S4Casting is a forecasting toolkit built around several deep learning models:
- State Space Models (SSMs) such as S4 and selective SSMs (a.k.a. S6 / Mamba-style models)
- Transformer variants
It is designed for medium voltage power forecasting tasks, including:
- Short-term forecasting – covering time horizons of days, typically used for forecasts up to 2 days ahead.
The repo contains:
- Training loop (distributed / torchrun-ready)
- Evaluation and benchmarking pipeline
- Config-driven experiments (CPU / CUDA configs)
- Inference scripts for running trained models on new data
You can install the repo following the installation guide:
You can find detailed documentation for each component in the following links:
You can find example notebooks in the notebooks/ directory, demonstrating how to use the S4Casting toolkit for forecasting tasks.
Future s4casting models will be integrated into the LF Energy OpenSTEF project in the near future. OpenSTEF provides reusable, automated machine‑learning pipelines for generating accurate and explainable short‑term grid load forecasts (up to 48 hours ahead). By integrating s4casting models into OpenSTEF, we can leverage shared infrastructure, align on common data and ML patterns, and reduce model‑specific complexity. At the same time, this integration expands the set of available forecasting models within OpenSTEF, improving overall maintainability, consistency, and reuse across implementations.
This project is licensed under the Mozilla Public License, version 2.0 - see LICENSE for details.
This project includes third-party libraries, which are licensed under their own respective Open-Source licenses. SPDX-License-Identifier headers are used to show which license is applicable. The concerning license files can be found in the LICENSES directory.
Please read CODE_OF_CONDUCT, CONTRIBUTING, GOVERNANCE and RELEASE for details on the process for submitting pull requests to us.
For citing our work please see CITATION.
Please read SUPPORT for how to connect and get into contact with the S4Casting project.