Predicting Electricity Distribution Feeder Failures Using Machine Learning Susceptibility Analysis

Philip Gross, Albert Boulanger, Marta Arias, David L. Waltz, Philip M. Long, Charles Lawson, Roger Anderson, Matthew Koenig, Mark Mastrocinque, William Fairechio, John A. Johnson, Serena Lee, Frank Doherty, Arthur KressnerAAAI 2006

Overview

Researchers from Columbia University and Consolidated Edison (Con Edison) have developed ROAMS (Ranker for Open-Auto Maintenance Scheduling), a machine learning system designed to rank New York City’s electricity distribution feeders by their susceptibility to failure. By employing a novel boosting algorithm called MartiRank, the system successfully identified 75% of actual summer failures within the top 25% of its "at-risk" rankings. This shift from reactive repairs to proactive, condition-based maintenance represents a significant advancement in urban grid management.

The Challenge of "Open Autos"

The New York City electricity grid is a marvel of engineering, yet it faces persistent challenges known as "Open Autos" (O/As). These are spontaneous failures of primary feeder cables—the high-voltage lines connecting substations to local transformers—that trigger automatic isolation to prevent wider damage.

Overview of the electricity distribution chain from generation to end-users.

During peak demand periods, such as summer heatwaves, a single feeder failure can shift the electrical load onto neighboring lines. This increased stress can lead to cascading failures. Traditionally, maintenance was either reactive or followed rigid, time-based schedules. Con Edison required a more sophisticated method to prioritize which of their approximately 1,000 feeders were most vulnerable before critical stress events occurred.

Methodology: The MartiRank Algorithm

The ROAMS system utilizes MartiRank, a specialized ranking algorithm based on Martingale Boosting. Unlike standard classification models that predict a binary "fail" or "not fail" outcome, MartiRank focuses on susceptibility ranking. It evaluates various attributes—such as cable age and recent load spikes—to sort feeders by their likelihood of failure.

Key advantages of MartiRank include:

  • Interpretability: Unlike "black-box" models, MartiRank produces results that engineers can audit. If a feeder is flagged as high-risk, operators can identify the specific contributing factors, such as a particular joint type or recent test results.
  • Optimization for Ranking: The algorithm is specifically designed to optimize the order of a list, ensuring that the most critical components appear at the top.

Data Integration and Analysis

The study leveraged a massive, heterogeneous dataset from Con Edison covering 941 underground primary feeders across Manhattan, Brooklyn, Queens, and the Bronx.

Substation infrastructure, a key component in the electricity distribution network.

The data categories included:

  • Static Data: Physical characteristics like cable age, section composition, and joint types.
  • Dynamic Data: Real-time electrical load data for feeders and transformers, processed at a rate of several hundred megabytes per day.
  • Historical Data: Past outage records and "Hi-Pot" (High Potential) test results, which involve applying high voltage to identify latent defects.

Key Results and Impact

The system's performance was measured using a normalized average rank of failures. During the summer of 2005, ROAMS achieved a performance score of 0.8 (where 1.0 is perfect and 0.5 is random).

Significant findings included:

  • Risk Concentration: 75% of actual failures occurred in feeders ranked in the top 25% of the susceptibility list.
  • Operational Efficiency: The top 10% of feeders accounted for 40% of all daily failures.
  • Real-Time Monitoring: The system was integrated into Con Edison’s control center, providing vulnerability updates every 20 minutes.
  • Maintenance Optimization: Feeders ranked as "low risk" consistently passed expensive Hi-Pot tests, allowing the utility to skip unnecessary testing and reduce equipment stress.

Future Directions

While ROAMS has proven highly effective, researchers are exploring several enhancements:

  • Automated Retraining: Developing dynamic algorithms to automate model updates when performance shifts.
  • Predictive Timing: Moving from ranking "which" feeder will fail to using Survival Analysis to predict "when" a failure might occur.
  • Causal Discovery: Investigating the underlying causes of failures to inform long-term grid engineering and infrastructure improvements.

Summary updated

BibTeX
@inproceedings{gross2006predicting,
  author = {Philip Gross and Albert Boulanger and Marta Arias and David L. Waltz and Philip M. Long and Charles Lawson and Roger Anderson and Matthew Koenig and Mark Mastrocinque and William Fairechio and John A. Johnson and Serena Lee and Frank Doherty and Arthur Kressner},
  title = {{Predicting Electricity Distribution Feeder Failures Using Machine Learning Susceptibility Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure