Yuanyuan Shi

University of Washington

Position: PhD Candidate
Rising Stars year of participation: 2018
Bio

Yuanyuan Shi is a fourth-year PhD candidate in the department of Electrical and Computer Engineering at the University of Washington, advised by Baosen Zhang. She is also pursuing a master’s degree in statistics. Previously, she received a bachelor’s degree in control and system engineering from Nanjing University in 2015. Her research focuses on relating machine learning with control theory to solve emerging and open problems in complex physical system control (e.g., power system transportation networks).

Real-Time Control for Complex Physical Systems: A Tractable Data-Driven Approach

Real-time Control for Complex Physical Systems: A Tractable Data-Driven Approach
Decisions on how to best operate large-scale complex physical systems such as power systems, large commercial buildings, and transportation networks are becoming increasingly challenging because of the growing system complexity and uncertainty. For example, controlling a large building currently requires one to have a model describing its dynamics. However, the cost, time, and effort associated with learning an accurate first-principles dynamical model of a large building would be enormous. Therefore, for simplicity and convenience, many control frameworks adopt a linear model that can be far away from the true system dynamics. To overcome the drawbacks of real-time control with first-principles models or oversimplified linear models, we propose a new data-driven predictive control framework. We leverage the data obtained from sensor measurements to construct an input convex neural network for the physical system dynamics where the weights between neurons are constrained to be positive, and some direct “passthrough” layers are added for better representation power. Without loss of generality, we proved that input convex neural network can represent all convex system dynamics and achieve comparable performance in fitting non-convex systems. Also, we proposed a projected gradient method in solving the constrained optimal control problem where the system model is described by neural networks. Experiment results show that our data-driven control framework can achieve 20 more energy reduction compared with previous method for large building system and reduce the modeling time from years to minutes. Currently, a prototype implementation of the input convex neural network control framework is being implemented at University of Washington EE building.