Untitled Document

Large scale autonomic systems can understand their own state and, if there are failures, recognize the situation then take actions to reconfigure. These systems can be applied to concerns in business and government about what is happening and what works best or what leads to failure. The application of machine learning systems can produce a significant return on investment in a wide range of projects, from live video scene tracking for security monitoring to prediction of effective financial investment patterns, using data on social networks of investors and analysts. Current and recent Con Edison projects illustrate how the Center is applying machine learning to crucial challenges confronted by the public utility providing electricity and gas to the New York City metropolitan area.

Con Edison’s Challenge: To ensure compliance with regulatory requirements and for legal reasons, Con Edison had saved over twenty years of records. New York City was one of the earliest major cities to be electrified over 180 years ago. Unlike more modern systems in cities like Tokyo and Las Vegas, most of the 30,000 plus cable sections of paper insulated lead sheathed cable, each a block long, are aging. Research Scientists at the Center are using machine learning capabilities to understand patterns (including 400 mgs of data about feeder failures daily), then predict which components are most susceptible to feeder failure and which are most likely to work well, based on age, location, weather and other conditions as well as historical events.
The Need to Understand Patterns of Failure: Because the maintenance and repairs of each of the 30,000+ cable section involves a series of tasks from pumping water and vermin out of manholes, shutting down power safely, pulling out old cable out of 3-16 splices at each end of the section and replacing them with new cable.
The Value of the Outcome: The work is increasing predictability of when and where feeders in the distribution system are susceptible to future failure. In 2005, CCLS algorithms, incorporating multiple attributes with the MartiRank method, predicted 40 percent of future outages for the data they had reviewed. As the computational systems “learn” more about the data, they reconfigure themselves. As the system generates models, CCLS research scientists refine rules, with the goal of moving towards 80 percent accuracy. This predictability gives Con Edison information needed to schedule maintenance in ways that will produce a geometric improvement in continuous power service to millions of customers and a reduction of costly emergency service.

Current and recent Con Edison Projects:

Columbia Learning System for Prioritization of Feeders for Long-Term Replacement Program for Cable and Splice Center of Excellence. Contract with Consolidated Edison.
Senior Team Members: Roger Anderson, David Waltz, Albert Boulanger, Phil Long

Columbia Decision Support System for the Manhattan Electric Control Center (MECC) 2005
Contract with Consolidated Edison.
Senior Team Members: Roger Anderson, David Waltz, Albert Boulanger, Phil Long

For more information, see article by Columbia’s Earth Institute, available online at http://www.earthinstitute.columbia.edu/news/2005/story06-01-05e.html