Anomaly Detection and Localization via Graph Learning

This technology introduces a graph learning-based framework for detecting and localizing anomalies in power systems using distributed PMU data. The core innovation lies in modeling the grid as a dynamic spatiotemporal graph, where each node represents a bus (or measurement point), and edges represent relationships between voltage measurements across the network.

The system constructs weighted graphs at successive time intervals using PMU voltage data. These graphs capture correlations between nodes, effectively encoding both spatial and temporal dependencies. A key component is the estimation of a generalized graph Laplacian matrix, which mathematically represents the structure and connectivity of the system. This matrix is iteratively optimized using statistical learning techniques to best reflect real-time system behavior.

Anomaly detection occurs in two stages. First, global connectivity of the graph is evaluated using the Fiedler value (second smallest eigenvalue). A significant drop in this value indicates abnormal system-wide behavior. For example, as shown in results (page 8), the anomaly segment exhibits a drastic connectivity reduction (λ₂ ≈ 0.002 vs. ~0.55 under normal conditions), signaling a disturbance.

Second, local connectivity analysis identifies the exact location of anomalies. Each node’s weighted degree is assessed, with lower connectivity indicating higher likelihood of fault or disturbance. In the case study (page 9), the system correctly identifies the faulted bus by detecting the lowest node connectivity.

The approach is validated using both simulated (IEEE 39-bus system) and real-world PMU datasets, achieving 97% detection accuracy and strong precision/recall performance.

Benefits

  • High Accuracy Detection: Achieves up to 97% accuracy with strong recall performance
  • Real-Time Monitoring: Leverages high-frequency PMU data for rapid anomaly detection
  • Precise Localization: Identifies exact fault locations using node-level analysis
  • Model-Free Approach: Avoids reliance on predefined thresholds or labeled datasets
  • Scalable Framework: Handles large-scale, distributed sensor networks
  • Enhanced Grid Resilience: Enables early detection of faults and disturbances

Applications

  • Smart grid monitoring and control
  • Fault detection in transmission and distribution networks
  • Renewable energy integration monitoring
  • Cyber-physical security in power systems
  • Industrial energy management systems
  • Critical infrastructure monitoring

Patents

This technology is patent pending in the US and is available for licensing/partnering opportunities.

Patent Information: