Topological Machine Learning Methods for Power System Responses to Contingencies

Brian Bush, Yuzhou Chen, Dorcas Ofori-Boateng, Yulia R. GelAAAI 2021 · DOI

Introduction

Modern power grids are among the most complex engineered systems in existence, characterized by non-linear dynamics and intricate connectivity. Ensuring the resilience of these grids against "contingencies"—unplanned events such as transmission line failures or transformer outages—is a critical task for utility operators. Traditionally, assessing the impact of such failures requires solving high-dimensional power-flow equations. While accurate, these numerical simulations are computationally intensive, making real-time assessment of millions of potential failure scenarios practically impossible.

This paper presents MCDPNet (Multi-Channel Deep Portfolio Networks), a novel framework that leverages Topological Data Analysis (TDA) and Convolutional Neural Networks (CNNs) to predict grid responses to contingencies with high accuracy and unprecedented speed.

Problem Statement

The primary challenge in contingency analysis is the "curse of dimensionality." A grid with thousands of components has an astronomical number of possible failure combinations. Brute-force simulation cannot keep pace with the rapid decision-making required during extreme weather events or cyber-physical attacks. Furthermore, standard network metrics often overlook the subtle structural changes that lead to localized blackouts. There is a pressing need for "reduced-form" models that can approximate the physics of the grid while providing near-instantaneous predictions.

Methodology

The MCDPNet framework bridges the gap between physical simulation and machine learning through two primary innovations: Topological Data Analysis and a specialized CNN architecture.

Overview of the MCDPNet framework integrating power grid data, topological measurements, and deep neural networks for outage prediction.

Topological Data Analysis (TDA)

Rather than relying solely on adjacency matrices, the authors employ persistent homology to characterize the grid's "shape." By analyzing how connectivity evolves across different scales, the model extracts:

  • Betti Numbers: $\beta_0$ (representing connected components) and $\beta_1$ (representing loops or cycles). These capture the fundamental robustness of the network's paths.
  • Bottleneck Distances: A metric used to quantify the structural divergence between the healthy grid and a post-contingency state.

MCDPNet Architecture

To process these features, the authors developed a multi-stage pipeline:

  1. Feature Selection via Autoencoders: An autoencoder-based ranking system identifies "communal information" across features, reducing redundancy and focusing the model on the most predictive topological and electrical properties.
  2. Feature-to-Image Transformation: A key innovation is the conversion of tabular data—such as reactance, Betti numbers, and flow—into normalized distance matrices. By treating these matrices as multi-channel images, the framework can utilize the spatial-pattern-recognition capabilities of CNNs.
  3. Hybrid Weighting Schemes: The model evaluates the grid under five distinct lenses: unweighted connectivity, reactance-weighted, capacity-weighted, base-flow, and residual capacity.

Dataset and Experimental Setup

The framework was validated using the ACTIVSg2000 synthetic power system, a 2,000-bus model representing a realistic large-scale grid (e.g., the Texas interconnect). The dataset comprises 281,273 solved power-flow simulations, covering a wide range of outage scenarios varying in location and intensity. This massive synthetic dataset provides the "ground truth" for training the neural networks.

Key Results

The experimental results demonstrate that MCDPNet significantly outperforms traditional machine learning baselines and provides a massive computational advantage over numerical solvers.

Sensitivity vs. Specificity plots comparing different weighting schemes and topological factors in predicting grid response.
  • Predictive Accuracy: The model achieved a classification accuracy of 92.51% in identifying contingencies that result in unserved load (blackouts). This exceeds the performance of Random Forests (88.23%) and standard Support Vector Machines (87.74%).
  • Regression Performance: For quantifying the magnitude of load loss, MCDPNet reached an $R^2$ of 92.13, a substantial improvement over standard DeepNets (87.31).
  • Computational Efficiency: Most notably, the trained model generates predictions approximately 1,000 times faster than traditional contingency analysis algorithms, enabling real-time monitoring and rapid response.

An unexpected finding was the efficacy of unweighted graphs. Simply analyzing the change in connectivity (topology) often proved as effective as, or more effective than, using complex electrical weighting schemes for binary failure prediction.

Conclusion and Future Directions

MCDPNet demonstrates that combining topological insights with deep learning can transform power system resilience analysis. By treating grid connectivity as a geometric problem, the framework achieves the speed necessary for real-time operations without sacrificing the accuracy required for critical infrastructure.

Future research will focus on transfer learning to allow models trained on one geographic region to be deployed in another with minimal retraining. Additionally, the integration of Graph Convolutional Networks (GCNs) may further enhance the model's ability to respect the underlying physical laws of the power system.

Summary updated

BibTeX
@inproceedings{bush2021topological,
  author = {Brian Bush and Yuzhou Chen and Dorcas Ofori-Boateng and Yulia R. Gel},
  title = {{Topological Machine Learning Methods for Power System Responses to Contingencies}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17791},
}
Figure