Multilayer GNN for Predictive Maintenance and Clustering in Power Grids Journal: Renewable Energy
Muhammad Kazim, Harun Pirim, Chau Le, Trung Le, Om Prakash Yadav · 2025
Read Multilayer GNN for Predictive Maintenance and Clustering in Power Grids Journal: Renewable Energy by Muhammad Kazim, Harun Pirim, Chau Le, Trung Le, Om Prakash Yadav (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-07298.txt.
Introduction
Kazim and four colleagues build a layered graph network to predict power grid failures in twenty twenty-five. Unplanned outages cost the United States economy a vast sum each year. Past forecast tools miss how failures link across place, time, and cause. The team used seven years of outage records from an Oklahoma power firm.
Their network has three layers, one each for place, time, and cause, merged by learned weights. It beat standard tools and single layer networks at thirty day forecasts. Dropping the cause layer hurt scores a lot, which shows its worth. The team also grouped substations into eight risk bands to guide repairs before failures strike.
Catalog record
| CREATOR | Muhammad Kazim, Harun Pirim, Chau Le, Trung Le, Om Prakash Yadav |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 60 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-07298
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Multilayer GNN for Predictive Maintenance and Clustering in Power Grids Journal: Renewable Energy Muhammad Kazim, Harun Pirim, Chau Le, Trung Le, Om Prakash Yadav Abstract Unplanned power outages impose significant economic costs, exceeding $150 billion annually in the U.S., driven by the limitations of current predictive maintenance (PdM) models that overlook spatial, temporal, and causal interdependencies in power grid failures. This study proposes a multilayer Graph Neural Network (GNN) framework to enhance PdM and enable resilience-based substation clustering, leveraging seven years of incident data from Oklahoma Gas & Electric (292,830 records across 347 substations). Our model integrates specialized GNN layers—Graph Attention Networks for spatial dependencies, Graph Convolutional Networks for temporal patterns, and Graph Isomorphism Networks for causal interactions—fused through attention-weighted embeddings. For PdM, the framework achieves a 30-day F1-score of 0.8935 ± 0.0258 0.8935\pm 0.0258 , outperforming industry benchmarks like XGBoost by 3.2% and Random Forest by 2.7%, and surpassing single-layer GNN variants by 10–15%. Ablation studies underscore the causal layer’s importance, with its removal dropping the F1-score to 0.7354 ± 0.0418 0.7354\pm 0.0418 . For resilience analysis, HierarchicalRiskGNN embeddings, clustered via HDBSCAN, identify eight distinct operational risk groups, with the highest-risk cluster (Cluster 5, 44 substations) showing an incident rate of 388.4/year and recovery time of 602.6 minutes, compared to low-risk clusters (e.g., Cluster 2) with rates below 62/year. Statistical analysis (ANOVA, p < 0.0001 p<0.0001 ) confirms robust inter-cluster separation, while our clustering outperforms K-Means and Spectral Clustering with a Silhouette Score of 0.626 (vs. 0.315) and a Davies-Bouldin index of 0.527 (vs. 0.806). This approach advances proactive grid management, offering superior predictive accuracy and resilience insights, with future applications in multiclass failure prediction and grid analytics, supporting sustainable energy infrastructure modernization. 1 Introduction Energy infrastructure forms the backbone of economic productivity and societal functioning. In the United States alone, unplanned power outages cost over $150 billion annually [ 1 ] . To mitigate such disruptions, predictive maintenance (PdM) and resilience-based clustering have emerged as essential tools for proactive grid management. PdM enables the early detection of equipment degradation, reducing unplanned downtime by up to 70% and lowering maintenance costs by 30% [ 2 ] . Clustering, in parallel, supports operational resilience through zoning strategies such as islanding and prioritized maintenance planning. These techniques, when integrated effectively, provide not only economic benefits but also enhanced reliability and response capability. Predictive analytics in energy infrastructure has demonstrated return-on-investment (ROI) ratios as high as 10:1 [ 1 ] , yet existing frameworks fall short of leveraging the full complexity of interdependent power systems. A core limitation of conventional PdM and clustering methods lies in their treatment of infrastructure assets as isolated units. Traditional models—whether rule-based, statistical, or even machine learning-based—often fail to capture the cascading nature of failures that propagate through interconnected grid components [ 3 , 4 , 5 , 6 ] . These approaches neglect spatial proximity, temporal degradation trends, and causal dependencies among assets, resulting in inaccurate forecasts and fragmented resilience strategies [ 7 ] . Even graph-based methods, which introduce topological awareness, typically rely on single-layer representations that inadequately model the heterogeneous relationships inherent in modern power grids [ 6 ] . Previous studies have explored a variety of graph neural network (GNN) architectures to model complex relational data. For instance, spatial-temporal GNNs have been widely applied to dynamic graph problems such as traffic forecasting and social network analysis [ 8 , 9 ] . These models effectively capture spatial dependencies between nodes and temporal dynamics over time, making them suitable for systems with evolving interactions. However, in domains like power grids, where relationships between substations may shift due to operational changes or failures, spatial-temporal GNNs often assume a fixed graph structure, limiting their adaptability. This motivates our approach, which incorporates multiple types of dependencies—spatial, temporal, and statistical co-failure relationships—to better model the resilience of power grid substations. To address these structural limitations, multilayer network theory provides a formal framework for modeling interdependent infrastructures by preserving both intra-layer and inter-layer dynamics [ 10 ] . In the context of power systems, nodes such as substations, transformers, and transmission lines often participate in diverse relational layers—ranging from physical proximity to temporal fault sequences and causal pathways. Multilayer modeling captures these relationships explicitly, overcoming the loss of information typical in aggregated graphs [ 11 , 12 , 13 ] . Applications of multilayer analysis in the energy sector have demonstrated its value in identifying critical nodes whose failures induce cascading disruptions [ 14 ] , and in characterizing the resilience of coupled gas-electricity systems [ 15 ] . These findings establish multilayer modeling as a powerful tool for understanding and mitigating systemic vulnerabilities in energy infrastructures. Building on this foundation, Graph Neural Networks (GNNs) have emerged as a scalable and data-driven approach for learning from graph-structured data in PdM and resilience tasks. In power grids, GNNs have been used for overload prediction [ 16 ] , equipment health indexing [ 17 ] , and cyber-physical interaction modeling [ 18 ] . However, most existing GNN-based PdM models in energy systems focus exclusively on spatial features and ignore the temporal and causal aspects critical to forecasting real-world failures. Moreover, these models generally employ single-layer graphs, missing the nuanced, layered interdependencies across system components [ 19 ] . As a result, they fail to capture how failures evolve, how causal chains propagate, and how risks concentrate across regions and subsystems. Multilayer GNNs offer a promising solution by integrating spatial, temporal, and failure propagation patterns into a unified predictive framework. Early multilayer GNN approaches, such as supra-graph modeling [ 20 ] , simplify the representation but overlook the semantics of individual layers. While multilayer GNNs like MGCN [ 21 ] and attention-based multi-relational GNNs [ 22 ] have advanced layer-wise information fusion, they are primarily designed for entity-type heterogeneity rather than the spatiotemporal-causal dynamics essential for power grid modeling. For instance, MGCN focuses on integrating multiple relation types but does not explicitly account for temporal evolution or causal propagation of failures. Similarly, Li et al. 2023 proposed a multi-channel GNN for event detection in power systems but overlooked explicit multilayer graph structures and relied on static type labels, limiting its adaptability to dynamic grid conditions. Zhang et al. 2024 demonstrated that multilayer GNNs can outperform single-layer variants by 10–15% in wind power forecasting. However, a cohesive, end-to-end framework for multilayer PdM in power systems—one that combines topology, time-series behavior, and failure causality—remains lacking. Unlike prior studies that apply multilayer GNNs to isolated tasks, such as wind power forecasting [ 24 ] , our approach uniquely unifies predictive maintenance and resilience clustering within a single scalable framework, providing both predictive accuracy and actionable operational insights for grid management. Clustering in multilayer power grids plays a critical role in resilience planning, particularly for segmenting infrastructure into operational zones based on vulnerability or failure patterns. Traditional clustering methods—such as K-Means, Spectral Clustering, or even single-layer Louvain methods—fail to incorporate inter-layer dependencies and often assume equal importance across all features. This fragmentation limits the ability to systematically prioritize risk and segment the grid in a way that aligns with both predictive maintenance outcomes and real-time operational constraints. While generalized multilayer clustering techniques [ 25 ]