PeakWeather 🌦️🌤️⛈️ Forecasting Use Cases

This repository provides code to replicate the experiments presented in the paper

PeakWeather: MeteoSwiss Weather Station Measurements for Spatiotemporal Deep Learning,
Daniele Zambon¹, Michele Cattaneo²³, Ivan Marisca¹, Jonas Bhend², Daniele Nerini², Cesare Alippi¹⁴.
¹ USI, IDSIA (Lugano, Switzerland), ² MeteoSwiss (Zurich, Switzerland), ³ ETH Zürich and EPFL, SDSC (Zurich, Switzerland), ⁴ PoliMi (Milan, Italy).

The paper presents PeakWeather, a new high-resolution spatiotemporal dataset of validated ground-based meteorological observations from Switzerland. The dataset includes:

• Meteorological observations from 302 stations across Switzerland, every 10 minutes from January 2017 to October 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. PeakWeather is associated with a Python library to download, load, and prepare the data.

Getting Started

The code has been tested on Debian GNU/Linux 11 with Python 3.10.

Requirements

To resolve all dependencies, we recommend using Anaconda and the provided environment configuration by running the following commands:

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 it afterwards with pip.

Model training and testing

To train and evaluate models, run:

# wind forecasting
python -m experiments.run_wind_prediction dataset=wind model=tcn 
python -m experiments.test_chronos_wind 
# temperature forecasting
python -m experiments.run_temp_prediction dataset=temp nwp_test_set=True model=mptcn_glob 
python -m experiments.test_chronos_temp 

The config directory stores all the configuration files used to run the experiments using Hydra. Available models are listed in config/model/ folder.

The code has been tested in a single-GPU setting.

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}, 
}