Source-linked AI summary
Graph Machine Learning: An Opportunity for Power Systems
Martin Sadric, Sebastian Pütz, Christian Nauck, Veit Hagenmeyer, Frank Hellmann, Dirk Witthaut, Benjamin Schäfer
TL;DR
Power systems need faster, topology-aware methods as complexity grows and conventional approaches can be too slow. This paper surveys nearly 800 GML studies, maps applications and benchmarks, and finds that GML has progressed beyond proof-of-concept while scalability, interpretability, and data limitations persist.
Problem
Power-system complexity and operational time constraints motivate topology-aware alternatives to model-based methods that may not scale sufficiently.
Method
The paper systematically surveys nearly 800 GML studies, maps applications across core power-system tasks, and assesses benchmarks through an ML-ready dataset requirements catalog.
Results
GML has progressed beyond proof-of-concept in power systems, supporting forecasting, monitoring, and decision-making, while scalability, interpretability, and data availability remain challenges.
Takeaways & Limitations
GML complements traditional power-system methods and offers a high-impact setting for developing robust, generalizable, and inductive models.
Takeaways & Limitations
Publicly accessible high-fidelity operational data are scarce, especially at lower grid levels, because privacy and security constraints limit disclosure.
Abstract
from arXiv · showhide
Modern power systems face growing operational complexity driven by the integration of renewable energy sources, decentralization, and the need for real-time decision-making across a wide range of timescales. Addressing these challenges traditionally relies on model-based methods that, while accurate, can be too slow for operational demands. Machine learning (ML) has therefore emerged as a faster, data-driven alternative. As grid topology plays a central role in power system operation, graph machine learning (GML) methods offer a natural framework for incorporating topological dependencies as an inductive bias. We survey nearly 800 papers at the intersection of GML and power systems, covering forecasting, state estimation, optimization, control, fault diagnosis, and cybersecurity. Power systems constitute an unusually rich benchmark setting for GML, as they combine hard physical constraints, multi-scale dynamics, safety-critical requirements, and scarce labeled data within a single, well-defined domain. Conversely, power systems can benefit from utilizing GML to complement classical solvers, as GML provide scalable, topology-aware approximations with promising generalization and computational efficiency. We identify open challenges, including limited real-world deployment and the need for interpretable models in safety-critical settings. Despite the rapidly growing number of publications, standardized benchmarks and open datasets remain scarce, leaving many results difficult to reproduce and undermining the long-term scientific credibility of the field. We further derive a structured requirements catalog for ML-ready power grid benchmarks, intended to guide future dataset development and improve reproducibility across studies. We call on the community to prioritize dedicated benchmark studies and the release of open datasets and models.
1. Introduction
The transition to decentralized, low-carbon power systems creates operational challenges that demand scalable, data-driven methods capable of capturing structured grid relationships. This review surveys GML applications, maps them to power-system tasks, assesses benchmark requirements, and identifies technical gaps and future opportunities.
- Motivation: The shift toward sustainable energy introduces decentralized generation, grid-expansion needs, reduced system inertia, and severe consequences when critical infrastructure fails.These changes increase the complexity and operational demands facing modern power systems.
- Motivation: GML methods can learn complex system behavior from historical data and deliver forecasts, control policies, and situational awareness at operational timescales.This supports the use of machine learning as a scalable, data-driven tool for power-system applications.
- Motivation: Graph Machine Learning operates directly on entire graphs, making it well-suited to relational, structured, and spatial characteristics in power-system data.The passage contrasts GML with conventional ML methods based on fixed-size feature vectors.
- Contributions: The review surveys research published through the end of 2025, maps studies to core power-system tasks, and identifies critical technical and methodological gaps.It builds on and extends an earlier survey while incorporating trends emerging since 2021 and emphasizing interdisciplinary perspectives.
- Structure: The paper establishes GML foundations, explains graph representations of power systems, surveys applications, assesses benchmarks, catalogs dataset requirements, and identifies future opportunities.These topics correspond to the review’s Sections 2–6.
2. Graph Machine Learning
Graph machine learning models graph-structured data through approaches including message passing, with permutation-equivariant representations supporting node-, edge-, and graph-level tasks. Power systems require topology-aware architectures that handle heterogeneous components, physically meaningful directional edge features, and known message-passing limitations.
- Foundations: GML encompasses graph kernels, spectral methods, and GNNs, while message passing currently dominates the power systems literature.The review uses GML to denote the broader field rather than GNNs alone.
- Message Passing: Message-passing models iteratively exchange information between neighboring nodes to update representations, and MPNNs remain widely used because of their flexibility, efficiency, and effectiveness.Node features and adjacency constitute the two main GNN inputs, with optional edge and graph-level features.
- Graph Tasks: Permutation-equivariant message passing makes node representations independent of neighbor ordering, while outputs can target nodes, edges, or entire graphs.Graph-level predictions require permutation-invariant pooling such as sum, mean, max, or learned pooling.
- Power-System Graphs: Power-system graphs require heterophilic modeling because buses, generators, loads, and branches differ physically, even though neighboring nodes can be similar in weather-driven forecasting.Homogeneous GNN approaches have been shown to inadequately represent power-flow settings.
- Limitations and Extensions: MPNNs face oversmoothing and oversquashing, which can make node embeddings indistinguishable and lose information from exponentially growing distant neighborhoods.These failure modes motivate architectural extensions addressing limited expressivity, temporal dynamics, physical constraints, and power-system-specific structure.
3. Graph Representation of Power Systems
Power-system GML represents grid components and interactions as graphs, but the appropriate abstraction, scale, and graph-construction paradigm depend on the task. Accurate modeling also requires physical parameters beyond topology alone, because graph structure cannot capture power-flow physics by itself.
- Abstraction levels: Buses are typically modeled as nodes and branches as edges, while the representation’s granularity varies with the task.At finer granularity, individual loads, generators, and busbars may each be represented as distinct nodes.
- Transmission, subtransmission, and distribution: Power systems span transmission, subtransmission, and distribution subsystems with decreasing voltage magnitude and distinct network structures.Transmission is high-voltage and long-distance, whereas distribution delivers electricity mainly to commercial and residential customers.
- Scale and graph size: 103–104 nodes characterize the order of magnitude of transmission-level European models, while practical distribution studies commonly use 102–104-node feeders or regional subnetworks.Distribution networks can in principle reach 107–108 nodes with individual end users, but complete topology at that resolution is rarely available.
- Beyond topology: Topology alone is insufficient for accurate power-system modeling because graph structure cannot capture power-flow physics.Physical parameters such as line impedance are necessary for accurate vulnerability assessment.
- Graph construction paradigms: GML graphs may encode physical topology or instead represent statistical or geographic relationships such as spatial proximity or correlation.Graph construction is therefore a modeling decision, and the resulting graph may not correspond to the physical grid.
4. Applications
Section 4 surveys GML applications across the grid’s operational logic, from state observation and forecasting to optimization, control, fault diagnosis, and cybersecurity. The survey reached 792 publications, with especially rapid growth in 2024 and 2025, while highlighting topology-aware gains in variable-topology optimization, scalable security assessment, and attack detection.
- Survey scope and organization: 792 publications were surveyed, including 234 in 2024 and 356 in 2025.Publication counts have grown rapidly across GML applications in power systems.
- Survey scope and organization: The applications are organized as observing the current state, predicting future conditions, acting through optimization and control, and defending against faults and adversarial threats.This ordering follows the grid’s operational logic.
- Optimization and control: GNNs outperform fully connected and convolutional models in variable-topology OPF, although simple fully connected networks suffice for fixed-topology problems.The comparison attributes the variable-topology advantage to GNNs’ ability to adapt to structural changes.
- Optimization and control: GNN-based OPF achieves near-optimal, near-feasible solutions robust to N-1 perturbations, with sub-linear scaling demonstrated up to 10,000 buses.The released OPFData dataset contains solved AC-OPF problems with N-1 perturbations for transparent architectural comparisons.
- Fault diagnosis and security assessment: GNN-based transient-stability assessment can be up to 1000 times faster than mathematical optimization while generalizing to unseen grid topologies.The framework classifies whether a medium-voltage grid satisfies the contingency criterion.
- Cybersecurity: Message-passing and spatiotemporal GNNs detect or localize false-data-injection and ramp attacks by exposing inconsistencies across measurements and topology, even with incomplete topology.These methods target attack origins and measurement integrity when classical residual-based detectors or full topological information are unavailable.
5. Datasets and Benchmarks
Credible progress in graph machine learning for power systems is constrained by scarce, fragmented data and a lack of shared, transparent benchmarks. Existing benchmarks cover static tasks more maturely than dynamic ones, motivating standardized graph representations, evaluation protocols, and reproducible releases.
- Benchmark motivation: Shared datasets and transparent benchmarks are essential because most studies use problem-specific settings and few release code, data, or trained models.Limited reporting of preprocessing and hyperparameter choices further hinders verification, comparison, and reuse.
- Data scarcity: Power-system data scarcity is severe: high-fidelity real-system models are rarely accessible, simulation is computationally prohibitive, and no framework jointly captures all relevant factors at high fidelity.The passage also notes that empirical operational data exists only in limited supply.
- GML for data scarcity: GML methods for imputation and synthetic-data generation show promise for addressing data constraints but remain proof-of-concept approaches that cannot replace curated reference datasets.Examples include exploiting network-mediated correlations for missing smart-meter readings and using GAN-GCN hybrids to generate realistic synthetic feeders.
- Coverage across tasks: Benchmark coverage is relatively mature for static power-flow and optimal-power-flow tasks but substantially more challenging for dynamic tasks involving transients, protection responses, and inverter behavior.Static dataset generation is comparatively straightforward because a single steady-state solve is computationally cheap.
- Benchmark requirements: GML-ready benchmarks should provide explicit graph semantics and standardized train, validation, and test protocols, including reproducible preprocessing and reporting requirements.The proposed criteria also call for clearly defined physical interpretations of node and edge features, hyperparameter search details, and shared code and trained models.
6. Challenges and Future Directions
GML methods often perform strongly in controlled settings but have limited transferability to operational power-system environments. Future work should develop integrated, hybrid, topology-adaptive, explainable, and realistically benchmarked methods for deployment.
- Overall future direction: GML methods must move beyond proof-of-concept demonstrations toward scalable, physically grounded, interpretable architectures rigorously evaluated on realistic benchmarks.The review calls for collaboration, shared benchmarking and reproducibility infrastructure, and attention to domain constraints.
- Emerging and future applications: Future power systems require integrated GML models that represent prosumers and coupled heating, transportation, and electricity sectors rather than isolated grid aspects.The review expects such integrated approaches to be challenging but substantially improve applicability.
- Emerging and future applications: Hybrid approaches should combine physics-based simulation reliability with GML representation capabilities while handling grids whose topologies evolve through expansion, faults, and reconfiguration.Real-world grids are not static, and topology changes can create combinatorially large configuration spaces, especially within substations.
- Emerging and future applications: GML should support grid expansion and upgrade planning, especially in distribution grids, and address storage placement, sizing, control, and market integration end-to-end.Synthetic dataset availability remains limited for distribution grids, while few studies address the full storage-optimization chain.
- Common GML challenges: Power-system GML must control overfitting, overcome oversmoothing and oversquashing, and address missing sensors, rare-event labels, restricted data access, and limited standardization.Reliable applications also require clarifying the expressivity needed by GML architectures and training strategies.
- Explainability: Nearly nine percent in 2021 and four to six percent thereafter describe the publication share for physics-informed modeling, indicating that PINN integration became an established component.Explainability remains essential for trust, debugging, scientific discovery, and deployment in critical power-system infrastructure.
7. Conclusion
Graph-based machine learning has moved beyond proof-of-concept use in power systems, supporting applications including forecasting, monitoring, decision-making, and grid analysis. The review positions GML as complementary to traditional methods while highlighting foundation models and unresolved benchmarking and deployment challenges.
- Conclusion: GML now supports real-world power-system applications, including forecasting, monitoring, and decision-making.The review describes this progress as extending beyond the proof-of-concept stage.
- Conclusion: Power systems provide a high-impact testing ground for robust, generalizable, and inductive GML models with strong relational structures.The passage also notes that applications span different homophily levels.
- Foundation models: GridFM exemplifies efforts to unify diverse grid tasks through a foundation model pretrained on large volumes of grid data using self-supervised learning.The model is designed specifically for the electric power grid and aims to capture underlying patterns.
- Relation to open challenges in GML: The review finds evidence that power systems address concerns about absent real-world GML applications and poorly constructed graphs.It contrasts these findings with limitations identified in recent GML literature.
- Conclusion: Applications from renewable-generation forecasting to optimal power flow and grid resilience support low-carbon electricity systems, but many results remain difficult to assess.The passage connects these applications to machine learning’s potential climate impact while noting unresolved evaluation concerns.
A. Literature Search Methodology
The review used a systematic Scopus-based search, supplemented by preprint databases and author-identified works, to locate relevant GNN applications in power grids and systems. Identified papers were manually screened at least by title and abstract, with coverage limited through December 31, 2025.
- Search strategy: The search queried Scopus and its beta preprint search for “Graph Neural Networks” combined with “Power Grid” or “Power Systems” across titles, abstracts, and keywords.The query targeted relevant publications on applying Graph Neural Networks to power grids and power systems.
- Search strategy: The preprint search drew on arXiv, bioRxiv, ChemRxiv, medRxiv, Research Square, SSRN, and TechRxiv to capture early-stage research.This complemented the structured database search with indexed preprint servers.
- Coverage and screening: The search covered publications through December 31, 2025 and included additional relevant works identified through the authors’ prior research experience.The authors supplemented database retrieval with works known independently from their research experience.
- Coverage and screening: Each identified paper was manually reviewed at minimum by title and abstract to assess its relevance.Manual screening followed the database and supplementary-work identification stages.
Declaration of Generative AI and AI-Assisted Technologies in the Manuscript Preparation Process
The authors used Claude for language, structure, abstract refinement, and visualization code, and Gemini solely for companion-webpage code; AI tools were not used for literature search or synthesis.
- AI-assisted technologies: Claude supported language and readability improvements, section structuring, abstract refinement, and code generation for Figures 3–6, which the authors verified and finalized.Gemini was used solely to generate code for the companion webpage.
- AI-assisted technologies: No AI-assisted tools were used in the literature search or synthesis process.
CRediT authorship contribution statement
The authors contributed across conceptualization, methodology, analysis, investigation, data curation, writing, supervision, project administration, and visualization.
- Martin Sadric led conceptualization, methodology, formal analysis, investigation, data curation, original-draft writing, and visualization.
- Sebastian Pütz contributed to conceptualization, methodology, investigation, original-draft writing, supervision, and project administration.
- Christian Nauck, Veit Hagenmeyer, Frank Hellmann, and Dirk Witthaut contributed to writing—review and editing.
- Benjamin Schäfer contributed to writing—review and editing and supervision.