Multilayer GNN for predictive maintenance and clustering in power grids.

Area of Science:

• Electrical Engineering
• Data Science
• Network Science

Background:

• Unplanned power outages incur significant economic losses.
• Effective prediction and management of these outages are crucial for grid reliability.
• Existing methods may not fully capture complex interdependencies within power grids.

Purpose of the Study:

• To develop a predictive framework for unplanned power outages using a multilayer graph neural network (GNN).
• To integrate spatial, temporal, and statistical co-failure patterns for enhanced prediction accuracy.
• To support both predictive maintenance and resilience planning for electrical utilities.

Main Methods:

• Utilized 7 years of operational data from 347 substations.
• Developed a multilayer GNN incorporating weighted graph convolutions and attention mechanisms.
• Encoded spatial, short-term co-occurrence, and statistically enriched co-failure patterns.
• Applied clustering to categorize substations into risk groups based on incident rates and recovery times.

Main Results:

• Achieved a peak 30-day F1 score of 0.8935 for predicting substations needing intervention.
• Successfully clustered substations into eight distinct risk groups.
• Identified the highest-risk group with over six times the incident rate of low-risk groups.

Conclusions:

• The GNN framework provides an integrated approach for outage prediction and resilience planning.
• Enables utilities to optimize maintenance scheduling, inspection prioritization, and grid-hardening investments.
• Offers a data-driven strategy to minimize outage impacts and improve overall grid reliability.