PLEASE NOTE THAT THIS IS DEPRECATED IN FAVOUR OF OCF_DATAPIPES!
Pre-prepare batches of data for use in machine learning training.
This code combines several data sources including:
- Satellite imagery (EUMETSAT SEVIRI RSS 5-minutely data of UK)
- Numerical Weather Predictions (NWPs. UK Met Office UKV model from CEDA)
- Solar PV power timeseries data (from PVOutput.org, downloaded using our pvoutput Python code.)
- Estimated total solar PV generation for each of the ~350 "grid supply points" (GSPs) in Britain from Sheffield Solar's PV Live Regional API.
- Topographic data.
- The Sun's azimuth and angle.
This repo doesn't contain the ML models themselves. Please see this page for an overview of the Open Climate Fix solar PV nowcasting project, and how our code repositories fit together.
From within the cloned nowcasting_dataset directory:
conda env create -f environment.yml
conda activate nowcasting_dataset
pip install -e .A (probably older) version is also available through pip install nowcasting-dataset
If you want to also install PVLive then use pip install git+https://github.com/SheffieldSolar/PV_Live-API
A pre commit hook has been installed which makes black run with every commit. You need to install
black and pre-commit (these will be installed by conda or pip when installing
nowcasting_dataset) and run pre-commit install in this repo.
To test using the small amount of data stored in this repo: py.test -s
To output debug logs while running the tests then run py.test --log-cli-level=10
To test using the full dataset on Google Cloud, add the --use_cloud_data switch.
Test using a docker file and database
docker stop $(docker ps -a -q)
docker-compose -f test-docker-compose.yml build
docker-compose -f test-docker-compose.yml run dataset
Use Satip to download native EUMETSAT SEVIRI RSS data from EUMETSAT's API and then convert to an intermediate file format.
Download PV timeseries data from PVOutput.org using our PVOutput code.
PV solar generation data from the UK. This dataset contains data from 1311 PV systems from 2018-01-01 to 2021-10-27. The time series of solar generation is in 5 minutes chunks. This data is collected from live PV systems in the UK. We have obfuscated the location of the PV systems for privacy.
Please use our nwp code to download UKV NWPs and convert to Zarr.
TODO - GSP
- Make an account at the USGS EarthExplorer website
- Create a region of the world to download data for, in our case, the spatial extant of the SEVIRI RSS image
- Select the data products you want, in this case SRTM elevation maps
- Download all the SRTM files that cover that area
There does not seem to be an automated way to do this selecting and downloading, so this might take awhile.
Copy and modify one of the config yaml files in
nowcasting_dataset/config/.
Run scripts/prepare_ml_data.py --help
to learn how to run the prepare_ml_data.py script.
Please see the data_sources/<modality>/<modality>_model.py files
(where <modality> is one of {datetime, metadata, gsp, nwp, pv,
satellite, sun, topographic}) for documentation about the different
data fields in each example / batch.
When we first started writing nowcasting_dataset, our intention was
to load and align data from these three datasets on-the-fly during ML
training. But it just isn't quite fast enough to keep a modern GPU constantly fed
with data when loading multiple satellite channels and multiple NWP
parameters. So, now, this code is used to pre-prepare thousands of
batches, and save these batches to disk, each as a separate NetCDF
file. These files can then be loaded super-quickly at training time.
The end result is a 12x speedup in training.
Thanks goes to these wonderful people (emoji key):
 Jack Kelly 💻 |
 Jacob Bieker 💻 |
 Peter Dudfield 💻 |
 Flo 💻 |
 Rohan Nuttall 💻 |
 Nasser Benabderrazik 💻 |
 Shanmukh Chava 💻 |
 Rishi Kumar Ray 💻 |
 JanEbbing 💻 |
This project follows the all-contributors specification. Contributions of any kind welcome!
