Research

Safe and Reliable Control of Cyber-Physical Systems

Learning-based control techniques are gradually being integrated into the operation of cyber-physical systems, from electric power systems to buildings to autonomous vehicles. The operation of these systems is governed by the laws of physics and is subject to various constraints. We explore how physics and operational constraints can be incorporated in the design of these optimization and control techniques to ensure the safe, reliable, and robust operation of these systems.

Privacy-Utility Tradeoffs in Mobile and IoT Systems

IoT and mobile devices are equipped with a myriad of sensors to monitor the environment and understand user behavior. The sheer amount of data collected by these sensors poses serious threats to our privacy. For example, building occupancy can be inferred from data collected by smart thermostats, people's health and wellness can be inferred from conversations recorded by smart home personal assistant devices, and their age and gender can be determined by a fitness tracking application running on their smartphone or smartwatch. To prevent these unwanted inferences and minimize distortion of useful data, we train machine learning models (e.g., generative adversarial networks, conditional diffusion models) on sensor data streams to obscure private information contained in the sensor data. These techniques enable data sharing with third party applications without compromising the user's privacy. See PrivDiffuser, PrivacyThroughDiffusion, ObscureNet, Blinder, and our privacy-preserving, latent space transformation technique.

Control Policy Transfer and Augmentation

Learning a high-quality policy—a mapping from perceived states to actions—through interaction with the real world takes a huge amount of time, especially in nontrivial control tasks, such as the optimal control of the HVAC system in buildings. To make this process less onerous, we proposed learning a finite pool of candidate control policies through interaction with a number of reference buildings in which training is affordable. We found empirically that for a sufficiently large pool of diverse policies, there exists a high-quality policy in the pool that would perform well if deployed to the target building. This motivated us to design a novel method for finding the policy that is expected to perform best when transferred to the target building. In follow-up work, we proposed a novel approach to augment black-box policy transfer methods for simultaneous minimization of energy consumption and thermal comfort constraint violation. This is an important problem because thermal comfort constraints are inherently specific to each building and cannot be considered during policy learning on reference buildings. We evaluated the proposed algorithm in various climate zones and showed that substantially energy savings can be achieved with theoretical guarantees for constraint satisfaction after a finite time.

System Identification with Physics-based Priors

Learning a dynamical system's model from a set of observations is crucial in many applications, from state estimation and control to detection and localization of topological changes, faults, and cyber attacks. This problem is known as the inverse problem. We have studied the inverse problem in the context of power distribution networks which are presently sparsely monitored. In our early work, we proposed an efficient algorithm for learning the topology and parameters of a radial distribution network from passive telemetry data when each node is equipped with a sensor. The physics-based prior that we used was the sparsity of the underlying graph. Our pioneering work considered the more realistic partially observable case, where only a subset of nodes are equipped with a sensor, and showed that the full model can be recovered under some assumptions.

In particular, we leverage the sparsity of power distribution networks along with the available sensor measurements and pseudo-measurements to (a) identify the network topology and model parameters, (b) estimate the state, (c) identify and pinpoint events to a small part of the network, and (d) monitor harmonic generation and propagation. In the context of buildings, we use time-series generated by smart thermostats to identify the thermal model of a home.

Adaptive Congestion Control: From the Internet to the Smart Grid

We addressed grid congestion from uncontrolled electric vehicle (EV) charging, a significant problem as EV penetration continues to rise. In our early work that received the 2024 Test of Time Award from ACM e-Energy, we extended decentralized algorithms for rate control in communication networks (Kelly, Maulloo and Tan 1998; Low and Lapsley 1999) to coordinate charging in a stable and fair fashion using explicit feedback. The proposed algorithm allows EV owners to plug in their car at will and have it charged at a variable rate, while sharing the limited available capacity of the network with other cars. We extended our early work by adapting the feedback controller's parameters to the network condition using a reinforcement learning (RL) agent. We showed through simulation that this learning-based control algorithm increases network utilization and alleviates congestion. This idea has become popular in recent years, leading to the design of adaptive controllers for datacenter networks and beyond.

In another line of work, we studied how a reputation-based scheme could encourage users to estimate their charging deadlines more conservatively.

Robust Sensor Fusion for Human Activity Recognition and Occupancy Estimation

Buildings are large-scale distributed systems housing hundreds of sensors, actuators, and processors. Yet they are largely unaware of the presence, preferences, and actions of the occupants. For example, the HVAC system typically runs on a static heating/cooling schedule that does not take occupancy into account, wasting a lot of energy in conditioning empty or partially-occupied spaces. We investigate how time-series data generated by heterogenous sensors can be fused to estimate the number of occupants and their thermal satisfaction, infer their activities, and localize them inside the building. Since sensor data streams are intermittent and noisy in real-world settings, we develop a multimodal fusion model that makes opportunistic use of the available sensor data, captures complementary information across different modalities, and is robust to sensor faults, network issues, and changes in the environment (see Centaur). Additionally, we study how this approximate knowledge of occupancy at the level of individual rooms can be incorporated in control loops.

Co-simulation of Smart Grids and Buildings

Energy systems are becoming increasingly complex with dynamics that span multiple timescales. Designing/validating new optimization and control algorithms for the smart grid requires simulation tools that accurately model the dynamics of the physical system and its interactions with various other domains, such as communication networks, building systems, transport systems, and electricity markets. We use concepts from operating systems and software engineering to design a modular, efficient, and secure co-simulation platform. Our smart grid co-simulation platform is called Maestro. We use this simulation platform to develop control and state estimation algorithms in addition to exploring cyber attacks (e.g., false data injection).

COBS is the simulation platform we developed for smart building. It is specifically designed to facilitate interactions between Reinforcement Learning (RL) agents and multiple building systems.

Teletraffic-based Sizing of Electric Power Grids

We used teletraffic theory to size power distribution networks just as it is used to size telecommunication access networks. Specifically, we proved the equivalence between a model of a distribution branch comprised of a transformer and storage that we want to size for a given underflow probability ε, and a queuing model that we want to size for a given overflow probability ε. Based on this equivalence, we showed how existing teletraffic analysis can be applied to jointly size transformers and storage.