Dartmouth Autonomy Seminar

"Responsible AI in Practice: A Realist Perspective from Industry"

Abstract

Electric heat pumps are one of the most important technologies for decarbonization,  providing the only viable alternative to fossil-fuel based space heating and cooling systems  in buildings.  Modern electric heat pumps are highly complex, mechatronic systems-of-systems, employing  variable-speed electric motors to drive compressors and fans, electronically actuated on/off  and expansion valves, and highly optimized heat exchangers, all interconnected and operated  to achieve energy-efficient and comfortable performance over a wide range of environmental conditions.

Control of these systems has been evolving for decades, but many challenges remain. Modern heat pumps operate over increasingly wide ranges of temperatures, heat loads, compressor and fan speeds, making robustness of feedback controllers an increasingly critical issue. Manufacturers offer a wide range of products, serving applications from a single room to a large commercial building.  Control architectures must scale to serve this range. The control problem is nonlinear, multivariable, subject to numerous constraints,  hybrid, and hierarchal, demanding careful attention to robustness with respect to model uncertainty.

In this talk, I will tell three stories about heat pump control.  First, I will explain how the  control algorithm currently in production was designed for a room air conditioner, which consists of a single outdoor unit and a single indoor unit.  Next, I will consider a model predictive controller for a  more complex house air conditioner, which has four to eight indoor units.  And finally I will show  an analysis of mass conservation for a dynamic model commonly used to represent thermofluid systems. It has been observed that fluid mass, in this case, the refrigerant mass, is not conserved during the course of a numerical simulation, and seems to "leak."  The analysis reveals the  root cause, and suggests a simple remedy rooted in classical control theory.  I will  close with a brief description of some current research interests at MERL, which include  estimation of unmeasured variables of interest in fluids problems.

The Dartmouth Autonomy Seminars explore how common principles of autonomy link fields such as robotics, economics, and cognition, and bring together academia and industry to discuss autonomous systems.

About the Speaker(s)

Scott Bortoff
Chief Scientist, MERL

Contact

For more information, contact Ada Yildirim at ada.yildirim.th@dartmouth.edu .