PeakWeather 🌦️🌤️⛈️ a use case in wind forecasting

This repository provides the code for replicating the wind forecasting experiments on the recently released PeakWeather dataset which has been presented in paper

PeakWeather: MeteoSwiss Weather Station Measurements for Spatiotemporal Deep Learning,
Daniele Zambon¹, Michele Cattaneo², Ivan Marisca¹, Jonas Bhend², Daniele Nerini², Cesare Alippi¹³.
¹ USI, IDSIA (Lugano, Switzerland), ² MeteoSwiss (Zurich, Switzerland), ³ PoliMi (Milan, Italy).

PeakWeather is a high-resolution spatiotemporal dataset of validated ground-based meteorological observations from Switzerland, and is associated with a Python library to download, load, and prepare the data. It includes

• Meteorological observations from 302 stations across Switzerland, every 10 minutes from Jan 2017 to Mar 2025
• Topographic features extracted from a 50m DEM for each station
• Operational NWP forecasts (ICON-CH1-EPS) co-located with observation points

The dataset supports a variety of tasks, including spatiotemporal forecasting, virtual sensing, and graph structure learning. Here, wind forecasting is considered.

Getting Started

Requirements

To solve all dependencies, we recommend using Anaconda and the provided environment configuration by running the command:

conda env create -f conda_env.yml
conda activate peakweather-env

Tip

Please, note that Conda packages for pytorch_geometric are currently not available for M1/M2/M3 macs; see PyG documentation. To install it, you can comment out PyG-related entries in conda_env.yml and install with pip, e.g.,

pip install torch_geometric==2.4 torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.2+cpu.html

Model training and testing

To train and evaluate a model for wind forecasting, run

python -m experiments.run_wind_prediction dataset=wind_1d model=<MODEL_NAME> 

Available models are listed in experiments/run_wind_prediction.py script; examples include tts_imp (STGNN), pers_st (PM-st), and icon (ICON).

Logging and tracking

To track the training, we rely on MLflow. To run it locally, execute the command below and open a browser at http://127.0.0.1:<PORT>/:

mlflow ui --port <PORT>

To run the experiment using your hosted MLflow server, specify the tracking URI with the overwrite: ++mlflow_tracking_uri=<YOUR_URI> when starting the experiment.

After training, the script automatically evaluates the model's predictive performance on the test set using predefined metrics.

Configuration files

The config directory stores all the configuration files used to run the experiment using Hydra.

Citation

If you use this code in your work, please cite our paper:

@misc{zambon2025peakweather,
  title={PeakWeather: MeteoSwiss Weather Station Measurements for Spatiotemporal Deep Learning}, 
  author={Zambon, Daniele and Cattaneo, Michele and Marisca, Ivan and Bhend, Jonas and Nerini, Daniele and Alippi, Cesare},
  year={2025},
  eprint={2506.13652},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2506.13652}, 
}