm automatically triggers a re-planning at the waypoints where wind data changes. The resulting path remains smooth and risk-free, with total flight time of 43.03 s, 15.13 s, and 44.50 s for the three segments. This demonstrates the robustness of our method to dynamic wind conditions.
Smoothness Enhancement
To further improve path smoothness, we apply B-spline interpolation at the turn points. This reduces abrupt direction changes, making the trajectory feasible for real UAV drone control. The smoothed paths maintain the same risk-free integrity and efficiency.
Conclusion
We have presented a comprehensive path optimization framework for UAV drones in urban wind fields. By combining CFD-based wind field gridding with an improved A* algorithm that adapts step size, checks line-of-sight, and incorporates wind-aware costs, we achieve significant improvements in safety (zero risk-zone encounters), efficiency (23.6%–42.2% flight time reduction), and path smoothness (48.5%–58.2% fewer turns). The method demonstrates strong adaptability across different building layouts and non-stationary wind scenarios. Future work will extend to multi-UAV coordination and real-time obstacle detection.