MPC LODL SESS

This repository contains the code (src directory) and result files (out directory) accompanying the paper “A Computationally Efficient End-to-End Learning Approach for Smart Energy Storage Systems”, accepted at IEEE ISGT Europe 2025.

It provides the implementation of the Locally Optimized Decision Loss (LODL) framework for Smart Energy Storage Systems (SESS). LODL learns a decision-aware loss function from MPC sensitivity analysis and uses it to train a recurrent forecasting model whose prediction errors are weighted according to their impact on control performance.

The SESS environment data and the recurrent prediction model are based on the MPC Predictor SESS Benchmark repository.

Note: The precomputed LODL target file (lodl_targets.pt) could not be uploaded to this repository, as it exceeds GitHub’s file size limit.
Please generate it locally using the provided scripts before running the experiments.

Citation

If you use this repository or find it helpful in your own research, please cite:

Note: This paper is accepted for publication and will appear in IEEE Xplore.

@inproceedings{Ludolfinger2025,
  title     = {A Computationally Efficient End-to-End Learning Approach for Smart Energy Storage Systems},
  author    = {Ludolfinger, Ulrich and Hamacher, Thomas and Martens, Maren},
  booktitle = {IEEE ISGT Europe 2025},
  year      = {2025},
  note      = {to appear}
}

Licensing

• The code and documentation are released under the MIT License.
• The electricity price data in res/data/ee_prices.csv are © Bundesnetzagentur | SMARD.de and redistributed under the CC BY 4.0 license.
• The files opsd_building_*.csv in res/data originate from the Open Power System Data project and are also redistributed under the CC BY 4.0 license.