Office 7002 Gates and Hillman Centers
Phone (412) 268-4733
Computer Science Department
My research focuses on computational approaches to sustainability and energy domains, focusing on core challenges arising in machine learning, optimization, and control in these areas. On the application side, my interests range from improving the efficiency of generation, controlling power in smart grids, and analyzing energy consumption in homes and buildings. To attack these problems I focus on techniques from machine learning, reinforcement learning, time series prediction, approximate inference, and convex optimization, amongst others. Some examples of specific projects I am working on include:
Home Energy Informatics. How can we understand a home's energy consumption by just looking at an (unlabelled) aggregated power signal for the whole home? And if we get such information, how can we use it to improve home efficiency? To address this problem, we are developing new methods in probabilistic inference for aggregation models, motif-based methods for HMM learning, and model predictive control for controlling appliances to conserve power. We have also worked on developing new data sets for this problem, to bring the tasks here into the ML community as a whole.
Power Routing for the Smart Grid. Power in the electrical grid is typically viewed as "unroutable;" unlike network traffic, for example, power flow must obey fundamental physical laws. This has raised many issues with efficiency (due to line capacity constraints) and economics in power grids, and has largely prevented the adoption of "CS-style" networking in power grids. However, recent advancements is power electronics have given us the power to control power flow in local ways. Our research is looking at control and networking methods that would enable arbitrary routing of power in electric grids, offering the potential to transform the smart grid into something that looks much more like computer networks.
Understanding National Energy Consumption. How many wind farms would it take to power the United States? While it's not hard to compute the answer from a raw power perspective, this ignores the indeterminacy of renewables and the losses associated with transmission. Our work looks to use the large amounts of publicly available data, including historical hourly power consumption and weather information, to arrive at an honest account of whether a given set of generators will actually be able to power the country (or smaller regions). Getting optimal answers to such problems requires solving large-scale optimization and resource allocation problems at very high speeds, to quickly give feedback to users of such a system.
Fundamental Algorithmic Advances. Many problems in energy and sustainability require new, massively scalable algorithms for probabilistic inference, learning and controlling dynamical systems, and optimization. We have several projects focused on these more "core" algorithmic areas, including: learning high-dimensional, sparse representations for dynamical systems; improving probabilistic inference with low-rank semidefinite programming; learning structured auto-regressive models for high-dimensional time series; and others.