Dymola-Enabled Reinforcement Learning for Real-time Generator Set-point Optimization

Aisling Pigott, Kyri Baker, Sergio A. Dorado-Rojas, and Luigi Vanfretti

• Author's copy

• URL

bib

@inproceedings{pigott2022,
  title = {Dymola-{{Enabled Reinforcement Learning}} for {{Real-time Generator Set-point Optimization}}},
  booktitle = {2022 {{IEEE Power}} \& {{Energy Society Innovative Smart Grid Technologies Conference}} ({{ISGT}})},
  author = {Pigott, Aisling and Baker, Kyri and {Dorado-Rojas}, Sergio A. and Vanfretti, Luigi},
  year = 2022,
  month = apr,
  pages = {1--5},
  issn = {2472-8152},
  doi = {10.1109/ISGT50606.2022.9817464},
}

Abstract

This paper introduces a reinforcement learning framework which determines real-time generator set-points in a dynamically changing environment in order to optimize a chosen objective. This is performed within the high-fidelity simulation environment Dymola, which utilizes the Modelica programming language to model complex systems. A case study is created using the OpenIPSL IEEE 9-bus dynamic model, with the objective of minimizing voltage deviations across the network. The reinforcement learning agent shows improvement in minimizing voltage deviations versus the default droop controlled governors without any explicit knowledge of the topology of the system or relative location of the controllers. The results indicate that reinforcement learning may be a useful tool for applications in model-free, real-time power systems dynamics and control. An open-source Python package is provided for the proposed framework with the present case-study as an example.

Back