In contemporary urban logistics, unmanned aerial vehicles (UAV drones) have emerged as a pivotal solution for circumventing ground traffic congestion and fulfilling last-mile delivery demands. However, the operational safety and efficiency of UAV drones in densely built urban environments are severely compromised by complex wind field disturbances, such as wind shear and strong crosswinds induced by building clusters. This study presents a novel framework that integrates high-fidelity wind field gridding with an enhanced path planning algorithm to address these challenges. We leverage Computational Fluid Dynamics (CFD) simulations, grounded in real-world Geographical Information System (GIS) data, to model the three-dimensional wind environment around complex building geometries. The continuous wind field is discretized into uniform cubic grids, creating a structured environment that quantifies wind vectors at each spatial location. To overcome the limitations of conventional A* algorithms, which often suffer from fixed search steps and a lack of wind adaptivity, we propose an improved method. This method incorporates a dynamic step size, a line-of-sight reachability check, and a wind-field-aware cost function to optimize flight time while ensuring safety. The algorithm dynamically adjusts the search granularity based on proximity to obstacles and the goal, and its cost function directly uses ground speed derived from the vector sum of airspeed and wind velocity. Experimental validation in a typical high-density district of Shenzhen demonstrates that our methodology successfully avoids hazardous high-wind zones. Critically, it reduces the average flight time by 23.6% and decreases the number of path waypoints (turns) by 58.2% compared to baseline wind-aware algorithms, thereby improving both trajectory smoothness and overall operational efficiency. The results confirm that the proposed framework provides a robust and efficient solution for navigation of UAV drones in challenging urban wind conditions.
1. Introduction
The rapid urbanization of modern cities has exacerbated traffic congestion, creating an urgent need for alternative transportation modalities. UAV drones, with their high autonomy and flexibility, are increasingly recognized as a transformative solution for urban parcel delivery, infrastructure inspection, and emergency response. However, the safe integration of UAV drones into low-altitude urban airspace faces a critical bottleneck: the complex and turbulent wind fields that develop around dense building arrays. The interaction of prevailing winds with varied architectural forms generates unpredictable wind patterns, including accelerated flow through street canyons, recirculation zones, and strong downdrafts. These phenomena pose significant risks to the flight stability and controllability of UAV drones. For instance, sudden wind gusts can cause position deviations and increase energy consumption, while persistent crosswinds can deviate a UAV drone from its intended path, leading to potential collisions or mission failure.
Existing research has extensively explored the interaction between wind and small UAV drones, confirming the amplification of positioning errors under gusty conditions and the increased energy demands when flying into a headwind. In the context of path planning, traditional methods like the A* algorithm are widely used for their efficiency and simplicity. However, conventional A* implementations typically rely on fixed step sizes and heuristic functions based on Euclidean distance, which are insufficient for the spatiotemporally varying nature of urban wind. While recent studies have begun to incorporate wind data for energy-efficient planning, they often operate in two-dimensional spaces or fail to fully integrate the three-dimensional heterogeneity of wind and obstacles. Furthermore, emerging methods such as reinforcement learning and model predictive control offer high adaptivity but suffer from high computational costs and training complexity in large-scale urban environments, making them less practical for real-time applications.
To bridge these gaps, this paper introduces an improved path optimization methodology specifically designed for the safe and efficient navigation of UAV drones in urban wind fields. The core contributions of this work are threefold. First, we establish a three-dimensional wind field simulation model that fuses GIS building data with CFD technology to accurately capture the spatial heterogeneity of wind. This continuous field is then discretized into a structured grid of uniform cubic cells, which serves as the operational environment for path planning. Second, we propose an enhanced A* algorithm that moves beyond single-objective optimization. Our method integrates a dynamic step size that adjusts based on the distance to the goal and nearby hazards, a line-of-sight reachability test to reduce path redundancy, and a wind-field-aware cost function that minimizes total flight time by calculating ground speed from the combined airspeed and wind vectors. Third, we conduct rigorous experiments using a real-world urban scene in Shenzhen, comparing our algorithm against both traditional and wind-aware baselines. The results demonstrate significant improvements in safety, flight time reduction, and path smoothness, validating the effectiveness of our integrated approach for the navigation of UAV drones.
2. The Impact of Urban Wind on UAV Drone Operations
Understanding the characteristics of low-altitude wind in densely built areas is paramount for planning safe routes. The interplay between prevailing winds and the built environment creates a highly non-uniform flow field. Buildings act as porous obstacles, channeling, accelerating, and deflecting wind. The resulting microclimate can lead to significant variations in wind speed and direction over very short distances. For a UAV drone operating in such an environment, these variations translate directly into varying levels of risk and energy expenditure.
Wind speeds in building-induced corridors or at building corners can exceed the ambient regional wind speed by a factor of two or more. When these velocities surpass the maximum operational wind speed of a UAV drone—for instance, 12 m/s for a DJI M350 or lower for lighter consumer models—the affected area becomes a flight-restricted zone or a hazard area. Beyond these absolute safety limits, relative wind direction profoundly affects flight efficiency. A UAV drone flying into a headwind experiences reduced ground speed, increasing flight time and energy consumption. Conversely, a tailwind enhances ground speed and improves efficiency. The difference in mission duration between a headwind and tailwind flight on the same route can be 20-30%. Therefore, an optimal path planning algorithm must not only avoid high-risk zones but also capitalize on favorable tailwinds while mitigating the negative impacts of adverse winds. This requires a detailed map of the wind field, which we achieve through a comprehensive gridding methodology.
3. Construction of a Gridded Wind Field Environment
3.1 Wind Field Simulation
The first step in our methodology is to determine the dominant wind characteristics of the target urban area. This is achieved by analyzing local meteorological data through a wind rose diagram, which displays the long-term frequency distribution of wind speed and direction. For our study region, we analyzed data from the Shenzhen Meteorological Bureau. The dominant wind direction was identified as north-northeast, with speeds typically in the range of 4-8 m/s.
To simulate the complex flow around buildings, we constructed a high-fidelity geometric model of the urban environment. Using 0.5-meter resolution satellite imagery and digital elevation models, we extracted the precise footprints and heights of all buildings in a 1 km by 0.9 km area. This geometric data was imported into a CFD pre-processor to create a three-dimensional solid model serving as the physical boundary for the simulation. The computational mesh was refined near building surfaces to capture local flow details, with a maximum cell size of 0.5 m on building surfaces and 0.3 m in the near-ground flight layer, while coarser cells were used away from buildings to balance computational cost. CFD simulations were run for three dominant wind speeds (4 m/s, 6 m/s, and 8 m/s) with the north-northeast direction. The output was a full three-dimensional dataset containing the wind velocity vector at all computational nodes.
3.2 Environmental Gridding
The continuous CFD wind field data is transformed into a structured, discrete environment suitable for path planning through a process of spatial gridding. The three-dimensional airspace is conceptualized as a large cuboid volume with dimensions X, Y, and Z. This volume is then partitioned into a set of uniform, non-overlapping cubic cells (voxels) of size Lgrid. The number of cells along each axis is determined by K = int(X/Lgrid), M = int(Y/Lgrid), and N = int(Z/Lgrid). The selection of Lgrid is a critical trade-off. A smaller grid size provides higher resolution but increases computational load. Based on the UAV drone’s dimensions (0.85 m wheelbase), the resolution of the CFD mesh, and a sensitivity analysis, we selected Lgrid = 2 m as the optimal granularity for our experiments.
Each grid cell, Ci(xi, yi, zi), is characterized by two key pieces of information: its center-point spatial coordinates and a representative wind velocity vector, Vw,i(Vx,i, Vy,i, Vz,i). For cells containing several CFD mesh nodes, the wind vector is averaged from all nodes within that cell. This structured grid provides a complete and quantifiable representation of the environment, allowing the path planner to query the wind conditions at any potential location. The grid serves as the foundational map for the improved A* algorithm, enabling it to compute wind-influenced costs for every potential move.
The complete process of wind field construction and gridding can be summarized by the following steps:
- Acquire building geometry from GIS data.
- Build a 3D solid model of the urban area.
- Define boundary conditions from wind rose data (e.g., dominant wind direction, speed).
- Run CFD simulation to obtain detailed wind vectors.
- Define grid cell size Lgrid.
- Partition the 3D airspace into K x M x N cubic cells.
- For each cell, average the CFD wind vectors to get a single representative vector Vw,i.
4. An Improved Path Planning Algorithm with Wind Field Awareness
4.1 Path Planning Model and Cost Function
The primary objective of our path planning model is to minimize the total flight time for a UAV drone from start to finish, while strictly adhering to a set of safety and operational constraints. The total flight time, T, is the sum of the time required to traverse each segment of the path:
$$ \min T = \sum_{i=1}^{n} \frac{dist(C_{i-1}, C_i)}{v_{g,i-1}} $$
where dist(Ci-1, Ci) is the Euclidean distance between two consecutive grid cells, and vg,i-1 is the ground speed of the UAV drone along that segment. The ground speed is a function of the UAV drone’s airspeed (Va) and the local wind velocity (Vw). Assuming a constant airspeed magnitude, the ground speed vector is the vector sum of the two: Vg = Va + Vw. The magnitude of the ground speed, vg, is derived from the vector geometry, which can be expressed as:
$$ v_g = v_a \cos\alpha + v_w \cos\beta $$
where α is the angle between the airspeed vector and the ground speed vector, and β is the angle between the wind vector and the ground speed vector. This model assumes the UAV drone’s heading is adjusted to maintain a course directly toward the next waypoint.
The optimization process is subject to several critical constraints:
- No-fly Zone Constraint: The path must not intersect with any grid cells classified as hazards. This includes building cells, a 4-meter safety buffer around buildings, and cells where the wind speed exceeds the UAV drone’s maximum operational wind speed (e.g., 12 m/s).
- Node Connectivity Constraint: Successive waypoints must be neighboring cells in the grid topology.
- Flight Altitude Constraint: The z-coordinate of any waypoint must lie within the designated altitude band.
4.2 The Improved A* Algorithm
The classic A* algorithm searches for an optimal path by minimizing a cost function f(n) = g(n) + h(n), where g(n) is the actual cost from the start node to node n, and h(n) is a heuristic estimate of the cost from node n to the goal. Our improvement enhances each of these components to be wind-aware and structurally efficient.
1. Dynamic Step Size: Instead of a fixed neighborhood search, we implement a dynamic step size, Si. This step size is inversely proportional to the distance to the nearest obstacle or hazard and proportional to the distance to the goal. This allows the algorithm to take larger, more efficient steps in open, safe areas and smaller, cautious steps near obstacles.
$$ S_i = \min(k \times S_0 + \mu, S_{min}) $$
Here, S0 is the base step size, Smin is the distance to the nearest hazard, and k is a coefficient representing the ratio of the distance from the current node to the goal to the total distance from the start to the goal.
2. Line-of-Sight Reachability: When expanding a node, the algorithm first checks if a line-of-sight exists between the current node’s grandparent and the potential new child node. If the straight path is free of obstacles and hazards, the grandparent node is directly connected to the child node, bypassing the parent. This reduces the number of intermediate waypoints and smooths the path.
3. Wind-Field-Aware Cost Functions: The core of the adaptivity lies in the cost functions, which are modified to represent time rather than distance.
$$ f(n) = g(n) + h(n) $$
$$ g(n) = \sum_{i=1}^{n} \frac{dist(C_{i-1}, C_i)}{v_{g,i-1}} $$
$$ h(n) = \sum_{k=n}^{goal} \frac{dist(C_{k-1}, C_k)}{v_{g,exp}} $$
In these equations, vg,i-1 is the calculated ground speed for the actual segment from the start, incorporating the local wind field. The heuristic h(n) uses an expected ground speed, vg,exp, to provide an optimistic estimate of the remaining time to the goal. This heuristic remains admissible and consistent, guaranteeing optimality.
The computational procedure of our improved A* algorithm is outlined below:
- Initialize: Create the start node with its coordinates, a zero initial time cost, and an estimated time to the goal. Initialize an open list (for nodes to be explored) and a closed list (for explored nodes). Load the gridded wind field map.
- Main Loop: While the open list is not empty, extract the node with the lowest f(n) value. If this node is the goal, reconstruct the path by backtracking through parent pointers.
- Node Expansion: For the extracted node, generate its neighbors using the dynamic step size mechanism. For each neighbor, if it exists among the grandparent and is reachable via a straight line, link directly to the grandparent.
- Cost Calculation & Update: For each feasible neighbor, calculate the wind-informed time cost to traverse the segment. If the neighbor is not yet in the open list, add it. If it is already in the open list but a better (shorter) path to it has been found, update its cost and parent.
- Return Path: Once the goal is reached, trace back the parent links to generate the complete, optimal flight path.
5. Experiments and Results Analysis
5.1 Experimental Setup
The experimental validation was conducted in a typical high-density urban district in Nanshan, Shenzhen, covering an area of 1.0 km by 0.9 km. The building density is 62%, with a maximum building height of 239 meters, creating a strongly heterogeneous wind field. The 3D building model was constructed from high-resolution satellite imagery. CFD simulations were run three times using the dominant north-northeast wind direction at speeds of 4 m/s, 6 m/s, and 8 m/s. For the path planning, a grid size of Lgrid = 2 m was used. The DJI M350 UAV drone was used as the reference platform with an assumed airspeed of 12 m/s and a maximum wind tolerance of 12 m/s. Three different start-to-goal waypoint pairs (Route 1, Route 2, Route 3) were designed to test the algorithm’s performance. We compared our improved algorithm against two baselines: (1) a traditional A* algorithm using Euclidean distance as cost (no wind awareness), and (2) a baseline wind-aware A* algorithm (which can avoid hazards) but without the dynamic step size or line-of-sight features.
| Parameter | Value |
|---|---|
| UAV Drone Airspeed | 12 m/s |
| UAV Drone Max. Wind Tolerance | 12 m/s |
| Minimum Flight Altitude | 80 m |
| Maximum Flight Altitude | 150 m |
| Grid Cell Size | 2 m |
| Building Safety Buffer | 4 m |
| UAV Drone Maximum Flight Time | 55 min |
5.2 Comparative Results
Experiment 1: Comparison with Traditional A* (No Wind Awareness)
As shown in Table 1, the traditional algorithm, which ignores the wind field, produced paths that penetrate hazardous wind zones, presenting serious safety risks. Our improved algorithm successfully avoided all such zones. Compared to the traditional method, our algorithm reduced the total flight time by over 42% and reduced the number of path waypoints (turns) by 48.5% for the test routes. This demonstrates a significant improvement in both safety and efficiency for navigating UAV drones.
| Algorithm | Route | Flight Distance (m) | Flight Time (s) | No. of Turns | Distance in Hazard (m) |
|---|---|---|---|---|---|
| Improved A* | Route 1 | 668.13 | 42.27 | 7 | 0 |
| Route 2 | 237.46 | 12.19 | 8 | 0 | |
| Route 3 | 406.85 | 20.52 | 32 | 0 | |
| Traditional A* (Euclidean) | Route 1 | 828.00 | 69.00 | 44 | 48.00 |
| Route 2 | 264.00 | 22.00 | 12 | 52.00 | |
| Route 3 | 464.00 | 38.67 | 35 | 28.00 |
Experiment 2: Comparison with Baseline Wind-Aware A*
This experiment compared our fully improved algorithm (with dynamic step and line-of-sight) against a baseline wind-aware algorithm that could avoid hazards but lacked these optimizations. The results are in Table 2. Both algorithms ensured safety by avoiding hazard zones. However, our improved algorithm demonstrated clear performance advantages.
| Algorithm | Route | Flight Distance (m) | Flight Time (s) | No. of Turns | Distance in Hazard (m) |
|---|---|---|---|---|---|
| Improved A* | Route 1 | 668.13 | 42.27 | 7 | 0 |
| Route 2 | 237.46 | 12.19 | 8 | 0 | |
| Route 3 | 406.85 | 20.52 | 32 | 0 | |
| Baseline Wind-Aware A* | Route 1 | 826.80 | 42.82 | 50 | 0 |
| Route 2 | 266.70 | 14.11 | 18 | 0 | |
| Route 3 | 1017.85 | 41.19 | 44 | 0 |
Our improved algorithm reduced the average flight time by 23.6% and, more significantly, reduced the average number of path waypoints by 58.2%. This substantial reduction in waypoints leads to a much smoother trajectory, which is easier for the flight controller of a UAV drone to execute and results in more stable flight. The dynamic step size allowed the algorithm to take longer paths in open areas, while the line-of-sight feature removed unnecessary intermediate nodes, directly contributing to this smoothness and efficiency.
Experiment 3: Multi-Scenario and Non-Stationary Wind Tests
To test the generalizability of our algorithm for UAV drones, we divided the study area into left and right sub-scenes with different building layouts. Our improved algorithm successfully planned collision-free, hazard-free paths in both scenes with similarly high performance, proving its robustness across varied urban morphologies.
Finally, we tested the algorithm under a non-stationary wind condition where the wind speed changed from 8 m/s to 6 m/s and then to 4 m/s along the flight path. Our algorithm, operating without interruption, automatically detected the change in the wind field data at the boundary nodes and triggered a replanning procedure. It smoothly and safely adapted its path to the new wind conditions, confirming its capability for real-time operation in dynamic environments.
6. Conclusion
This study has successfully developed and validated a novel path optimization framework for the navigation of UAV drones in complex urban wind fields. The key findings can be summarized as follows:
- Integrated Wind Field Modeling: The fusion of GIS data with CFD simulations provides a highly accurate and quantifiable representation of the three-dimensional wind environment. The subsequent gridding of this continuous field into a structured grid is essential for enabling direct, cell-level evaluation of wind impact by path planning algorithms.
- Enhanced A* Algorithm Performance: The proposed improvements to the A* algorithm—specifically the integration of a dynamic step size, a line-of-sight reachability check, and a wind-field-aware time-based cost function—collectively achieve a synergistic optimization of safety, efficiency, and path quality. Compared to a standard wind-aware algorithm, our method achieves a 23.6% reduction in average flight time and a 58.2% reduction in path waypoints for UAV drones, leading to smoother and more easily executable trajectories.
- Robustness and Adaptability: The algorithm demonstrates outstanding robustness across different urban layouts (multi-scenario tests) and exceptional adaptability to dynamically changing wind conditions. Its ability to trigger automatic replanning in non-stationary wind fields without mission interruption underscores its potential for real-world operations of UAV drones in urban environments.
In conclusion, our research provides a practical and effective solution for the safe and efficient low-altitude flight of UAV drones. Future work will focus on extending this framework to multi-drone cooperative path planning and integrating real-time sensor data for dynamic obstacle avoidance, further enhancing the autonomy and safety of urban UAV drone operations.