Thomas Catanach

California Institute of Technology

Everyday interconnections intrigue Thomas Catanach. “Network systems have become the backbone of our everyday lives,” says the Department of Energy Computational Science Graduate Fellowship recipient. Whether the power grid, telephone system, Internet or more, “they’re everywhere. They’re very complicated, and they’re essential to how we live,” two qualities that drive Catanach’s research into understanding the labyrinths around us.

The problem: It’s often difficult to comprehend or control a network because there’s limited information about what’s happening in it. “You can only observe so much of the world and you have to work out the rest,” says Catanach (pronounced CATanack).

The California Institute of Technology doctoral candidate initially considered going into particle physics. As a high school student in the Dallas area, he did a science fair project studying cosmic rays. He continued working in that vein as a University of Notre Dame undergraduate.

But Catanach also was drawn to the computational methods used in the field. During a summer research experience at the Santa Fe Institute, he worked on modeling financial networks, seeking ways to quantify their resiliency and probability of failure. Catanach was introduced to the power grid – the network application he focuses on now – during another summer program, his 2014 DOE CSGF practicum at Los Alamos National Laboratory.

The electric grid poses special challenges. Utilities must constantly judge its status, estimating demand at any minute and increasing or cutting generation to meet it. The rise of renewable energy resources like wind and solar, with their accompanying irregularity, makes the problem even more thorny.

Catanach says the practicum changed how he thinks of network analysis. Before, he had considered mostly static problems, in which the system changes little. With power grids, he must model a web with both static and rapidly changing conditions. The problem also is complicated because sensors on the system provide limited information about what’s happening in it.

To address what looks like “a fairly simple problem of how you translate data that you’re getting about a system into understanding what’s actually going on inside that system,” Catanach applies Bayesian inference. This mathematical technique weighs the validity of a network model or a set of possible network models, making judgments about which best reflects the real network based on incoming data. It establishes a probability distribution of how likely a model or set of models is correct under different conditions then updates that distribution as new information arrives.

Catanach: Flow Chart

Formulating the problem in those terms is relatively easy, Catanach says. What’s hard is solving it computationally. In his research under advisor Jim Beck, Catanach employs two approaches.

The first, Markov Chain Monte Carlo (MCMC), samples model scenarios from the probability distribution based on information about it. “MCMC works very well when you have a bunch of data and you’re trying to learn from it,” Catanach says, so it works well when networks can be modeled as static.

The second approach, filtering, works better when networks are changing quickly. Filtering techniques estimate a network’s state while considering uncertainty and other factors and continually updating the estimate with new data.

“My approach to these network systems is separating the different types of components we’re trying to estimate,” Catanach says. When applied to the electric grid, this layered architecture identifies rarely changing parts like line impedances and dynamic parts like sudden spikes in demand or line failures. His technique applies the proper method to each component to estimate the network’s state more quickly and efficiently than standard techniques.

“You can break up almost any system into these different components,” apply the right algorithm given the time scale at which they work best, and integrate the two results, Catanach says. “You can’t use the same MCMC algorithm that you want to estimate these static parameters to estimate something that changes” quickly, like a system failure when lighting strikes a power line. That requires methods that learn and integrate new data quickly.

Catanach’s technique should be applicable to many kinds of networks, including transportation, supply chains or epidemics. After graduating in 2017, he hopes to work on extending them to these or similar applications, probably as a postdoctoral researcher at a DOE national laboratory.

“I really like the work I’ve done with Los Alamos,” Catanach says. “I like that environment a lot. I think it’s a natural fit for the types of problems I like to think about.”

Image caption: A schematic of Thomas Catanach’s layered approach for fast power system estimation and control. Sensor data is used to update the system model through estimation and to identify disturbances by detecting changes at points in the system. Using the current model, predictions are made to test model accuracy and to determine control actions.