PVNet 3.0.37

PVNet

PVNet 2.1

This project is used for training PVNet and running PVnet on live data.

PVNet2 is a multi-modal late-fusion model that largely inherits the same architecture from PVNet1.0. The NWP and satellite data are sent through some neural network which encodes them down to 1D intermediate representations. These are concatenated together with the GSP output history, the calculated solar coordinates (azimuth and elevation) and the GSP ID which has been put through an embedding layer. This 1D concatenated feature vector is put through an output network which outputs predictions of the future GSP yield. National forecasts are made by adding all the GSP forecasts together.

Experiments

Link to our workshop paper coming soon

Some very rough working notes on this model are here.

Some slightly more structured notes on deliberate experiments we have performed with PVNet are here

Setup / Installation

git clone https://github.com/openclimatefix/PVNet.git
cd PVNet
pip install .

The commit history is extensive. To save download time, use a depth of 1:

git clone --depth 1 https://github.com/openclimatefix/PVNet.git

This means only the latest commit and its associated files will be downloaded.

Next, in the PVNet repo, install PVNet as an editable package:

pip install -e .

Additional development dependencies

pip install ".[dev]"

Getting started with running PVNet

Before running any code in within PVNet, copy the example configuration to a configs directory:

cp -r configs.example configs

You will be making local amendments to these configs. See the README in configs.example for more info.

Datasets

As a minimum, in order to create batches of data/run PVNet, you will need to supply paths to NWP and GSP data. PV data can also be used. We list some suggested locations for downloading such datasets below:

GSP (Grid Supply Point) - Regional PV generation data
The University of Sheffield provides API access to download this data: https://www.solar.sheffield.ac.uk/pvlive/api/

Documentation for querying generation data aggregated by GSP region can be found here: https://docs.google.com/document/d/e/2PACX-1vSDFb-6dJ2kIFZnsl-pBQvcH4inNQCA4lYL9cwo80bEHQeTK8fONLOgDf6Wm4ze_fxonqK3EVBVoAIz/pub#h.9d97iox3wzmd

NWP (Numerical weather predictions)
OCF maintains a Zarr formatted version the German Weather Service's (DWD) ICON-EU NWP model here: https://huggingface.co/datasets/openclimatefix/dwd-icon-eu which includes the UK

PV
OCF maintains a dataset of PV generation from 1311 private PV installations here: https://huggingface.co/datasets/openclimatefix/uk_pv

Connecting with ocf_datapipes for batch creation

Outside the PVNet repo, clone the ocf-datapipes repo and exit the conda env created for PVNet: https://github.com/openclimatefix/ocf_datapipes

git clone --depth 1 https://github.com/openclimatefix/ocf_datapipes.git
conda create -n ocf_datapipes python=3.10

Then go inside the ocf_datapipes repo to add packages

pip install ".[dev]"

Then exit this environment, and enter back into the pvnet conda environment and install ocf_datapies in editable mode (-e). This means the package is directly linked to the source code in the ocf_datapies repo.

pip install -e <PATH-TO-ocf_datapipes-REPO>

Generating pre-made batches of data for training/validation of PVNet

PVNet contains a script for generating batches of data suitable for training the PVNet models. To run the script you will need to make some modifications to the datamodule configuration.

Make sure you have copied the example configs (as already stated above):

cp -r configs.example configs

Set up and config example for batch creation

We will use the example of creating batches using data from gcp: /PVNet/configs/datamodule/configuration/gcp_configuration.yaml Ensure that the file paths are set to the correct locations in gcp_configuration.yaml.

PLACEHOLDER is used to indcate where to input the location of the files.

For OCF use cases, file locations can be found in template_configuration.yaml located alongside gcp_configuration.yaml.

In these configurations you can update the train, val & test periods to cover the data you have access to.

With your configuration in place, you can proceed to create batches. PVNet uses hydra which enables us to pass variables via the command line that will override the configuration defined in the ./configs directory.

When creating batches, an additional config is used which is passed into the batch creation script. This is the datamodule config located PVNet/configs/datamodule.

For this example we will be using the streamed_batches.yaml config. Like before, a placeholder variable is used when specifing which configuration to use:

configuration: "PLACEHOLDER.yaml"

This should be given the whole path to the config on your local machine, such as for our example it should be changed to:

configuration: "/FULL-PATH-TO-REPO/PVNet/configs/datamodule/configuration/gcp_configuration.yaml" `

Where FULL-PATH-TO-REPO represent the whole path to the PVNet repo on your local machine.

Running the batch creation script

Run the save_batches.py script to create batches with the following example arguments as:

python scripts/save_batches.py datamodule=streamed_batches +batch_output_dir="./output" +num_train_batches=10 +num_val_batches=5

In this function the datamodule argument looks for a config under PVNet/configs/datamodule. The examples here are either to use "premade_batches" or "streamed_batches".

Its important that the dates set for the training, validation and testing in the datamodule (streamed_batches.yaml) config are within the ranges of the dates set for the input features in the configuration (gcp_configuration.yaml).

If downloading private data from a gcp bucket make sure to authenticate gcloud (the public satellite data does not need authentication):

gcloud auth login

For files stored in multiple locations they can be added as list. For example from the gcp_configuration.yaml file we can change from satellite data stored on a bucket:

satellite:
    satellite_zarr_path: gs://solar-pv-nowcasting-data/satellite/EUMETSAT/SEVIRI_RSS/v4/2020_nonhrv.zarr

To satellite data hosted by Google:

satellite:
    satellite_zarr_paths:
      - "gs://public-datasets-eumetsat-solar-forecasting/satellite/EUMETSAT/SEVIRI_RSS/v4/2020_nonhrv.zarr"
      - "gs://public-datasets-eumetsat-solar-forecasting/satellite/EUMETSAT/SEVIRI_RSS/v4/2021_nonhrv.zarr"

Datapipes is currently set up to use 11 channels from the satellite data, the 12th of which is HRV and is not included in these.

Training PVNet

How PVNet is run is determined by the extensive configuration in the config files. The following configs have been tested to work using batches of data created using the steps and batch creation config mentioned above.

You should create the following configs before trying to train a model locally, as so:

In configs/datamodule/local_premade_batches.yaml:

_target_: pvnet.data.datamodule.DataModule
configuration: null
batch_dir: "./output" # where the batches are saved
num_workers: 20
prefetch_factor: 2
batch_size: 8

In configs/model/local_multimodal.yaml:

_target_: pvnet.models.multimodal.multimodal.Model

output_quantiles: [0.02, 0.1, 0.25, 0.5, 0.75, 0.9, 0.98]

#--------------------------------------------
# NWP encoder
#--------------------------------------------

nwp_encoders_dict:
  ukv:
    _target_: pvnet.models.multimodal.encoders.encoders3d.DefaultPVNet
    _partial_: True
    in_channels: 10
    out_features: 256
    number_of_conv3d_layers: 6
    conv3d_channels: 32
    image_size_pixels: 24

#--------------------------------------------
# Sat encoder settings
#--------------------------------------------

# Ignored as premade batches were created without satellite data
# sat_encoder:
#   _target_: pvnet.models.multimodal.encoders.encoders3d.DefaultPVNet
#   _partial_: True
#   in_channels: 11
#   out_features: 256
#   number_of_conv3d_layers: 6
#   conv3d_channels: 32
#   image_size_pixels: 24

add_image_embedding_channel: False

#--------------------------------------------
# PV encoder settings
#--------------------------------------------

pv_encoder:
  _target_: pvnet.models.multimodal.site_encoders.encoders.SingleAttentionNetwork
  _partial_: True
  num_sites: 349
  out_features: 40
  num_heads: 4
  kdim: 40
  pv_id_embed_dim: 20

#--------------------------------------------
# Tabular network settings
#--------------------------------------------

output_network:
  _target_: pvnet.models.multimodal.linear_networks.networks.ResFCNet2
  _partial_: True
  fc_hidden_features: 128
  n_res_blocks: 6
  res_block_layers: 2
  dropout_frac: 0.0

embedding_dim: 16
include_sun: True
include_gsp_yield_history: False

#--------------------------------------------
# Times
#--------------------------------------------

# Foreast and time settings
history_minutes: 60
forecast_minutes: 120

min_sat_delay_minutes: 60

sat_history_minutes: 90
pv_history_minutes: 60

# These must be set for each NWP encoder
nwp_history_minutes:
  ukv: 60
nwp_forecast_minutes:
  ukv: 120

# ----------------------------------------------
# Optimizer
# ----------------------------------------------
optimizer:
  _target_: pvnet.optimizers.EmbAdamWReduceLROnPlateau
  lr: 0.0001
  weight_decay: 0.01
  amsgrad: True
  patience: 5
  factor: 0.1
  threshold: 0.002

In configs/local_trainer.yaml:

_target_: lightning.pytorch.trainer.trainer.Trainer

accelerator: cpu # Important if running on a system without a supported GPU
devices: auto

min_epochs: null
max_epochs: null
reload_dataloaders_every_n_epochs: 0
num_sanity_val_steps: 8
fast_dev_run: false
accumulate_grad_batches: 4
log_every_n_steps: 50

And finally update defaults in the main ./configs/config.yaml file to use your customised config files:

defaults:
  - trainer: local_trainer.yaml
  - model: local_multimodal.yaml
  - datamodule: local_premade_batches.yaml
  - callbacks: null
  - logger: csv.yaml
  - experiment: null
  - hparams_search: null
  - hydra: default.yaml

Assuming you ran the save_batches.py script to generate some premade train and val data batches, you can now train PVNet by running:

python run.py

Backtest

If you have succesfully trained a PVNet model and have a saved model checkpoint you can create a backtest using this, e.g. forecasts on historical data to evaluate forecast accuracy/skill. This can be done by running one of the scripts in this repo such as the UK gsp backtest script or the the pv site backtest script, further info on how to run these are in each backtest file.

Testing

You can use python -m pytest tests to run tests