Graph machine learning represents a significant opportunity for power systems facing unprecedented operational complexity. As renewable energy sources proliferate and grids become increasingly decentralized, traditional model-based methods struggle to deliver real-time insights at the speed modern operations demand. A comprehensive academic survey examining nearly 800 papers at the intersection of graph ML and power systems identifies how these methods naturally accommodate the topological structure of electrical grids, creating what researchers call an "inductive bias" that traditional machine learning cannot provide.
The applications span the full spectrum of grid operations: load forecasting, state estimation, optimization, control algorithms, fault detection, and cybersecurity threat identification. GML methods deliver topology-aware approximations that scale efficiently across networks of varying sizes and configurations, offering faster computation than classical solvers while maintaining reasonable accuracy for many operational tasks.
Power systems present an unusually rich proving ground for graph ML research. Grids combine hard physical constraints with multi-scale dynamics ranging from sub-second stability events to seasonal demand patterns. Safety-critical requirements demand high reliability, yet labeled training data remains scarce—challenges absent from many other ML application domains.
Despite rapid publication growth, critical implementation gaps persist. Real-world grid deployments remain limited, with most work confined to simulation environments. Safety-critical applications demand interpretable models that explain their decisions, yet many GML approaches function as black boxes. Most significantly, the field lacks standardized benchmarks and openly available datasets, making reproduction of published results difficult and undermining scientific credibility.
The research community identifies a structured requirements catalog for ML-ready grid benchmarks, emphasizing the urgent need for openly available datasets and reproducible studies. Organizations controlling grid data must prioritize dataset release, and researchers must publish standardized benchmarks that enable fair comparison across methods. Without these foundational elements, graph machine learning risks remaining academically interesting but operationally impractical—unable to bridge the gap from research to grid control rooms.
