OpenSTEF provides automated machine learning pipelines to deliver accurate and explainable forecasts of the load on the grid for the next 48 hours. Infused with expert energy knowledge, the platform incorporates external predictors such as weather data and market prices, ensuring validation and training of machine learning models to provide precise load predictions.

Open

Description

The energy transition poses new challenges to all parties within the energy sector. Grid operators, grappling with the upsurge in renewable energy and heightened electrification, find their grid capacities nearing physical limitations. Therefore, it is imperative to forecast grid load in the upcoming hours to days, enabling the anticipation of local congestion and thereby optimal utilization of existing assets.

OpenSTEF provides a complete software stack specifically engineered to forecast the load on the electricity grid for the next hours to days. Given a timeseries of measured (net) load or generation, a fully automated machine learning pipeline is executed which delivers a probabilistic forecast of future load. This is applicable to energy consumption, renewable generation, or a combination of the two. OpenSTEF does not stop at forecating: it validates input data, combines measurements with external predictors such as weather data and market prices, trains any scikit-learn compatible machine learning model, and delivers the forecast via both an API and an (expert) graphical user interface. The entire stack, crafted on open-source technology and adhering to standards, is organized in a microservice architecture optimized for cloud-deployment.

The Dutch DSO Alliander started the Short-Term-Forecasting project to anticipate congestion in the distribution grid, to allow for grid safety analysis in the transmission grid and to enable smart grid innovations to locally balance supply and demand within the constraints of the grid. The objective of opensourcing the stack is two-fold: provide an industry standard for generating and evaluating forecasts in the operational time-domain, as well as allow for structured collaboration.

Framework features

• Prediction job: input configuration for a task or pipeline. It contains for example the location and forecasting horizon.

• Feature engineering: is called by pipelines to select required features for training/forecasting based on the configuration from the prediction job (e.g. create new features for energy load of yesterday, last week).

• Machine learning: is called by pipelines to perform training, forecasting, or evaluation based on the configuration from the prediction job (e.g. train an XGB quantile model).

• Pipelines: OpenSTEF contains automated machine learning pipelines. Can be called to perform training, forecasting or evaluating by giving input data to the pipeline.

• Data validation: called by pipelines to validate data.

• Confidence estimates: there are two methods for confidence estimation: quantile regression and standard deviation.

• Energy splitting: ability to split the load forecast into solar, wind and load. Uses the Domain Adaption for Zero Shot Learning in Sequence (DAZLS), which is a technique which transfers knowledge from complete information substations to incomplete information substations.

Find the latest information on the project on the projects website.

Want to join the discussion? Join our Teams channel.

You can also watch a video about OpenSTEF instead of reading about the project

Documentation

Technical documentation can be found here

Roadmap

The roadmap can be found here

Technical steering committee

More information on the OpenSTEF Technical steering committee can be found here.

Repositories

License

This project is licensed under the Mozilla Public License, version 2.0.

Contributing

Please read CODE_OF_CONDUCT.md, CONTRIBUTING.md and PROJECT_GOVERNANACE.md for details on the process for submitting pull requests to us.

Contact

Please read SUPPORT.md for how to connect and get into contact with the OpenSTEF project