S4Casting

S4Casting is an open source forecasting toolkit I have been working on as part of my R&D work at Alliander, a Dutch energy network operator. It applies time series foundation models to energy forecasting, built around deep learning architectures such as State Space Models (S4 and selective SSMs, a.k.a. S6 / Mamba-style models) and Transformer variants.

The toolkit is designed for medium voltage power forecasting, with a focus on short-term forecasts up to two days ahead. It includes a distributed training loop, an evaluation and benchmarking pipeline, config-driven experiments for CPU and CUDA, and inference scripts for running trained models on new data.

Future S4Casting models will be integrated into the LF Energy OpenSTEF project, which provides reusable, automated machine learning pipelines for accurate and explainable short-term grid load forecasts.

The project is licensed under the Mozilla Public License 2.0 and maintained under the Alliander open source organisation on GitHub.