Roadmap 2025

Vision

Our open-source initiative aims to develop a cutting-edge machine learning pipelines for short-term energy load forecasting on the grid. These pipelines are designed for seamless adoption, boasting exceptional quality and harnessing built-in expert knowledge. The aim is to make OpenSTEF the industry standard for short term energy forecasting.

Fostering an active community, our project encourages collaborative efforts, fostering continuous improvement of the code base through contributions from multiple stakeholders.

Unique selling points

• Easy adoptability

•  

  • Automated pipelines make is very easy to train a model and make a forecast

  • The example notebooks provide an easy way to get familiar with the model

• Simple interface

•  

  • Extendable

  • Explainable

• Proven/benchmarked high forecasting quality

•  

  • Easily benchmarked by backtest

  • Implemented at companies such as Alliander and RTE

• Build-in expert knowledge on energy systems

•  

  • Close connection to academic knowledge (latest scientific innovation)

  • Feature engineering

  • Energy splitting

• Fully open-source

•  

  • Community of experts

• Flexibility to be able to implement in any enterprise environment

• Scalability

• Data standard on how to handle forecast (standardization)

Plans community 2025

Alliander - KTP team

• Improve forecast quality

• Improve forecast uncertainty estimation over horizons

  • focus on intraday using new weather sources

• Investigate potential of open sourcing deep learning model

• Investigate potential to include Dagster in the reference implementation

RTE

• 2025 S1 : Wind power forecast

  • We aim at reproducing of wind power forecast algorithm with Openstef fullstack solution. The challenges are :

•  

  •  

    • Use GroupedRegressor

    • Get model input with multiple tAhead (not only the latest)

    • Scaling load by installed capacity before fit and after predict

    • Extract multiple locations weather forecast

    • Using DAGSTER (or a challenger) to predict/train over 2500+ prediction jobs.

• 2025 S2 : National load forecast

  • We aim at using OpenSTEF python package only to develop our new prototype, which implies the following operations in Openstef :

•  

  •  

    • Use GroupedRegressor (one model per hour of the day

    • Using rpy2 to fit and train R/mgcv ::gam & qgam ML models

    • Agregation of predictions from 3 weather forecasts sources

RTE-i

• evaluate openSTEF predictions over the year

Sigholm

• make plans to use openSTEF as Sigholm's main forecasting tool

• looking into collaborating with Swedish DSO to implement openSTEF there

• Continue to enhance international compatibility of openSTEF

• collaborate on implementing S4 or other model improvements

Previous roadmaps 

2024:

Alliander - KTP team

• Improve forecast quality

• Dagster

  • May influence the reference implementation

  • Add maybe second reference implementation

  •  

    • Larger scale implementation

    • Add to description

•  

  • MLOPs architecture

• Improved DAZLs model (Q1)

• Update example notebooks

Firan

• Improve forecasting quality of peaks, with a very short time horizon

RTE

• Predict wind power on OpenSTEF.

  • Predict every hour for 72 hours.

  • 2500 windfarms.

• Implement new machine learning model

RTE-i

• Create a complete POC for demonstration purposes (Q1/Q2)

• Present a webinar on OpenSTEF (22nd of March, Q1)

Sigholm

• Deploy OpenSTEF on Sigholm cloud environment

• If requirements are met, switch to OpenSTEF as main forecasting tool.

• Demo to Sigholm customers

• Integrate with other Sigholm products.

Shell

• Implementation OpenSTEF-dbc realtime

Planned milestones

2022/2023:

Input Alliander:

Forecasting topics at Alliander 2022:

• Scale up ‘Contingencies’; locations with active capacity management

• T-prognoses customers; Including those in our load/generation forecasts

• Forecasting load at non-measured secondary substations

• Forecasting Reactive power (with few measurements)

Milestones related directly to OpenSTEF: (version 2022-02-21):

Input RTE:

Open