Recovery of UAV Swarm-Enabled Collaborative Beamforming in Low-Altitude Wireless Networks Under Wind Field Disturbances

Geng Sun, Chenbang Liu, Jiahui Li, Guannan Qu, Shuang Liang, Jiacheng Wang, Changyuan Zhao, Keyi Chen, Dusit Niyato

TL;DR

This paper introduces a real-time optimization framework to rescue collaborative beamforming performance in UAV swarms when wind disturbances knock the drones out of position. By deploying a novel deep reinforcement learning algorithm called PPO-LA (Proximal Policy Optimization with LSTM and Adaptive learning rate), the system dynamically adjusts the antenna excitation weights of the UAVs to maintain high signal directivity and suppress interference without requiring energy-intensive physical repositioning.

What Problem Does This Paper Address?

Deploying a swarm of Unmanned Aerial Vehicles (UAVs) to act as a Virtual Antenna Array (VAA) is a powerful way to extend communication range and boost energy efficiency in low-altitude wireless networks. Through collaborative beamforming (CB), individual drones coordinate their transmissions to focus signal power directly toward a ground base station.

However, low-altitude airspace is plagued by unpredictable wind fields—specifically constant wind, wind shear, and atmospheric turbulence. These wind forces inevitably push UAVs away from their ideal 3D coordinates. Because beamforming relies on precise phase synchronization, even minor positional displacements distort the beam pattern. This results in a drop in main lobe directivity and the rise of unwanted "sidelobes," which cause signal interference and security vulnerabilities.

Traditional error-correction methods either assume static environments or require extensive training data for specific wind profiles, making them brittle when faced with rapidly changing, real-world wind dynamics.

System overview showing a UAV swarm forming a virtual antenna array to communicate with a ground base station under wind disturbances.

Methods

Instead of forcing the UAVs to expend valuable battery power physically fighting the wind to stay in a perfect geometric grid, the authors propose a software-based compensation strategy. They formulate a long-term, real-time optimization problem that maximizes beam directivity ($D$) and minimizes the maximum sidelobe level ($M$) purely by adjusting the excitation current weights (amplitude and phase) of each UAV's antenna.

To solve this highly non-convex and stochastic problem, they developed PPO-LA, an enhanced version of the Proximal Policy Optimization algorithm featuring two key innovations:

  1. LSTM Integration: Standard reinforcement learning algorithms operate on the Markov assumption—that the next state depends only on the current state. However, wind disturbances possess temporal patterns. By embedding Long Short-Term Memory (LSTM) layers into the policy network, the agent can capture historical wind trends to predict and preemptively compensate for positional drift.
  2. Adaptive Learning Rate: The highly stochastic nature of wind fields creates volatile gradient landscapes during training. Integrating an Adam-based adaptive learning rate allows the algorithm to automatically scale its step sizes based on the local geometry of the loss surface, ensuring stable and rapid convergence.

The authors also established a rigorous mathematical framework modeling three distinct wind conditions: constant wind velocity, linear wind shear (speed variations relative to height), and turbulent wind (stochastic fluctuations modeled via the Dryden turbulence model).

Datasets

The framework was evaluated using simulated 3D environments rather than real-world datasets. The authors constructed two distinct synthetic benchmarks to evaluate algorithmic scalability:

  • A small-scale swarm consisting of 8 UAVs
  • A large-scale swarm consisting of 16 UAVs

Both setups operated within a $50 \times 50 \times 50 \text{ m}^3$ airspace with simulated wind profiles generated via standard aerodynamic and turbulence equations.

Key Insights and Ideas

  • Software Compensation > Hardware Correction: Compensating for phase errors by tweaking electromagnetic wave parameters in software is far more energy-efficient for a swarm than constantly firing drone rotors to maintain rigid physical formations in a windstorm.
  • Exploiting Temporal Momentum: Wind isn't entirely random; it has fluid momentum. Breaking the standard Markovian constraints of DRL by adding memory (LSTM) allows the network to "anticipate" where the drone will be pushed next, enabling proactive parameter tuning.
  • The Scalability Penalty: As the swarm scales from 8 to 16 nodes, the maximum achievable directivity jumps significantly (from ~8.25 dB to 15 dB). However, the fluctuation range of the beam pattern also expands. This points to an "error accumulation effect," where individual positioning errors compound exponentially through array factor multiplication in larger swarms.

Key Results

In comprehensive simulations, the proposed PPO-LA algorithm successfully recovered degraded beamforming performance and consistently outperformed standard benchmark algorithms (DDPG, DQN, SAC, and standard PPO).

  • Convergence and Stability: PPO-LA demonstrated faster convergence speeds and achieved higher, more stable cumulative rewards than standard PPO, which suffered from high performance volatility due to environmental noise.
  • Directivity and Sidelobe Control: In the 8-UAV setup, PPO-LA maintained a stable directivity value of $8.25 \pm 0.15 \text{ dB}$ while keeping sidelobes suppressed. In the 16-UAV setup, it successfully scaled to deliver a stabilized directivity of $15 \text{ dB}$ and a reduced maximum sidelobe level of $-3.5 \text{ dB}$.
  • Turbulence Resilience: The algorithm showed its most dramatic comparative advantage in turbulent wind scenarios—the most chaotic environment tested—proving that the LSTM architecture successfully modeled complex spatio-temporal displacement patterns where static algorithms failed.

Limitations and Future Work

While the simulation results are promising, the authors acknowledge certain limitations that pave the way for future research:

  • Purely Simulated Validation: The framework has yet to be validated on physical hardware, where real-world constraints like imperfect Channel State Information (CSI) estimation and antenna hardware impairments would introduce additional noise.
  • Static Trajectories: The current model adjusts antenna weights for UAVs hovering around fixed nominal positions. Future work will aim to explore joint trajectory design and beamforming, allowing the drones to optimize both their physical flight paths and their electrical transmission parameters simultaneously.

Summary updated

Abstract

Unmanned aerial vehicle (UAV) swarms utilizing collaborative beamforming (CB) in low-altitude wireless networks (LAWN) demonstrate significant potential for enhanced communication range, energy efficiency, and signal directivity through the formation of virtual antenna arrays (VAA). However, environmental disturbances, particularly wind fields, significantly degrade CB performance by introducing positional errors that disrupt beam patterns, thereby compromising transmission reliability. This paper investigates the critical challenge of maintaining CB performance in UAV-based VAAs operating in LAWN under wind field disturbances. We propose a comprehensive framework that models the impact of three distinct wind conditions (constant, shear, and turbulent) on UAV array performance, and formulate a long-term real-time optimization problem to maximize directivity while minimizing maximum sidelobe levels through adaptive excitation current weight adjustments. To address the inherent complexity of this problem, we propose a novel proximal policy optimization algorithm with long short-term memory (LSTM) structure and adaptive learning rate (PPO-LA), which effectively captures temporal patterns in wind field disturbances and enables real-time adaptation without requiring extensive prior training for specific wind conditions. Our simulation results demonstrate that the proposed PPO-LA algorithm successfully recovers degraded CB performance across various wind scenarios, and thus significantly outperforming benchmark algorithms.
BibTeX
@article{sun2026recovery,
  author = {Geng Sun and Chenbang Liu and Jiahui Li and Guannan Qu and Shuang Liang and Jiacheng Wang and Changyuan Zhao and Keyi Chen and Dusit Niyato},
  title = {{Recovery of UAV Swarm-Enabled Collaborative Beamforming in Low-Altitude Wireless Networks Under Wind Field Disturbances}},
  journal = {IEEE Trans. Vehicular Technology},
  year = {2026},
  doi = {10.1109/tvt.2026.3654588},
}

Loading related content...

Figure