Posted on Leave a comment

GRL for Sequential Distribution System Restoration Learning


T. Zhao and J. Wang, “Learning Sequential Distribution System Restoration via Graph-Reinforcement Learning,” in IEEE Transactions on Power Systems, vol. 37, no. 2, pp. 1601-1611, March 2022, doi: 10.1109/TPWRS.2021.3102870,


DSR As a basic resilient model for system operators, a distribution service restoration algorithm provides an ideally coordinated and performance improvement on restoration. Restoration problem helps in Improved coordination between control switches and generators a model-based control system is often created to solve restoration problem but they depend on the accurate methods causes poor scalability.

To tackle these issues the author of this paper develops a novel GRL architect for the restoration issue in order to address these constraints. We connect the architecture of the power system to a GCA, which does an interplay between controllable devices and captures the process of network restoration in power networks. Graph convolutional layers generate latent characteristics across graphical power networks, which are then used to train the control policy for network restoration through DRL


Issues in the previous methods:

  •   Limitations include dependency on accurate or approximate power system methods, leading to no guaranteed solutions.
  • Computational time increases due to a large number of controllable components.
  •  Solutions often lack scalability.
  • Decentralized methods rely on precise information, limiting their effectiveness.
  •  Some consider DRL-based approaches, but they face challenges with growing distribution system scales, parameters, viable choices, and reconfiguration steps.
  • Using a single agent to gather central data is expensive and time-consuming.


So, the author of this paper proposes a GRL architect which tackle the aforementioned challenges.

The author proposes restoration problem Using a routing model that adheres to the MAS framework. The sequential restoration model is one of the DSR issues covered by the suggested formulation. The reward of this case is described as:

Ri,t(st, at) = P Ll,t × Δt


Without the knowledge of the power system characteristics, a novel GRL framework is developed to train efficient control methods for DSR issues sequentially by utilizing the graphical power system model and RL’s model-free functionality. It combines the power system topology with the GCN architecture. So, DGs extract latent graphical characteristics and abstract the mutual relations in order to understand the intricate repair process.


Figure of learning framework suggested in this paper from the study by T. Zhao(2022)


case studies are demonstrated to confirm the learning-based restoration framework’s efficiency in terms of scalability and optimality. We begin by outlining the system.
and setups for algorithms. Next, its scalability and optimality are illustrated and contrasted with benchmark methods. The changed version of IEEE 123-node and IEEE 8500-node is used to demonstrate the exceptional performance of this technology, comparative studies are carried out, and the IEEE 8500-node test systems are used to confirm its scalability.

image of results from the study by T. Zhao(2021)


  • Proposed a GRL framework to solve DSR problems.
  • GCNs records the impact of graphical user interface and network reconfiguration on controllable devices. Next, utilizing the latent characteristics generated by the DRL method, an efficient control policy for DSR is learned by GCN.
  • Case studies on IEEE 123- and 8500-node test systems show that proposed G-RL framework sign improves the efficiency and scalability of conventional RL algorithms, (DQN and MARL).

Sakthivel R

I am a First-year M.Sc., AIML student enrolled at SASTRA University

I possess a high level of proficiency in a variety of programming languages and frameworks including Python. I have experience with cloud and database technologies, including SQL, Excel, Pandas, Scikit, TensorFlow, Git, and Power BI

Leave a Reply

Your email address will not be published. Required fields are marked *