nowcasting-forecast 0.2.6

Live forecast for the OCF nowcasting project

Nowcasting Forecast

Making live forecasts for the nowcasting project. The aim is to produce Grid Supply Point (GSP) Solar generation Forecasts. This is done by loading live PV results, Numerical Weather Predictions (NWPs) and Satellite images, and running this data through various different ML models.

You can run the application locally by:

python nowcasting_forecast/app.py --db-url='sqlite:///test.db'

Directories and files

The following shows the main files

+-- nowcasting_forecast
|   +-- config
|       +-- mvp_v0.yaml
|       +-- mvp_v1.yaml
|   +-- models
|       +-- nwp_solar_simple.py
|       +-- nwp_solar_simple_trained
|           +-- model.py
|           +-- nwp_simple_trained.py
|   +-- app.py
|   +-- batch.py
|   +-- dataloader.py
+-- scripts
+-- tests

☀️ nowcasting_forecast

app.py is the main entry point for running the forecast module. It contains the following arguments:

• --db-url: the database url that the forecasts will be saved too
• --fake: Create fake forecast, mainly used for testing

The app has three main steps

1. Make batched data
2. Run forecasting model to make forecasts
3. Save forecasts to the database

batch.py is used to load the data and create batches of data. This uses mainly used ManagerLive

dataloader.py is used to load the batched data in an efficient way. This is current a stripped down version of this dataloader.

The config directory contains configurations that are used to load different data sources. The configuration is used by batch.py to convert large amounts of data into batched data, which is then ready for ML models.

The database directory contains database models and functions to interact with the database. See below for more information

The model directory contains ml models to make forecasts. See below for more information

Models

NWP Simple

This takes the average 'dswrf' for each example and then divides this by 10 to give a rough model for MW per GSP

configuration: mvp_v0.yaml

NWP Simple trained

CNN model of 'dswrf' channel in NWP. 6 CNN layers then 4 fully connected layers.

Training run: https://app.neptune.ai/o/OpenClimateFix/org/predict-pv-yield/e/PRED-951/charts

configuration: mvp_v1.yaml

🩺 Testing

Tests are run by using the following command

docker-compose -f test-docker-compose.yml run forecast

These sets up postgres in a docker container and runs the tests in another docker container. This slightly more complicated testing framework is needed (compared to running pytest) as some queries can not be fully tested on a sqlite database

🛠️ infrastructure

.github/workflows contains a number of CI actions

1. linters.yaml: Runs linting checks on the code
2. release.yaml: Make and pushes docker files on a new code release
3. test-docker.yaml': Runs tests on every push

The docker file is in the folder infrastructure/docker/

The version is bumped automatically for any push to main.

Environmental Variables

• DB_URL: The database url which the forecasts will be saved too
• NWP_PATH: Override NWP data path. THis is useful when running this locally.
• FAKE: Option to make fake/dummy forecasts
• MODEL_NAME: Optional of 'nwp_simple' or 'nwp_simple_trained'