In a paper published in the journal Nature Communications, researchers unveiled WindSeer, a novel approach to real-time, high-resolution wind predictions essential for aviation safety and other fields.
Overview of the wind prediction pipeline. A First we generate labelled flows utilizing a CFD simulation. B Then WindSeer is trained with measurements along randomly sampled piecewise linear trajectories to predict the dense flow. C During deployment the wind estimates from the UAV or wind measurement towers together with the known topography serve as the input to WindSeer. Image Credit: https://www.nature.com/articles/s41467-024-47778-4
Introducing the deep neural network-based model WindSeer addressed the limitations of current weather models, enabling accurate predictions even with sparse data. Its adaptability and ability to predict wind fields over diverse terrains without retraining mark a significant advancement in wind prediction technology, promising safer and more efficient operations across various sectors.
Background
Past work has explored various methods for wind prediction, from numerical weather prediction (NWP) models to computational fluid dynamics (CFD) simulations and data assimilation techniques. However, these approaches often need help with high computational costs or reliance on extensive observation data.
While artificial intelligence (AI)-based methods have shown promise in accelerating wind flow computations, they typically require privileged information or dense measurement coverage. Additionally, traditional methods need help to capture the complex dynamics of wind flow around steep terrain features, limiting their applicability in real-world scenarios.
WindSeer Development Overview
The pipeline was developed to train and deploy WindSeer, a convolutional neural network (CNN) architecture to predict dense time-averaged wind and turbulence around complex terrain. The training commenced with generating a dataset featuring dense flows over Swiss terrain patches using a Reynolds-averaged Navier–Stokes (RANS) computational fluid dynamics (CFD) solver.
Subsequently, WindSeer was trained using these labeled flows to simulate local wind measurements along randomly generated trajectories, incorporating added noise. Trained WindSeer was then evaluated on held-back CFD-simulated flows, real wind data from measurement campaigns, and wind data measured by multiple sUAVs around mountainous terrain.
The team utilized the open-source solver OpenFOAM, employing a steady-state RANS model and the k − ϵ two-equation turbulence closure for the CFD wind data generation. Terrain patches were extracted from the GeoVite service, yielding 563 patches, each with multiple wind speed flow solutions. The team computed CFD solutions with irregular meshes resampled to regular grids for consistent input to WindSeer.
Geometric transformations were applied to the CFD flows to augment the WindSeer training data, enhancing the dataset's quality and size. Data augmentation aimed to prevent overfitting and improve model generalization. The input to WindSeer comprised four volumetric channels, including terrain encoding and sparse, noisy wind measurements, while the labels consisted of three-dimensional predicted velocity and turbulence kinetic energy (TKE) from the CFD ground truth flows.
Model training involved an encoder-decoder CNN architecture with skip connections based on the U-Net architecture. The loss function applied during training was a scaled version of the mean squared error (MSE) loss, which balanced the loss between different samples and channels. The model was trained using the Adam optimizer for 3000 epochs, with the learning rate quartered every 700 epochs.
The analysts conducted inference time experiments on an Orin AGX single-board computer and performed small unmanned aerial vehicle (sUAV) flight tests at three test sites in Switzerland. WindSeer predictions were compared to wind data from measurement campaigns and sUAV flights, assessing prediction accuracy under various real-world conditions. Additionally, a range of error metrics, including relative error and speedup error, were defined to evaluate WindSeer's performance comprehensively. All processed CFD data and real campaign/sUAV flight measurements are publicly available, facilitating further research and validation.
WindSeer Validation Overview
WindSeer, a CNN, underwent rigorous evaluation across various experiments to validate its real-time wind prediction capabilities. In the first experiment group, WindSeer's performance was assessed using held-out CFD-simulated flows, demonstrating its ability to capture complex flow patterns around diverse terrains accurately.
The model exhibited strong predictive capabilities, particularly in regions where operational altitudes for sUAVs and wind turbine hub heights are typically situated, despite challenges such as confounding factors near the ground.
In the second experiment group, WindSeer was evaluated on real-world wind data from measurement campaigns conducted in different terrains. In most cases, the model outperformed an averaging baseline, demonstrating its effectiveness in predicting wind magnitude, vertical wind, and TKE. WindSeer consistently performed across various terrains, providing valuable insights for safer and more efficient sUAV trajectory planning.
The performance was tested in the third experiment group using noisy wind measurements from multiple fixed-wing sUAVs, mimicking real-world scenarios. Despite high input noise levels, WindSeer exhibited superior predictive capabilities, particularly in predicting variations in vertical wind. The model's ability to provide accurate predictions in dynamic and noisy environments highlights its potential for practical deployment in sUAV operations. WindSeer also demonstrated low-latency inference times on flight-grade hardware, confirming its suitability for real-time wind prediction applications.
Conclusion
To sum up, the comprehensive evaluation of WindSeer underscored its robustness and efficacy in real-time wind prediction across diverse scenarios. From CFD-simulated flows to real-world measurement campaigns and sUAV flight tests, WindSeer consistently demonstrated its ability to predict wind dynamics accurately.
Its performance surpassed baseline methods, showcasing its potential for enhancing safety and efficiency in sUAV operations and wind energy applications. With low-latency inference times on flight-grade hardware, WindSeer emerged as a promising solution for practical deployment in dynamic environments, paving the way for wind monitoring and forecasting technology advancements.
Using GPT-4o for Web Archiving