Temporal Graph Neural Networks for Failure and Cause Prediction in Transmission Networks

Overview

Developed temporal Graph Neural Network (GNN) models for large-scale transmission infrastructure in collaboration with Pacific Gas and Electric Company (PG&E).

The project focuses on transforming raw asset inventories, span connectivity data, and historical outage records into graph-structured datasets to enable structure-level failure localization and cause prediction under highly imbalanced, real-world operational conditions.

The framework supports proactive reliability planning by shifting from reactive failure analysis to predictive, topology-aware risk assessment.

Technical Contributions

1. Scalable Graph Construction Pipeline

Designed and implemented a utility-scale graph-building engine that converts raw utility records into structured power-grid graphs:

• Nodes: transmission structures (poles/towers)
• Edges: physical spans connecting structures
• Node features: design attributes, age, environmental exposure layers, insulator aggregates
• Edge features: conductor properties, span geometry

Constructed event-level graph instances with shared topology and event-specific failure labels, enabling temporal train/validation/test splits aligned with operational timelines.

2. Multi-Task Temporal GNN Modeling

Developed multi-task learning architectures to jointly perform:

• Node-level failure localization (identify failed structure within large graph)
• Graph-level failure cause classification (≈10–12 operational cause categories)

Core modeling components:

• Graph Attention Network (GATv2) backbone
• Temporal data splits based on outage timestamps
• Pointer-style localization objective to handle extreme class imbalance
• Candidate-set sampling using k-hop neighborhoods
• Class-balanced and focal-style losses for rare-event learning

Designed topology-aware evaluation metrics including Recall@k and graph-distance-to-true-failure to reflect operational usefulness rather than naive accuracy.

3. Coarse-Graph and Feature Selection Experiments

Implemented supernode-based coarse graph representations to:

• Study information bottlenecks under aggregation
• Perform distance-aware soft labeling
• Evaluate top-k performance across reduced graph resolutions

This experimentation quantified trade-offs between computational efficiency and spatial localization precision.

4. Asset Lifecycle and Insulator Analytics

Developed analytical modules to examine:

• Insulator replacement vs addition detection
• Age distribution and material evolution trends
• Data currency validation across decades of records

Generated lifecycle summaries and structured reports to inform graph feature engineering and reliability insights.

5. Connectivity and GIS-Based Network Analytics

Conducted large-scale graph diagnostics:

• Connected component analysis
• Degree distributions and structural bottleneck identification
• Temporal failure trend analysis
• OpenStreetMap-based interactive GIS visualization of structures and failure clusters

Enabled domain experts to interpret learned patterns in spatial and network context.

Impact

• Converted heterogeneous utility asset records into scalable graph datasets with ~10⁵ nodes per network instance
• Enabled structure-level predictive failure localization under extreme class imbalance
• Integrated topology-aware metrics aligned with utility operational decision-making
• Provided interpretable spatial analytics linking network structure to failure patterns
• Supported transition toward predictive reliability planning and early-warning systems

Relevance

Demonstrates deployment of advanced graph machine learning in real-world utility operations.

Bridges:

• Infrastructure asset analytics
• Graph deep learning
• Temporal failure modeling
• GIS-based network interpretation
• Reliability engineering for large-scale power systems

Establishes a reusable blueprint for graph-based predictive maintenance in critical infrastructure networks.