Lf Energy launches modular OpenSTEF 4.0 forecast suite

LF Energy has released OpenSTEF 4.0, restructuring its open source short-term energy forecasting software.

OpenSTEF is already used in production by Dutch distribution system operator Alliander and by Sweden's Sigholm, in work linked to about 40% of the country's district heating supply.

The new version splits the software into separate packages, each covering a distinct part of the forecasting process. These include core data structures, forecasting models, a metalearning layer, a backtesting and evaluation framework called BEAM, and a package for pretrained foundation models.

LF Energy hosts the project as part of its open source energy software portfolio. Development involves a broader group of grid operators, technology suppliers, researchers and energy specialists.

Modular rebuild

The redesign shifts OpenSTEF from a tightly integrated codebase to a modular structure that users can fit into existing technology environments. Teams can adopt individual components rather than deploy the full stack, which may appeal to utilities and energy companies with established data engineering and machine learning operations.

At the centre of the release are three additions not previously part of the project. The first is a metalearning ensemble system that combines outputs from several forecasting models instead of relying on a single approach.

Different model types tend to perform differently under changing conditions. Tree-based models can capture more complex non-linear relationships, while linear models can extrapolate beyond the range of the training data, making the trade-off a recurring issue for operational forecasting teams.

OpenSTEF 4.0 addresses this with an ensemble model that runs multiple forecasters in parallel and decides, quantile by quantile, which model to trust more. The software includes two methods for combining those outputs: a weighted-average model and a stacked model that places another regressor on top of the base predictions.

Foundation model

The second major addition is a pretrained foundation-model forecaster. It can generate a forecast for a meter it has not seen before without requiring separate training for each target, according to LF Energy.

That could reduce operational overhead for organisations managing large numbers of assets or meters by removing the need to train and store individual models across thousands of endpoints. It may also help with new assets or sites that have only a short data history.

The foundation model sits alongside more traditional machine learning methods rather than replacing them. Project maintainers describe the software as model-agnostic, with classical models, ensembles and foundation models all available depending on the forecasting task and the characteristics of the data.

Model testing

The third addition is BEAM, a framework for backtesting, evaluation and analysis. It is intended to help users compare forecasting models more systematically and reduce the risk of flawed testing methods.

In energy forecasting, model comparisons can be undermined by inconsistent test windows, mismatched data splits and data leakage, where a model is inadvertently exposed to information that would not have been available in live operation. BEAM is designed to separate backtesting, evaluation and analysis so each stage can be rerun independently.

The framework also breaks down results by lead time, forecast generation time and rolling time windows. It includes controls intended to ensure that a model does not access future data during backtesting, LF Energy said.

Operational use

OpenSTEF has already moved beyond experimentation into operational use. Alliander runs the software at thousands of grid locations, while Sigholm uses it in Sweden's district heating sector.

Those deployments point to broader demand for software that helps utilities and energy operators forecast demand, congestion and other short-term shifts across increasingly complex networks. Short-term forecasting is becoming more important as grids absorb more distributed energy resources, electrification adds new load patterns, and operators seek finer control over flexibility.

The release also reflects a broader trend in open source infrastructure for energy systems, with utilities and technology firms sharing software components that previously would have been built in-house or acquired through proprietary platforms. LF Energy positions itself as a neutral host for such projects under the Linux Foundation umbrella.

Version 4.0 is available now through Python package installation tools, while users of the previous major release can continue running version 3 if needed. Updated documentation includes user guides, examples and a migration guide for teams moving from the older version, the maintainers said.

OpenSTEF 4.0 has been rebuilt around modular packages including openstef-core, openstef-models, openstef-meta, openstef-beam and openstef-foundation-models, according to LF Energy.