Introduction
As the Radio Frequency (RF) spectrum becomes increasingly congested, Visible Light Communication (VLC) has emerged as a high-bandwidth, energy-efficient alternative. Deploying VLC via Unmanned Aerial Vehicles (UAVs) offers flexibility, but introduces a unique challenge: ambient light from city infrastructure. While this light helps meet illumination requirements for ground users, it simultaneously acts as significant interference for the VLC data link. This paper proposes a novel framework that uses Gated Recurrent Units (GRUs) to predict ambient light patterns, allowing for optimal UAV positioning that minimizes transmit power.
System Model and Methodology
The proposed system addresses the spatial-temporal dynamics of ambient light through a three-stage approach designed to optimize communication efficiency while maintaining required illumination levels.
- Spatial Modeling with GMM: To manage the complexity of urban light distributions, the researchers employ a Gaussian Mixture Model (GMM). This allows a complex light map to be represented by a small set of parameters (means, variances, and weights), significantly reducing the dimensionality of the input data for the prediction model.
- Temporal Prediction with GRU: A Gated Recurrent Unit (GRU) network is trained on historical light data from the NOAA/NCEI Earth Observations Group. The GRU forecasts how the GMM parameters will evolve, effectively predicting future "bright" and "dark" spots in the environment.
- Convex Optimization: With the predicted light distribution, the UAV deployment is formulated as a power-minimization problem. By proving the problem is convex, the authors use the dual method to allow each UAV to independently determine its optimal $(x, y)$ coordinates and transmit power.
Key Results and Performance
The framework was validated using real-world nighttime remote sensing data and simulated environments. A key finding is the "Bright Spot" strategy: counter-intuitively, results show that UAVs should hover over areas with strong ambient illumination. Although interference is higher in these regions, the proximity to users and the ease of meeting illumination requirements result in lower overall transmit power.
The results demonstrate significant efficiency gains over traditional deployment methods:
- Power Efficiency: The proposed GRU-based deployment achieves up to a 22.1% reduction in transmit power compared to baseline deployments that ignore ambient light distribution.
- Near-Optimal Accuracy: The performance gap between the GRU-predicted deployment and a theoretical "perfect knowledge" scenario (using actual future data) is less than 3.4%, proving the robustness of the prediction.
- Prediction Reliability: The GRU model reached high accuracy after 100 epochs, with an average Root Mean-Square Error (RMSE) of 3.65 on test data.
Conclusion and Future Work
By integrating deep learning-based prediction with convex optimization, this research provides a scalable solution for VLC-enabled UAV networks. While the current model focuses on communication power, the authors acknowledge that future iterations should incorporate UAV mobility energy consumption and 3D placement (varying altitude) to further enhance real-world applicability in dynamic urban environments.
Summary updated