University of Texas
Morgan Kelley says her mother suggested going into chemical engineering. “I didn’t really know what else I wanted to do” and “I did like chemistry,” says Kelley, now a process systems Ph.D. candidate at the University of Texas at Austin.
It was a good choice. Kelley earned a ChemE degree at Arizona State University. Now, as a Department of Energy Computational Science Graduate Fellowship (DOE CSGF) recipient, she uses high-performance computing (HPC) to optimize industrial production schedules, which involves managing demand response (DR) to cut operating costs and enable higher renewable energy production on the electric grid.
Kelley has long been interested in employing energy, especially from renewable sources, to improve lives. At ASU, she helped design and build solar-powered lights for schools in Fiji. Students could join the modular units her team produced to light their classroom and then separate and take them home to illuminate their studies.
Kelley also participated in the National Academy of Engineering’s Grand Challenges Scholars Program, taking courses in environmental sustainability, energy systems, and energy and technology policy.
“That very much inspired me” to pursue doctoral studies, Kelley says. She likes that her research, with Michael Baldea and Ross Baldick, emphasizes changing habits rather than buying equipment. “We can just adjust our behaviors, the way we interact with the power grid, and make a big difference.”
While residential consumers sometimes can’t do that – no one wants to forgo air conditioning when it’s 100 degrees – manufacturers have flexibility. They also are true needle-movers, Kelley and colleagues say in a 2017 Applied Energy paper, accounting for 32 percent of U.S. electricity consumption in 2013.
The researchers test their methods on air-separation units, which pull apart oxygen, nitrogen and argon from the atmosphere. “If we can predict what our electricity prices will be we can design a plant schedule so that when electricity is cheap, they overproduce and store this extra product,” such as liquid nitrogen, for distribution when the plant lowers production during peak power demand and price.
In a sense, the reserved chemicals are batteries, storing material instead of power. That makes DR useful for coping with solar and wind – time-varying energy sources.
But chemical plants can’t easily start and stop production, so a schedule that appears optimal may not be safe or feasible. A viable solution must meet conditions, called constraints, for large and small factors that may change over time.
A useful model must look far enough ahead to let the plant accelerate when power is cheap and slow down when it’s expensive. But the more distant the time horizon, the less certain the outcome and the more factors the program must consider. Variables pile up, multiplied by the time steps calculated, yet the code must run on desktop computers, since few chemical plants have supercomputers.
Kelley reduced the model, omitting variables that have little or no impact on the outcome. Using data from a first-principles air-separation-unit simulation, she framed the problem as a mixed-integer linear program, using low-order data-driven models that calculate both continuous and binary variables.
Kelley tested her code on supercomputers at UT Austin’s Texas Advanced Computing Center, running dozens of scenarios that varied electricity prices and other factors. It originally took up to 97 hours to run but now takes less than a minute – on a desktop computer.
The calculations showed DR could cut electricity use by 3 to 15 percent – a result that should improve when the scheduling problem considers an air-separation-unit network.
Kelley also studied optimization on her 2018 Argonne National Laboratory practicum with Sven Leyffer. She focused on uncertain parameters, or factors, affecting the solution.
Kelley studied chance constraints, in which researchers set a minimum probability tolerance for meeting a necessary condition subject to an uncertain parameter. For example, plant operators may want to be 95 percent sure they can satisfy demand and will seek operating schedules that meet or exceed that number.
Uncertain parameter values are represented with probability distributions, which capture the likelihood that each will occur. Kelley’s solution uses the cumulative density function, summing probabilities of events across the distribution to find aggregate values that meet the minimum tolerance.
Kelley hasn’t directly used the technique in her thesis research, but she’s tapped chance constraints to deal with uncertain parameters in her models.
Kelley is interning – virtually – in data science at Dell Technologies near Austin and may stay there after her 2021 graduation or seek a postdoctoral research post instead.
Image caption: Industrial demand response scheduling uses electricity price and grid-side emissions data, enabling load-shifting and reductions in emissions, operating costs and stress on the grid. Credit: Morgan Kelley.