Learning the Optimal Power Flow: Environment Design Matters

This is the accompanying repository to the publication "Learning the Optimal Power Flow: Environment Design Matters" by Thomas Wolgast and Astrid Nieße.

Note that most source code implemented for this paper can be found in https://gitlab.com/thomaswolgast/drl (RL algorithms) and https://github.com/Digitalized-Energy-Systems/opfgym (RL environments, warning: repository re-named and moved from GitLab to GitHub), which are work-in-progress and therefore continued in different repositories.

Especially the environment framework was significantly improved in the meantime. Please visit the GitHub repository above if you want to work with the most recent version

Installation

All experiments were performed with python 3.8. In an virtualenv, run pip install -r requirements.txt to install all dependencies in the right version at publication time (not the most recent).

Note: torch sometimes makes problems and needs to be installed manually before performing the previous step.

Repository structure

• run.sh: A list of the commands performed to reproduce all experiments done for this publication. Should not be run all at once! Overall computation time will be multiple weeks. Use this file to copy-paste single commands from. Will automatically create a data/ folder with the result files.
• LICENSE: The license used for this work (MIT).
• requirements.txt: Reference to the two previously mentioned repositories for simple installation.
• src/: The source code to aggregate the results and create the figures of this exact publication. The source code for running the experiments is in external repositories.
• compact_results/: Collection of JSON files with compact representations of the results with different metrics. (created with create_results.py and used by boxplots.py)
• results/: Optional folder that can be downloaded from Zenodo. Contains model weights, hyperparameter information, and agent performances in the course of training (e.g. MAPE every 30k steps). The model weights are required to run create_results.py. The training performances are required to plot performance during training, if required. Download this folder from Zenodo and place it in this repository to work with the data.

Attention: Sometimes, the VoltageControl environment is called QMarketEnv because of a name change during development. Both names are synonymous in the context of this work.

Contact

For questions or feedback, contact the first author Thomas Wolgast (thomas.wolgast@uol.de).