WindSeer: Real-time volumetric wind prediction over complex terrain aboard a small UAV

TL;DR

WindSeer tackles real-time, high-resolution wind prediction over complex terrain using sparse, noisy measurements by training a CNN on CFD-generated flows and deploying it in a multi-resolution, terrain-aware framework. The method uses a four-channel input that encodes terrain and sparse wind observations to predict a four-channel output, including $W_x$, $W_y$, $W_z$, and $TKE$, without requiring privileged boundary conditions. It demonstrates zero-shot sim-to-real transfer across CFD test flows and real measurement campaigns, and validates real-time onboard inference on a Jetson Orin to support sUAV planning and safety. The work highlights the ability to predict detailed near-ground wind fields at meter-scale resolutions, enabling improved flight planning and wind-energy micrositing, while outlining limitations related to training-domain scale, temporal variability, and sensor calibration for future improvements.

Abstract

Real-time high-resolution wind predictions are beneficial for various applications including safe manned and unmanned aviation. Current weather models require too much compute and lack the necessary predictive capabilities as they are valid only at the scale of multiple kilometers and hours - much lower spatial and temporal resolutions than these applications require. Our work, for the first time, demonstrates the ability to predict low-altitude wind in real-time on limited-compute devices, from only sparse measurement data. We train a neural network, WindSeer, using only synthetic data from computational fluid dynamics simulations and show that it can successfully predict real wind fields over terrain with known topography from just a few noisy and spatially clustered wind measurements. WindSeer can generate accurate predictions at different resolutions and domain sizes on previously unseen topography without retraining. We demonstrate that the model successfully predicts historical wind data collected by weather stations and wind measured onboard drones.

Figures (17)

• Figure 1: 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.
• Figure 2: CFD experiment. A) Terrain and input wind measurements (red arrows) with their respective prediction error (B). High prediction errors can be observed close to the ground or on the lee side of the terrain. C) Wind and turbulence prediction performance on the CFD dataset over the full test set (blue) and with 2000 random trajectories for the three different terrains shown in A. While most of the terrains result in uni-modal error distributions (1,2), more complex ones can have a second mode for samples from a complex flow region, indicated by the red box in (3). D) Density scatter plots comparing the label and the predictions for each predicted property using the terrain and input pair presented in A) (2).
• Figure 3: Measurement campaigns experiment. The mast locations and elevation maps for the Bolund (A), Askervein (C), and Perdigão (E) campaigns. The tower positions are colored by the average prediction error when using that specific mast as the input to predict the wind. In the Askervein and Perdigão case some masts did not provide a valid measurement for that experiment. (B, D, F) show the wind directions for the different experiments for each terrain.
• Figure 4: Measurement campaign results. Measured wind compared to the predictions aggregated over all predictions for the Bolund (A, 32 predictions [4 experiments, 8 masts]), Askervein (B, 182 predictions [13 experiments, 14 masts]), and Perdigão (C, 9120 predictions [240 experiments, 38 masts]) campaigns. In D the evolution of the prediction error and correlation of the wind norm S for WindSeer (WS) using the 5 minute and 1 hour averaged data together with the measurements from the TSE04 tower as a reference are shown. We show the results aggregated over all the 38 predictions using the different masts as input and the scores using only the TNW11 and V01 tower data for the prediction.
• Figure 5: Measurement campaigns prediction line and sUAV predictions along the path. (A-C) Predictions and measurements along characteristic lines with a constant height for each experiment with the baseline averaging method (AVG) and WindSeer (WS). Three predictions using different input masts are shown for each model and experiment. The asterisk * indicates that no measurement was available for that respective mast at the queried height and the closest one was picked. The uncertainty of the measurements is displayed by the standard deviation of the raw high-rate data. (D) The predictions from EZG A along the flight paths from EZG A and B for the first Chasseral flight.