Online Optimization and Learning for Sustainable Cyber-Human-Physical Systems

Abstract

Due to the increasing connectivity of modern technology and extensive data availability, large-scale societal systems have massive potential to improve reliability, reduce operational costs, ensure safety, and decrease their carbon footprint. However, the increasing complexity of these systems also brings forth new challenges that traditional engineering fields are not well equipped to handle. Rather, interdisciplinary solutions from many fields including optimization, control theory, communications, signal processing, economics, power engineering, and transportation engineering are required. Specifically, my research has focused on the upcoming Cyber-Human-Physical-Systems (CHPS) within this domain (e.g., transportation systems, power grids, and smart cities with cooperative building communities). A Cyber-Human-Physical System is any real-world system that involves physical components that are controlled by both algorithms and human input. For example, these systems operate by way of physical components and infrastructure, utilize algorithms for the control and optimization of the infrastructure, and rely on human input. Additionally, these societal systems seek to maximize profit or social welfare while operating subject to constraints and subject to inherently stochastic environments.

This thesis is focused on developing optimization frameworks and machine learning strategies to improve the operation of these complex modern societal infrastructure systems in uncertain environments. Namely, 1) leveraging recent advancements in online optimization for system scheduling with applications to electric vehicle (EV) charging and community energy storage (CES) management, 2) using the machine learning framework called Thompson Sampling for the design of effective price signals for an electricity aggregator passively learning customers’ price sensitivities while running a load shaping program and providing theoretical safety guarantees on critical infrastructure constraints, and 3) optimizing real-world workplace EV charging in an online fashion and scheduling the charging/routing of a real electric bus fleet to minimize operational costs.