Massachusetts Institute of Technology
Miles Lubin calls himself a methodology person.
“If people are doing things the wrong way, it just bothers me,” he chuckles. “So I’m trying to develop new techniques” to solve optimization problems, which seek the best route to an answer from a myriad of possibilities influenced by a multitude of variables.
The application that Lubin, a fourth-year Department of Energy Computational Science Graduate Fellowship (DOE CSGF) recipient, usually focuses his algorithms on is the power grid supplying electricity to homes and businesses. Utilities play a constant balancing act, producing only as much energy as consumers demand at any time. Too much or too little power in the system can cause damaging overloads, brownouts and blackouts.
The grid optimization problem has become even more thorny with the steady addition of solar and wind energy. Utilities must integrate that renewable electricity and manage the system “given the uncertainty that comes from how much wind is going to blow, how much energy you’re going to get from solar power.”
Lubin first dove into the problem at Argonne National Laboratory, near Chicago. He was an undergraduate applied mathematics student at the University of Chicago when he took a high-performance computing (HPC) course from a professor with ties to the lab.
“I ended up thinking that high-performance computing is pretty cool,” Lubin says. He interned at the lab one summer, continued working with his mentors as he finished school and then took a short-term post after graduation.
Although Lubin hadn’t planned on graduate school, working at Argonne convinced him a Ph.D. was within his reach and would advance his career. He entered the operations research program at the Massachusetts Institute of Technology, where he works with Juan Pablo Vielma.
Lubin still collaborates with Argonne researchers on power grid problems that only HPC can address. These problems sample thousands of projected grid operation scenarios under varying wind and solar generation conditions. “You want to do the best in the average case, so you have to consider all these” scenarios. “It could be an optimization problem with a billion variables,” a task so difficult few people would even consider attacking it. The Argonne research demonstrated HPC can solve these problems for certain kinds of models.
Lubin has taken a different tack on the problem since his 2014 DOE CSGF practicum at Los Alamos National Laboratory. Working with researchers Russell Bent, Michael Chertkov and Scott Backhaus, Lubin has sought ways to formulate the power grid problem so it’s solvable on less powerful computers – even laptops.
To do that, the researchers changed the way models deal with uncertainty. For example, under the usual approach a model might be designed so every acceptable scenario strictly respects limits on the amount of power sent over transmission lines, limiting uncertainty in the solution.
“The model that we were looking at turned that into a constraint that says I want to respect the transmission limits with high probability” rather than certainty, Lubin says. By constructing the model so the constraint holds only with high probability, the problem becomes easier to solve.
That doesn’t mean that the grid will instantly fail if power sent over a line exceeds the limit. Transmission capacities are soft constraints, Lubin says, and lines can handle overages for short periods.
With fellow MIT graduate students Iain Dunning and Joey Huchette, Lubin also developed JuMP, software in the Julia programming language that deals with optimization problems. “I think of it as designing the tools that I would like to use in order to solve these problems,” Lubin says. Many of the Los Alamos researchers he introduced to JuMP now use it regularly. At Argonne, he and fellow researchers developed a JuMP extension that makes it easier for engineers to formulate and solve optimization problems.
Lubin is still looking for the best ways to do things. His latest project tackles another variety of optimization puzzle: mixed-integer convex problems, a class of problems in which one may impose the challenging combinatorial constraint that certain variables can only be whole numbers.
The work started, Lubin says, from a discussion he had with Los Alamos postdoctoral researcher Emre Yamangil, who now works for social media company Snapchat. The algorithm implementation they devised is called Pajarito, for a mountain near the lab.
Lubin is continuing his power grid work but expects the new project will become his doctoral thesis subject. He hopes to graduate in spring 2017 and then seek a university position.
Image caption: A single-axis tracking flat-plate photovoltaic array near Sacramento, California. Credit: Warren Gretz, National Renewable Energy Laboratory.