This page contains the data to train and evaluate WindSeer predicting the steady state wind around complex terrain.

We generated flow data over real terrain patches with a pipeline based on the open source solver OpenFOAM with the RANS model and the popular k − ε two-equation turbulence closure. The automated pipeline ingests terrain patches and outputs the time-averaged flow solutions for multiple wind speeds. We extracted 563 terrain patches each with an extent of 1.5 km × 1.5 km from the GeoVite service, which provides access to the swissALTI3D digital elevation map for Swiss researchers, with a lateral resolution of 0.5 m. The terrain patches exhibit at least one side with near-constant elevation allowing us to simulate a formed boundary layer flow (logarithmic profile) entering into the domain from that face. Some terrains allowed for multiple flow directions leading to 866 terrain/flow direction pairs. The vertical extent of the simulation domain was three times the height difference of the terrain with a lower bound of 1100 m minimizing the boundary effects on the flow. Each case was simulated with up to 15 different wind speeds if the automatic meshing succeeded, resulting in 7361 executed CFD runs. Only solutions that met a required optimization tolerance were accepted as fully converged solutions, which was the case in 92.9 % of the runs. We enhanced our dataset with one zero-velocity flow for each terrain that had at least one converged CFD simulation, resulting in a total of 7285 flows.

The CFD solutions are computed on an automatically generated irregular mesh with OpenFOAM’s SnappyHexMesh utility. We resampled each case up to a height of 1100 m to a regular 91 × 91 × 96 grid resulting in a resolution of 16.5 m horizontally and 11.5 m vertically.

The code is available at https://github.com/ethz-asl/WindSeer and all files are structured such they can be read by the corresponding script in the repository.

The train, validation, and test set contain the CFD-simulated data to train and evaluate WindSeer. The dataset are split across terrain to ensure that WindSeer predicts well on previously unseen topographies.

We also evaluate WindSeer on real wind data either collected from static wind sensors as part of large scale measurement campaigns, such as the Askervein, Bolund, and Perdigao campaigns, or onboard wind estimates from small UAVs. We also provide these datasets already converted to the HDF5 format and preprocessed such that they can directly be used by the scripts in the repository.

Train – 127 GB – Dataset used to train WindSeer

Validation – 52 GB – Dataset used to evaluate WindSeer during the training.

Test – 19 GB – Dataset used to evaluate WindSeer on the CFD data.

Askervein – 570 MB – Converted dataset of the Askervein measurement campaign

Bolund – 581 MB – Converted dataset of the Bolund measurement campaign

Perdigao - 2017-05-09 – 1.1 GB – Converted dataset of the Perdigao measurement campaign for 2017-05-09

Flight data – 30 MB – Postprocessed wind and position estimates from all data collection flights.

Models – 2.1 GB – Trained models.

Video – 389 MB – High-quality version of the paper video.