Topology-Aware Reinforcement Learning over Graphs for Resilient Power Distribution Networks

Roshni Anna Jacob, Prithvi Poddar, Jaidev Goel, Souma Chowdhury, Yulia R. Gel, Jie Zhang

TL;DR

This paper introduces a "topology-aware" reinforcement learning framework designed to keep the lights on during extreme weather or cyberattacks. By integrating Topological Data Analysis (TDA)—specifically persistence homology—into a graph-based RL model, the researchers created a system that understands the deep structural "DNA" of a power grid. This approach allows for faster, more effective grid reconfiguration and load shedding, outperforming standard graph-based AI by providing up to 6% more power delivery and significantly fewer voltage violations.

What Problem Does This Paper Address?

When a storm knocks out power lines or a cyberattack disrupts a substation, power distribution networks (DNs) must "self-heal" by reconfiguring switches and shedding non-essential loads to prevent a total collapse. Traditionally, this is treated as a complex math problem (mixed-integer programming) that is too slow for real-time response.

While Reinforcement Learning (RL) has been proposed as a faster alternative, standard RL models struggle when the grid's topology changes drastically during an outage. The "state-action space" becomes massive and chaotic. Conventional Graph Neural Networks (GNNs) only look at immediate neighbors (who is physically connected to whom), often missing the "higher-order" structural patterns—like loops and voids—that determine how a grid can be successfully partitioned into resilient, self-sustaining islands.

Methods

The authors propose a framework called PH-GCAPCN (Persistence Homology-enhanced Graph Capsule Convolutional Network). The core components include:

  1. Persistence Homology (PH): This is a tool from TDA that analyzes the grid at multiple scales. It tracks how topological features (like connected components or cycles) "persist" as you vary a distance threshold. This creates a "Persistence Diagram"—a signature of the grid's structural health.
  2. Topological Edge Reweighting: Instead of just treating every connection equally, the model calculates the 2-Wasserstein distance between the persistence diagrams of different neighborhoods. It then reweights the graph's edges so that nodes with similar structural roles have stronger "logical" connections, even if they aren't physically next to each other.
  3. Graph Capsule Convolutional Neural Network (GCAPCN): This specialized GNN processes the reweighted graph to learn complex node embeddings.
  4. Proximal Policy Optimization (PPO): A robust RL algorithm that uses these embeddings to decide which switches to flip and which loads to shed to maximize energy supply while keeping voltages stable.

Datasets

The researchers used a modified IEEE 123-bus distribution test feeder. They enhanced this standard benchmark by:

  • Adding 13 sectionalizing switches and 9 tie switches.
  • Integrating 7 grid-forming Distributed Energy Resources (DERs) and 6 grid-feeding DERs (like solar/wind).
  • Training Set: 10,000 unique outage scenarios generated by simulating random failures within subgraphs of the feeder.
  • Testing Set: 300 disjoint outage scenarios (split into three sets of 100) to ensure the model could generalize to unseen disasters.

Key Insights and Ideas

  • Beyond Physical Proximity: The "Aha!" moment here is that physical connectivity isn't the only thing that matters in a crisis. By using TDA to reweight edges, the GNN can aggregate information from nodes that perform similar functional roles in the network's restoration, regardless of their physical distance.
  • Stability through Topology: Incorporating TDA didn't just make the model more accurate; it made the learning process more stable. The "topology-aware" model converged faster and showed less variance during training compared to the baseline.
  • Implicit Radiality: Interestingly, the authors found that they didn't need to strictly enforce "radiality" (the tree-like structure utilities usually prefer). The RL agent naturally learned that radial topologies were the most efficient for restoration based on the physics of the environment.

Key Results

The PH-enhanced model was compared against a baseline Graph RL model (GCAPCN) across 300 test cases:

  • Higher Rewards: The PH-GCAPCN model achieved 9-18% higher cumulative rewards.
  • Better Power Delivery: It delivered 4-6% more energy to customers during outages.
  • Improved Stability: It resulted in 6-8% fewer voltage violations, meaning the power delivered was of higher quality and safer for equipment.
  • Win Rate: In head-to-head comparisons, the topology-aware model "won" (achieved a higher reward) in over 82% of all test scenarios.

Limitations and Future Work

While the results are promising, the authors note a few areas for growth:

  • Interpretability: Like many deep learning models, it can be hard to "see" exactly why the agent chose a specific switching sequence. Future work will focus on making these AI decisions more transparent for grid operators.
  • Generator Dispatch: The current study focuses on switching and load shedding. The authors plan to extend the framework to include the active dispatch of generators and DERs during outages.
  • Computational Overhead: Calculating persistence diagrams adds a layer of complexity. While inference is still fast (around 0.018 seconds), scaling this to massive transmission-level grids will require further optimization.

Summary updated

Abstract

Extreme weather events and cyberattacks can cause component failures and disrupt the operation of power distribution networks (DNs), during which reconfiguration and load shedding are often adopted for resilience enhancement. This study introduces a topology-aware graph reinforcement learning (RL) framework for outage management that embeds higher-order topological features of the DN into a graph-based RL model, enabling reconfiguration and load shedding to maximize energy supply while maintaining operational stability. Results on the modified IEEE 123-bus feeder across 300 diverse outage scenarios demonstrate that incorporating the topological data analysis (TDA) tool, persistence homology (PH), yields 9-18% higher cumulative rewards, up to 6% increase in power delivery, and 6-8% fewer voltage violations compared to a baseline graph-RL model. These findings highlight the potential of integrating RL with TDA to enable self-healing in DNs, facilitating fast, adaptive, and automated restoration.
BibTeX
@misc{jacob2026topology,
  author = {Roshni Anna Jacob and Prithvi Poddar and Jaidev Goel and Souma Chowdhury and Yulia R. Gel and Jie Zhang},
  title = {{Topology-Aware Reinforcement Learning over Graphs for Resilient Power Distribution Networks}},
  year = {2026},
  eprint = {2603.06964},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.06964},
}
Figure