Sustainable energy issues pose one of the largest challenges facing society: 84% of the world's energy currently comes from fossil fuels, raising major issues with climate change, energy security, and the long-term availability of these sources.  Although energy domains span a huge range of different areas, a common theme in many modern energy tasks is the availability of large amounts of data, and the need to learn models, make inferences, and control the system based upon this data.  These are problems that require new methods in machine learning, probabilistic inference, and control, and where such algorithms can have a profound impact on the energy space.  In this talk I will at two particular tasks spanning different extremes of energy consumption and generation and show how new algorithmic methods can play a pivotal role in each.

First, on the energy consumption side, I will present new techniques for energy disaggregation, the task of taking an aggregate (e.g. a whole-home) power signal and decomposing it into separate devices.  This ability helps us understand, in much greater detail than would be possible otherwise, how energy is consumed in a building, and studies have shown that just presenting this information to users can directly lead to large energy savings.  Unlike previous approaches to this problem, which typically focus on classifying single "events" in the power signal, my work considers models that look jointly at the entire signal and exploit the rich temporal structure of the data.  The key technical challenge here is the task of making inferences in these high-dimensional, factorized, temporal models, and I will present new algorithms I have developed, based upon convex relaxations of inference, that greatly outperform existing approaches on this task.  Second, on the energy generation side, I will present work on maximizing power output for wind turbines in low-wind conditions.  Existing approaches typically focus on model-based optimization, but aerodynamic models of wind turbines can be highly inaccurate.  Model-free reinforcement learning approaches, in contrast, are able to adaptively control the turbine in an online fashion without a model of the system, but existing methods, such as policy gradient approaches, typically use data very inefficiently.  I will present a novel policy learning approach, based upon trust-region optimization, which is able to maximize power using much less data than existing learning techniques, and which is able to produce 30% more power than the model-based approach on an experimental wind turbine.


Faculty Host: Ashutosh Saxena


B17 Upson Hall

Tuesday, February 21, 2012

Refreshments at 3:45pm in the Upson 4th Floor Atrium


Computer Science


Learning, Inference, and Control for Sustainable Energy

Zico Kolter