🚨🚨🚨 WeatherBench 2 has been released. It provides an updated and much improved benchmark including more comprehensive and more easily accessible datasets.🚨🚨🚨
If you are using this dataset please cite
Stephan Rasp, Peter D. Dueben, Sebastian Scher, Jonathan A. Weyn, Soukayna Mouatadid, and Nils Thuerey, 2020. WeatherBench: A benchmark dataset for data-driven weather forecasting. arXiv: https://arxiv.org/abs/2002.00469
This repository contains all the code for downloding and processing the data as well as code for the baseline models in the paper.
Note! The data has been changed from the original release. Here is a list of changes:
- New vertical levels. Used to be [1, 10, 100, 200, 300, 400, 500, 600, 700, 850, 1000], now is [50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000]. This is to be compatible with CMIP output. The new levels include all of the old ones with the exception of [1, 10].
- CMIP data. Regridded CMIP data of some variables was added. This is the historical simulation of the MPI-ESM-HR model.
If you have any questions about this dataset, please use the Github Issue feature on this page!
Model | Z500 RMSE (3 / 5 days) [m2/s2] | T850 RMSE (3 / 5 days) [K] | Notes | Reference |
---|---|---|---|---|
Operational IFS | 154 / 334 | 1.36 / 2.03 | ECWMF physical model (10 km) | Rasp et al. 2020 |
Rasp and Thuerey 2020 (direct/continuous) | 268 / 499 | 1.65 / 2.41 | Resnet with CMIP pretraining (5.625 deg) | Rasp and Thuerey 2020 |
IFS T63 | 268 / 463 | 1.85 / 2.52 | Lower resolution physical model (approx. 1.9 deg) | Rasp et al. 2020 |
Weyn et al. 2020 (iterative) | 373 / 611 | 1.98 / 2.87 | UNet with cube-sphere mapping (2 deg) | Weyn et al. 2020 |
Clare et al. 2021 (direct) | 375 / 627 | 2.11 / 2.91 | Stacked ResNets with probabilistic output (5.625 deg) | Clare et al. 2021 |
IFS T42 | 489 / 743 | 3.09 / 3.83 | Lower resolution physical model (approx. 2.8 deg) | Rasp et al. 2020 |
Weekly climatology | 816 | 3.50 | Climatology for each calendar week | Rasp et al. 2020 |
Persistence | 936 / 1033 | 4.23 / 4.56 | Rasp et al. 2020 | |
Climatology | 1075 | 5.51 | Rasp et al. 2020 |
You can follow the quickstart guide in this notebook or lauch it directly from Binder.
The data is hosted here with the following directory structure
.
|-- 1.40625deg
| |-- 10m_u_component_of_wind
| |-- 10m_v_component_of_wind
| |-- 2m_temperature
| |-- constants
| |-- geopotential
| |-- old
| | `-- temperature
| |-- potential_vorticity
| |-- relative_humidity
| |-- specific_humidity
| |-- temperature
| |-- toa_incident_solar_radiation
| |-- total_cloud_cover
| |-- total_precipitation
| |-- u_component_of_wind
| |-- v_component_of_wind
| `-- vorticity
|-- 2.8125deg
| |-- 10m_u_component_of_wind
| |-- 10m_v_component_of_wind
| |-- 2m_temperature
| |-- constants
| |-- geopotential
| |-- potential_vorticity
| |-- relative_humidity
| |-- specific_humidity
| |-- temperature
| |-- toa_incident_solar_radiation
| |-- total_cloud_cover
| |-- total_precipitation
| |-- u_component_of_wind
| |-- v_component_of_wind
| `-- vorticity
|-- 5.625deg
| |-- 10m_u_component_of_wind
| |-- 10m_v_component_of_wind
| |-- 2m_temperature
| |-- constants
| |-- geopotential
| |-- geopotential_500
| |-- potential_vorticity
| |-- relative_humidity
| |-- specific_humidity
| |-- temperature
| |-- temperature_850
| |-- toa_incident_solar_radiation
| |-- total_cloud_cover
| |-- total_precipitation
| |-- u_component_of_wind
| |-- v_component_of_wind
| `-- vorticity
|-- baselines
| `-- saved_models
|-- CMIP
| `-- MPI-ESM
| |-- 2.8125deg
| | |-- geopotential
| | |-- specific_humidity
| | |-- temperature
| | |-- u_component_of_wind
| | `-- v_component_of_wind
| `-- 5.625deg
| |-- geopotential
| |-- specific_humidity
| |-- temperature
| |-- u_component_of_wind
| `-- v_component_of_wind
|-- IFS_T42
| `-- raw
|-- IFS_T63
| `-- raw
`-- tigge
|-- 1.40625deg
| |-- geopotential_500
| `-- temperature_850
|-- 2.8125deg
| |-- geopotential_500
| `-- temperature_850
`-- 5.625deg
|-- 2m_temperature
|-- geopotential_500
|-- temperature_850
`-- total_precipitation
To start out download either the entire 5.625 degree data (175G) using
wget "https://dataserv.ub.tum.de/s/m1524895/download?path=%2F5.625deg&files=all_5.625deg.zip" -O all_5.625deg.zip
or simply the single level (500 hPa) geopotential data using
wget "https://dataserv.ub.tum.de/s/m1524895/download?path=%2F5.625deg%2Fgeopotential_500&files=geopotential_500_5.625deg.zip" -O geopotential_500_5.625deg.zip
and then unzip the files using unzip <file>.zip
. You can also use ftp
or rsync
to download the data. For instructions, follow the download link.
IMPORTANT: The format of the predictions file is a
NetCDF dataset with dimensions [init_time, lead_time, lat, lon]
. Consult the notebooks for examples. You are
stongly encouraged to format your predictions in the same way and then use the same evaluation functions to ensure
consistent evaluation.
The baselines are created using Jupyter notebooks in notebooks/
. In all notebooks, the forecasts are saved as a
NetCDF file in the predictions
directory of the dataset.
An example of how to load the data and train a CNN using Keras is given in notebooks/3-cnn-example.ipynb
. In
addition a command line script for training CNNs is provided in src/train_nn.py
. For the baseline CNNs in the
paper the config files are given in src/nn_configs/
. To reproduce the results in the paper run e.g. python -m src.train_nn -c src/nn_configs/fccnn_3d.yml
.
Evaluation and comparison of the different baselines in done in notebooks/4-evaluation.ipynb
. The scoring is done
using the functions in src/score.py
. The RMSE values for the baseline models are also saved in the predictions
directory of the dataset. This is useful for plotting your own models alongside the baselines.
The dataset already contains the most important processed data. If you would like to download a different variable , regrid to a different resolution or extract single levels from the 3D files, here is how to do that!
The workflow to get to the processed data that ended up in the data repository above is:
- Download monthly files from the ERA5 archive (
src/download.py
) - Regrid the raw data to the required resolutions (
src/regrid.py
)
The raw data is from the ERA5 reanalysis archive. Information on how to download the data can be found here and here.
Because downloading the data can take a long time (several weeks), the workflow is encoded using Snakemake. See Snakefile
and the configuration files for each variable in scripts/config_ {variable}.yml
. These
files can be modified if additional variables are required. To execute Snakemake for a particular variable type
: snakemake -p -j 4 all --configfile scripts/config_toa_incident_solar_radiation.yml
.
In addition to the time-dependent fields, the constant fields were downloaded and processed using scripts /download_and_regrid_constants.sh
To obtain the operational IFS baseline, we use the TIGGE Archive. Downloading the data for Z500 and T850 is done in scripts/download_tigge.py
; regridding is done in scripts /convert_and_regrid_tigge.sh
.
The T21 baseline was created by Peter Dueben. The raw output can be found in the dataset. To regrid the data scripts /convert_and_regrid_IFS_TXX.sh
was used.
To download historical climate model data use the Snakemake file in snakemake_configs_CMIP
. Here, we downloaded data from the MIP-ESM-HR
model. To download other models, search for the download links on the CMIP website and modify the scripts accordingly.
If you would like to extract a single level from 3D data, e.g. 850 hPa temperature, you can use src /extract_level.py
. This could be useful to reduce the amount of data that needs to be loaded into RAM. An example
usage would be: python extract_level.py --input_fns DATADIR/5.625deg/temperature/*.nc --output_dir OUTDIR --level 850