Research map / Graph learning / GNN architectures
PowerGraph: A power grid benchmark dataset for graph neural networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces PowerGraph, a set of GNN-ready power grid datasets (power flow, optimal power flow, cascading failure) with ground-truth explanations, and benchmarks GNN architectures and explainers on them.
Problem
Lack of publicly available graph datasets for training and benchmarking GNNs on power grid analysis tasks (power flow, OPF, cascading failures) and for evaluating explainability with real-world ground truth.
Approach
Builds graph datasets from MATPOWER power flow/OPF simulations and Cascades AC cascading-failure simulations on IEEE24, IEEE39, IEEE118 and UK grids, and uses the cascading-failure edges as ground-truth explanation masks for graph-level classification.
Key results
- Transformer models achieve node-level PF MSE of about 2e-05 to 4.4e-05 across grids, and OPF MSE of 3.8e-05 to 9e-05.
- Graph-level binary balanced accuracy reaches 0.98 to 0.998 (e.g., IEEE118 GIN 0.9982), but graph-level regression R2 peaks at 0.43 on IEEE24 and falls below 0.26 on other grids.
- GSAT is the only explainer that is both accurate and faithful, reaching balanced accuracy around 0.9 on IEEE24 and about 0.8 on UK; none recover cascading edges on IEEE39 or IEEE118.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a Transformer architecture rather than other benchmarked GNN architectures | PowerGraph node- and graph-level tasks · Power flow, optimal power flow, and cascading failure analysis across four grids | prediction performance | Helped |
| Use GCN rather than other benchmarked GNN architectures | PowerGraph power flow tasks · Power flow across power grids | prediction performance | Hurt |
| Use GNN regression for cascading-failure demand-not-served prediction | PowerGraph graph-level regression · IEEE24 and other power systems | R2: 0.43; below 0.26 | Hurt |
| Use GSAT to identify cascading edges | PowerGraph explainability benchmark · IEEE24 | balanced accuracy: 0.9 | Helped |
| Use GSAT to identify cascading edges | PowerGraph explainability benchmark · UK | balanced accuracy: almost 0.8 | Helped |
| Use explainability methods to identify cascading edges | PowerGraph explainability benchmark · IEEE39 and IEEE118 | balanced accuracy | Hurt |
| Use gradient-based explainers rather than other explainers | PowerGraph explainability benchmark · Three IEEE grids | faithfulness | Helped |
| Use Occlusion or GradCAM | PowerGraph explainability benchmark · UK | faithfulness | Helped |
| Use the Truth explainer rather than learned explainers | PowerGraph explainability benchmark · All four power grids | balanced accuracy: 1 | Helped |
| Use the Inverse explainer | PowerGraph explainability benchmark · All four power grids | balanced accuracy: 0.5 | Hurt |
Surprising findings
- Regression of demand not served remains poor (R2 below 0.26 for most grids), in contrast to high classification accuracy.
- Human-centric accuracy and model-centric faithfulness often disagree; ground-truth cascading edges are not faithful to the model on IEEE39 and IEEE118.
Limitations
- Uses IEEE test systems and synthetic cascading-failure simulations, limiting generalizability to real grids.
- Graph-level regression of demand not served performs poorly.
- Explainer performance is low on IEEE39 and IEEE118 due to few Category A instances and low cascading-edge ratio.
Experimental setup
Datasets: PowerGraph (IEEE24, IEEE39, IEEE118, UK)
Benchmarks: PowerGraph
Metrics: mean squared error balanced accuracy R2 fidelity+ balanced accuracy of explanations
Baselines: GCNConv GATConv GINEConv Gradient Boosted Trees Occlusion Saliency Integrated Gradients Grad-CAM GNNExplainer PGMExplainer SubgraphX RCExplainer
Builds on: MATPOWER Cascades GraphFramEx GSAT TransformerConv
Closest papers by meaning
- PF∆: A Benchmark Dataset for Power Flow under Load, Generation, and Topology Variations (NeurIPS 2025 · similarity 0.87)
- Power Grid Anomaly Detection via Hybrid LSTM-GIN Model (Student Abstract) (AAAI 2024 · similarity 0.87)
- Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification (NeurIPS 2024 · similarity 0.83)
- SGNNBench: A Holistic Evaluation of Spiking Graph Neural Networks on Large-scale Graphs (NeurIPS 2026 · similarity 0.83)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.83)
- Disentangled Table-Graph Representation for Interpretable Transmission Line Fault Location (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.