TLDR: G-WindCast is a new deep learning framework that uses signal strength variations from 5G GNSS signals to retrieve and forecast three-dimensional atmospheric wind fields in real-time. It demonstrates high accuracy and robustness, outperforming ERA5 reanalysis data in some scenarios, and is scalable, requiring as few as 100 stations for effective localized forecasting. This approach offers a cost-effective and high-resolution solution for weather forecasting, aviation safety, and disaster risk reduction.
Accurate information about atmospheric wind fields is vital for many areas, including weather forecasting, aviation safety, and reducing disaster risks. However, getting high-resolution wind data has always been difficult due to limitations in traditional observation methods, remote sensing techniques, and the high computational cost and biases of numerical weather prediction (NWP) models.
A new deep learning framework called G-WindCast has been introduced to tackle these challenges. This innovative system uses variations in signal strength from 5G Global Navigation Satellite System (GNSS) signals to both determine and predict three-dimensional (3D) atmospheric wind fields. The framework employs advanced deep learning models, specifically Forward Neural Networks (FNN) and Transformer networks, to identify complex, non-linear, and spatiotemporal connections between GNSS-derived features and how winds behave.
The core idea behind G-WindCast is that GNSS signals are affected by the atmosphere as they travel. These changes, usually seen as errors in positioning, can actually provide valuable clues about the atmospheric state, including wind variations. By collecting and processing signal intensity data from dense ground-based GNSS receiver networks, G-WindCast extracts these atmospheric indicators. Deep learning then helps the system learn from historical data to infer current wind conditions and forecast future wind changes.
Preliminary results from G-WindCast are very promising. The system shows good accuracy in both retrieving current wind fields and forecasting short-term winds, up to a 30-minute lead time. Its performance metrics are comparable to high-resolution NWP outputs in certain situations. The model also proves to be robust across different forecast durations and atmospheric pressure levels. Its predictions for wind speed and direction align better with actual observations compared to concurrent ERA5 reanalysis data, which is a widely used dataset.
One of the significant advantages of G-WindCast is its cost-effectiveness and scalability. The research demonstrates that the system can maintain excellent performance for localized forecasting, even with a significantly reduced number of GNSS stations. For example, using around 100 stations, which is about one-twelfth of the total stations used in some experiments, was sufficient for strong wind speed predictions 30 minutes ahead, making it highly practical for applications like airport local forecasting.
This interdisciplinary approach highlights the transformative potential of using non-traditional data sources and deep learning for advanced environmental monitoring and real-time atmospheric applications. While the study used a limited dataset in terms of temporal and spatial coverage, future work aims to explore larger datasets and integrate multiple meteorological observations to further enhance prediction performance.
Also Read:
- AI Accelerates Discovery of Atomic Fine Structure
- Navigating Unpredictable Roads: An Analysis of AI’s Reliability in Traffic Control
For more detailed information, you can refer to the full research paper available at Communications to Circulations: 3D Wind Field Retrieval and Real-Time Prediction Using 5G GNSS Signals and Deep Learning.
